🛰️ Global AI Alpha

120 notable papers · updated 2026-08-17 18:02 UTC

News — AI & startups

Drops & buzz

Trending open models — Hugging Face

ModelCountryTaskDownloadsLikes
Qwen/Qwen3.8-27B🇨🇳 Chinaimage-text-to-text415,03910622
unsloth/Qwen3.8-27B-GGUF🇨🇳 China2,727,6091605
meta-models/Muse-Glimmer-30B🏳️ Otherimage-text-to-text334,0991652
Qwen/Qwen3.8-2.4T-A95B🇨🇳 Chinatext-generation9,4651033
Lightricks/LTX-2.5🏳️ Otherimage-to-video465,5291084
MiniMaxAI/MiniMax-Music3🇨🇳 Chinatext-to-audio10,375890
MiniMaxAI/MiniMax-H3🇨🇳 Chinaimage-text-to-video2,403,2384077
deepseek-ai/DeepSeek-V4-Pro-0813🇨🇳 Chinatext-generation25,006568
Qwen/Qwen3.8-27B-FP8🇨🇳 Chinaimage-text-to-text495,646521
orcarouter/Qwen3.8-27B-Uncensored-FP8🇨🇳 Chinaimage-text-to-text15,812419
deepseek-ai/DeepSeek-V4-Flash-0731🇨🇳 Chinatext-generation1,978,2983493
lightx2v/Minimax-h3-Turbo🇨🇳 Chinaimage-to-video264,351583
unsloth/Muse-Glimmer-30B-GGUF🏳️ Otherimage-text-to-text755,125465
nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4🇺🇸 UStext-generation231,271305
JonathanColetti/Qwen3.8-27B-Uncensored-GGUF🇨🇳 Chinatext-generation357,701282

Model capability — LMArena Elo

Frontier capability gap — US vs Chinese open-weight

1439146114821504152603-1903-3004-1004-2205-1405-2506-0606-1807-0208-05best US (1509.0)best CN open (1485.0)gap: 24.0 Elowas 45 on 03-19LMArena human-preference Elo · best US vs best Chinese open-weight model

Leaderboard (top 15) · as of Aug 12, 2026

#ModelLabCountryWeightsEloVotes
1claude-fable-5Anthropic🇺🇸 USclosed1506.021,533
2claude-opus-4-6-highAnthropic🇺🇸 USclosed1505.072,516
3claude-opus-4-7-highAnthropic🇺🇸 USclosed1502.060,331
4muse-spark-1.2 (xHigh)Meta🇺🇸 USclosed1498.03,280
5claude-opus-4-6Anthropic🇺🇸 USclosed1497.076,474
6claude-opus-4-7Anthropic🇺🇸 USclosed1494.061,419
7claude-opus-5-highAnthropic🇺🇸 USclosed1493.020,030
8qwen3.8-maxAlibaba🇨🇳 Chinaclosed1491.07,004
9gemini-3.7-flash-highGoogle🇺🇸 USclosed1490.05,744
10claude-opus-5-maxAnthropic🇺🇸 USclosed1489.09,679
11muse-spark-1.1Meta🇺🇸 USclosed1489.016,928
12kimi-k3-maxMoonshot🇨🇳 Chinaclosed1489.011,969
13muse-sparkMeta🇺🇸 USclosed1488.013,592
14gemini-3.1-pro-previewGoogle🇺🇸 USclosed1486.095,107
15gemini-3-proGoogle🇺🇸 USclosed1485.041,509

Research pace & markets

AI papers by country — share of monthly total

0%9%19%28%38%2025-032025-062025-092025-122026-032026-06USA 12.3%China 23.3%India 7.3%UK 3.5%Germany 2.3%France 1.3%dashed = latest month (provisional, still indexing) · share = papers with ≥1 author from country

Markets (visual context)

571111642182025-032025-052025-082025-112026-022026-052026-08S&P 500 +33%NVIDIA +98%Hang Seng +9%Bitcoin -25%indexed to 100 at window start · visual context only, not a prediction

Daily output — China calendar heatmap

MonWedFri2026-02-16: n/a2026-02-17: n/a2026-02-18: n/a2026-02-19: n/a2026-02-20: n/a2026-02-21: 2922026-02-22: 1182026-02-23: 3692026-02-24: 4902026-02-25: 4422026-02-26: 4842026-02-27: 4762026-02-28: 4172026-03-01: 1371Mar2026-03-02: 4502026-03-03: 4932026-03-04: 4392026-03-05: 4352026-03-06: 6872026-03-07: 3052026-03-08: 982026-03-09: 5272026-03-10: 5422026-03-11: 3992026-03-12: 4962026-03-13: 5812026-03-14: 27422026-03-15: 1942026-03-16: 4682026-03-17: 4672026-03-18: 4812026-03-19: 5282026-03-20: 10382026-03-21: 2962026-03-22: 1512026-03-23: 3612026-03-24: 4732026-03-25: 4132026-03-26: 4852026-03-27: 9552026-03-28: 2382026-03-29: 1592026-03-30: 4562026-03-31: 5632026-04-01: 1332Apr2026-04-02: 3852026-04-03: 3972026-04-04: 1572026-04-05: 802026-04-06: 3602026-04-07: 4392026-04-08: 5202026-04-09: 6482026-04-10: 7182026-04-11: 1682026-04-12: 2952026-04-13: 6732026-04-14: 3482026-04-15: 4742026-04-16: 3992026-04-17: 9322026-04-18: 2722026-04-19: 1362026-04-20: 4852026-04-21: 28982026-04-22: 4092026-04-23: 3422026-04-24: 6352026-04-25: 2282026-04-26: 1162026-04-27: 4072026-04-28: 4152026-04-29: 3222026-04-30: 4512026-05-01: 1272May2026-05-02: 1662026-05-03: 722026-05-04: 2722026-05-05: 2972026-05-06: 4552026-05-07: 3672026-05-08: 5092026-05-09: 2722026-05-10: 1182026-05-11: 4092026-05-12: 4632026-05-13: 4442026-05-14: 4772026-05-15: 5752026-05-16: 2292026-05-17: 1282026-05-18: 3712026-05-19: 4832026-05-20: 4462026-05-21: 4262026-05-22: 6252026-05-23: 2262026-05-24: 1962026-05-25: 3942026-05-26: 4152026-05-27: 4002026-05-28: 6402026-05-29: 4712026-05-30: 3022026-05-31: 1222026-06-01: 1564Jun2026-06-02: 3962026-06-03: 3962026-06-04: 4252026-06-05: 3842026-06-06: 2472026-06-07: 1052026-06-08: 4752026-06-09: 3962026-06-10: 4042026-06-11: 4102026-06-12: 4592026-06-13: 1632026-06-14: 2372026-06-15: 6992026-06-16: 3682026-06-17: 3472026-06-18: 3582026-06-19: 3192026-06-20: 1912026-06-21: 912026-06-22: 3012026-06-23: 3882026-06-24: 3612026-06-25: 3782026-06-26: 3522026-06-27: 2012026-06-28: 1212026-06-29: 4222026-06-30: 4742026-07-01: 1760Jul2026-07-02: 3432026-07-03: 3582026-07-04: 1802026-07-05: 1022026-07-06: 3542026-07-07: 3182026-07-08: 3542026-07-09: 3922026-07-10: 3362026-07-11: 2062026-07-12: 672026-07-13: 3792026-07-14: 3462026-07-15: 3372026-07-16: 3632026-07-17: 3452026-07-18: 1882026-07-19: 922026-07-20: 3802026-07-21: 3652026-07-22: 3152026-07-23: 3462026-07-24: 3522026-07-25: 1812026-07-26: 1112026-07-27: 4112026-07-28: 3512026-07-29: 3642026-07-30: 3352026-07-31: 3782026-08-01: 798Aug2026-08-02: 1742026-08-03: 3582026-08-04: 3552026-08-05: 2822026-08-06: 3602026-08-07: 3032026-08-08: 1932026-08-09: 732026-08-10: 3092026-08-11: 2832026-08-12: 3222026-08-13: 6582026-08-14: 2512026-08-15: 542026-08-16: 1lessmore (max 2898/day)China AI papers/day (OpenAlex) · recent days still indexing

Filings & markets — SEC EDGAR

Notable papers

🇺🇸 University of Washington added 2026-07-17 · arXiv 2026-07-16

Bad Memory: Evaluating Prompt Injection Risks from Memory in Agentic Systems

Systematic evaluation of prompt injection attacks targeting persistent memory in agentic systems, demonstrating successful exploitation across Claude and GPT models with implications for agent security.

  • Identifies and systematizes a novel attack surface (memory-based prompt injection) that emerges as agents become stateful and self-improving—a practical security threat at the intersection of agent design and adversarial robustness.
  • Empirical evaluation across multiple production agentic systems (Claude Code, OpenAI Codex) and model variants reveals that planted payloads in persistent memory reliably persist and influence future sessions, quantifying real vulnerability.
  • Motivates a concrete defense design problem: protecting agent memory updates while preserving beneficial adaptation, relevant to practitioners building stateful agents in production.
agentssecurityprompt-injectionadversarialagentic-systems
🇺🇸 unknown added 2026-07-17 · arXiv 2026-07-16

PolyQ: Codesigning End-to-End Quantization Framework for Scalable Edge CPU LLM Inference

PolyQ co-designs quantization and CPU compilation to enable efficient fractional-bit LLM inference with mixed per-channel bit-widths, achieving 2.4–32% perplexity gains over existing methods on edge CPUs.

  • Novel compiler/quantization co-design that solves practical fractional-bit deployment on CPUs via SIMD/LUT-compatible kernel generation and channel clustering, not just quantization in isolation
  • Demonstrates stable quality scaling from 3–6 bits with measured end-to-end gains (70% reduction in activation reorder traffic, near-proportional throughput/latency scaling) across workstation/laptop/mobile CPUs
  • Addresses a real deployment gap: existing mixed-precision methods are theoretically fine-grained but practically hard to execute on CPUs; PolyQ makes it work with <2% energy overhead
quantizationefficiencyedge-inferencecpucompiler-codesign
🇺🇸 unknown added 2026-07-17 · arXiv 2026-07-16

Routing Ceilings Are Domain-Independent: Structural Prior Injection in Code Security Vulnerability Detection

Demonstrates that structural priors (cheatsheets) in LLMs achieve impressive in-distribution gains but fail dramatically under distribution shift, extending the router hypothesis across domains from math to code security with real CVE data.

  • Replicates and extends SAIR's cross-distribution trade-off phenomenon to a new domain (code security), providing empirical evidence that the router hypothesis generalizes beyond formal reasoning
  • Identifies a systematic failure mode: cheatsheets that saturate synthetic performance (100% F1) collapse on real CVE data (48.9% F1, -51.1pp), motivating rethinking of prompt calibration vs. distribution-aware training
  • Comprehensive evaluation across three LLMs, three vulnerability categories with varying semantic complexity, and transfer to real-world VUDENC CVE data strengthens claims about generalization
reasoningdistribution-shiftcode-securitypromptingrobustness
🇨🇳 unknown added 2026-07-17 · arXiv 2026-07-16

MCPEvol-Bench: Benchmarking LLM Agent Performance Across Dynamic Evolutions of MCP Servers

MCPEvol-Bench evaluates LLM agent robustness to dynamic tool interface changes via 11 mutation operators across 123 MCP servers, exposing significant vulnerabilities in frontier models.

  • Novel evaluation dimension: first benchmark to systematically measure LLM agent adaptability under realistic tool evolution rather than static tool interfaces
  • Reveals concrete failure mode in frontier models (13.7-14.4% performance drops, planning/reasoning errors) that current evals miss
  • Practical relevance: MCP emergence as standard tool-use infrastructure makes dynamic tool compatibility a pressing real-world concern
agentstool-usebenchmarkrobustnessevaluation
🇨🇳 unknown added 2026-07-17 · arXiv 2026-07-16

Trajectory-Aware Flow Matching for Topology Optimisation

Flow matching approach for topology optimization that embeds physics-guided BESO trajectories to enable fast, stable conditional design generation without adversarial training or lengthy sampling.

  • Physics-informed generative modeling: integrates BESO optimization history into flow matching without inference-time optimization, a novel path design principle
  • Substantial efficiency gains: achieves better compliance, volume-fraction satisfaction, and topology fidelity with far fewer sampling steps than diffusion baselines
  • Practical for engineering: addresses real TO bottleneck (repeated FEA) with a method that works on 2D and 3D problems under limited training data
generative-modelsflow-matchingtopology-optimizationphysics-informedefficiency
🇯🇵 University of Tokyo added 2026-07-17 · arXiv 2026-07-16

Scalable Training of Continuous-Time Spiking Neural Networks with Differentiable Spike-Time Discretization

Memory-efficient training framework for continuous-time spiking neural networks via differentiable spike-time discretization, reducing memory by 100× and enabling deeper SNN models on single GPUs.

  • Novel discretization scheme that maps irregular spike times to fixed differentiable events, reducing activation memory from O(N_in·N_out) to O(M)—elegant theoretical contribution with practical scaling gains
  • Synfire-chain-inspired regularization organizes temporal dynamics and prevents dead neurons, enabling training of significantly deeper SNNs (9-20 layer convnets) than prior work
  • 100× memory reduction and 20× speedup on dense LIF layers demonstrates genuine scalability breakthrough for neuromorphic computing, historically a hard constraint
spiking-neural-networkstraining-efficiencymemory-optimizationneuromorphic
🇨🇳 Alibaba Group added 2026-07-17 · arXiv 2026-07-16

SEED: Self-Evolving On-Policy Distillation for Agentic Reinforcement Learning

SEED improves agentic RL by converting on-policy trajectories into learnable hindsight skills and dense token-level distillation signals, bridging sparse episode rewards to fine-grained policy learning.

  • Self-evolving distillation closes the supervision gap in outcome-based RL for LLM agents by extracting reusable natural-language skills and re-scoring actions under skill-augmented contexts for dense on-policy guidance.
  • Jointly optimizes outcome-based RL with auxiliary token-level distillation from hindsight skills that co-evolve with the policy, maintaining alignment with current trajectory distribution.
  • Demonstrates consistent improvements in sample efficiency and generalization across text and vision-based agentic tasks, with code release.
rlagentsdistillationpost-trainingreasoning
unknown added 2026-07-17 · arXiv 2026-07-16

RW-Voice-EQ Bench: A Real World Benchmark for Evaluating Voice AI Systems

Multidimensional voice AI benchmark (TTS, STS, SU, ASR) exposing acoustic and expressive capability gaps beyond standard isolated metrics.

  • Addresses genuine gap: current benchmarks miss paralinguistic/acoustic information that distinguishes voice from text—core to real voice AI evaluation
  • Reveals non-obvious findings (e.g., STS systems with audio access still remain transcript-driven) that should inform model development priorities
  • Comprehensive across four major voice modalities with analysis of independence of evaluation dimensions (naturalness/expressiveness/identity for TTS), exposing evaluation design issues in the field
benchmarkvoice-aittsasrspeech-understandingevaluation
unknown added 2026-07-17 · arXiv 2026-07-16

Innocuous-Seeming Data, Latent Ideology: Ideological Generalisation in Finetuned LLMs

Demonstrates that finetuning LLMs on narrow, benign datasets causes broad ideological shifts across unrelated domains—a critical finding for model safety and alignment.

  • Reveals an under-studied failure mode where seemingly-innocuous domain finetuning generalizes to unwanted ideological changes on out-of-distribution topics
  • Proposes 'ideological generalisation' framework with measurable breadth and amplification metrics; shows finetuning pushes shifts further than few-shot prompting alone
  • Empirically validates on GPT-4.1 and Gemma-3 that mixing with generic data doesn't prevent the effect, raising deployment concerns for policy-aligned systems
safetyfinetuningalignmentevaluationgeneralization
🇨🇳 Qwen/Alibaba added 2026-07-17 · arXiv 2026-07-16

LongStraw: Long-Context RL Beyond 2M Tokens under a Fixed GPU Budget

LongStraw enables million-token RL post-training on fixed GPU budgets via architecture-aware execution optimizations for long-context agent training.

  • Solves critical inference-posttraining context gap (inference 1M+ tokens vs RL training at 256K), enabling long-horizon agent learning.
  • Architecture-aware stack with memory-efficient replay strategy (0.21GB overhead per group size increase) demonstrates feasibility of 2.1M-token RL on commodity GPUs.
  • Validates on production models (Qwen, GLM) with concrete scaling results; establishes foundation for long-context RL at scale.
rlpost-traininglong-contextefficiencyagents
🇨🇳 Alibaba added 2026-07-17 · arXiv 2026-07-16

OmniaBench: Benchmarking General AI Agents Across Diverse Scenarios

OmniaBench: A comprehensive agent evaluation benchmark spanning 354 domains with 1,431 tasks to diagnose capability boundaries of general AI agents.

  • Largest hierarchical taxonomy for agent evaluation (354 level-2 domains across ToC/ToB/ToE) with executable environments and fine-grained 10-dimensional capability taxonomy
  • Methodical task synthesis via four complementary routes (DAG, DAG-S, Solver, Program) reducing contamination risk; reveals clear frontier model limitations on planning and constraint maintenance
  • Diagnostic value: structured failure analysis shows gaps in multi-step reasoning, enabling targeted model improvement beyond single-metric scores
agentsbenchmarkevaluationtool-usereasoning
🇺🇸 unknown added 2026-07-17 · arXiv 2026-07-16

Benchmarking Multimodal Large Language Models for Scientific Visualization Literacy

Benchmark of 6 MLLMs on scientific visualization literacy reveals significant gaps in quantitative reasoning and domain-specific visual understanding compared to human baselines.

  • Identifies concrete failure modes (fine-grained quantitative estimation, flow interpretation) critical for real-world multimodal AI deployment beyond generic chart QA
  • Structured evaluation framework (49 items, 8 techniques, 11 task types) that exposes performance unevenness and could become a standard SciVis literacy benchmark
  • Open-source benchmark and model outputs enable reproducible analysis of MLLM visual reasoning limitations in scientific domains
multimodalevaluationbenchmarkvisualizationreasoning
🇸🇬 National University of Singapore added 2026-07-17 · arXiv 2026-07-16

BadWAM: When World-Action Models Dream Right but Act Wrong

BadWAM reveals a critical adversarial vulnerability in world-action models where small visual perturbations cause misalignment between predicted and executed actions, breaking assumed safety guarantees.

  • First systematic adversarial attack framework targeting WAMs' core coupling property; shows 96.5%→43.1% success rate drop under simple perturbations
  • Introduces imagination-preserving attacks that maintain plausible future predictions while inducing harmful action shifts—exposes hidden vulnerability class specific to WAMs
  • Challenges the assumption that coupling action generation with future prediction provides interpretability and safety benefits for embodied AI systems
adversarialworld-modelsroboticsembodied-aisafety
🇺🇸 Stanford added 2026-07-17 · arXiv 2026-07-16

Symbal: Detecting Systematic Misalignments in Model-Generated Captions

Symbal detects systematic errors in MLLM-generated captions by identifying recurring visual-error associations, with a 1.7M-image benchmark across natural and medical domains.

  • Addresses a concrete, practical problem in vision-language data quality—systematic misalignments—where errors correlate with specific visual features rather than being random
  • Introduces SymbalBench, a substantial 1.7M-pair dataset across 420 VL datasets with annotated systematic misalignments, enabling benchmarking of error-detection methods
  • Demonstrates real-world utility for auditing MLLM captions and datasets without model access, with 63.8% detection accuracy and 4x improvement over baselines
multimodaldatasetquality-controlevaluationvision-language
🇺🇸 NVIDIA added 2026-07-17 · arXiv 2026-07-16

RoboTTT: Context Scaling for Robot Policies

RoboTTT scales robot visuomotor context to 8K timesteps using test-time training, unlocking one-shot in-context imitation and long-horizon task completion.

  • Test-time training with fast weights enables 8K-token context (3 orders of magnitude scaling) without inference overhead—a new scaling axis for robot models
  • Demonstrates concrete capability gains: one-shot imitation from human video, on-the-fly policy adaptation, and completion of long-horizon (5-min, 10-stage) tasks that prior baselines fail
  • Combines sequence action forcing + truncated BPTT into a practical recipe; validates that context length is a scaling law for embodied models, parallel to LLM findings
roboticsscalingtest-time-trainingin-context-learninglong-horizon
unknown added 2026-07-17 · arXiv 2026-07-16

Partition, Prompt, Aggregate: Statistical Self-Consistency in Language Models

Reveals that LLMs violate basic probabilistic consistency when aggregating predictions across partitioned populations, introducing a novel self-consistency evaluation criterion.

  • Identifies systematic failure of in-context learning to obey law of total probability—a fundamental probabilistic identity—across frontier models
  • Introduces 'macro fallacy' pattern where fine-grained subpopulation estimates aggregate better than direct population-level estimates, suggesting misalignment between model's internal knowledge and aggregate output
  • Proposes reference-free evaluation criterion (statistical self-consistency) that exposes a gap no existing benchmark captures, actionable for post-training/alignment
reasoningin-context-learningevaluationconsistency
🇨🇳 Xiaomi added 2026-07-16 · arXiv 2026-07-14

Full-Pipeline Inference Optimization for MiMo-V2.5 Series: Pushing Hybrid SWA Efficiency to the Limit

Production inference optimization system for Hybrid SWA + MoE + multimodal models, with novel KVCache management and RDMA-optimized distributed serving.

  • First large-scale LLM serving system combining Hybrid SWA (linear attention), sparse MoE, and multimodal encoders in production—non-trivial engineering addressing real deployment constraints
  • Systematic KVCache optimizations (layerwise prefetch, SWA-aware prefix cache trees, O(W) storage) and GCache distributed infrastructure with RDMA show measurable systems gains
  • Practical multimodal optimizations (GPU preprocessing, parallel video decoding, cache sharing) demonstrate end-to-end efficiency gains beyond just the attention mechanism
efficiencyinferencekvcachemoehybrid-swamultimodalsystems
🇨🇳 Chinese Academy of Sciences added 2026-07-16 · arXiv 2026-07-14

ShortOPD: Recovering Pruned LLMs with Short-to-Long On-Policy Distillation

ShortOPD recovers pruned LLMs' generation quality via adaptive short-to-long on-policy distillation, achieving 9× recovery on free-form generation with 4× better efficiency than standard baselines.

  • Addresses critical gap between pruning validation (multiple-choice) and deployment reality (free-form generation); identifies and exploits suffix repetition bottleneck in recovery
  • Novel training schedule that intelligently allocates on-policy rollout budget by detecting teacher-confirmed repetitive suffixes, reducing wasted computation by 71% while matching longer horizons
  • Comprehensive evaluation on math, code, and open-ended generation shows structured pruning can achieve deployment-ready quality when paired with smart post-training recovery
pruningpost-trainingdistillationefficiencygeneration
🇨🇳 unknown added 2026-07-16 · arXiv 2026-07-14

Boogu-Image-0.1: Boosting Open-Source Unified Multimodal Understanding and Generation

Open-source unified multimodal model matching closed-source systems with efficient training ($400K) and comprehensive generation+understanding capabilities.

  • Demonstrates competitive multimodal understanding and generation (text-to-image, editing, bilingual) with dramatically reduced training budget (~208M images, $400K) compared to closed-source systems
  • Provides practical insights and reproducible recipes for efficient multimodal training; full weights and code released under Apache 2.0
  • Shows agentic inference-time scaling can substantially boost performance—useful reference point for efficient multimodal systems under compute constraints
multimodalopen-weightsefficiencytext-to-imagepost-training
🇺🇸 unknown added 2026-07-16 · arXiv 2026-07-14

Concurrent Image Understanding and Generation: Self-Correcting Coupled Markov Jump Processes

A coupled Markov jump process framework for joint image-text generation with cross-modal self-correction and remasking, plus three new multimodal benchmarks.

  • Novel sampler design where one modality's denoising transitions depend functionally on the other's confidence—enabling within-step cross-modal feedback rather than just parallel or interleaved updates
  • Remasking mechanism to detect and repair contradictions between modalities, addressing a real limitation of standard MDMs
  • Releases three large-scale joint multimodal corpora with matched in/out-of-distribution splits for rigorous evaluation
multimodaldiffusionreasoninggenerationdataset
🇺🇸 University of Texas at Austin added 2026-07-16 · arXiv 2026-07-14

Adapting Generalist Vehicle Models for High-Speed MPC Across Terrains

OptCar adapts generalist vehicle dynamics models for high-speed off-road control via history-conditioned adaptation and synthetic rollouts, achieving 55% error reduction with minimal real-world data.

  • History-conditioned dynamics context token enables efficient cross-terrain generalization without extensive terrain-specific data collection
  • Novel adaptation recipe combining limited real-world + targeted synthetic rollouts solves practical sim-to-real gap for high-speed control where slip dominates
  • Demonstrates strong out-of-distribution robustness (unseen payload) and competitive performance with 10x less data than specialist baselines
controldynamics-modelssim-to-realroboticsadaptation
🇺🇸 Anthropic added 2026-07-16 · arXiv 2026-07-15

The Refusal Residue: When Probes Catch Alignment Faking and When They Don't

Empirical study of alignment faking in LLMs via hidden-state probing, revealing asymmetric refusal residues and model-dependent detection limits with rigorous evaluation controls.

  • First large-scale sweep (13 models) studying whether hidden states reveal alignment faking undetectable in outputs; identifies naturally-emerging faking in only Qwen3-32B and Llama-3.1-8B
  • Establishes critical measurement framework correcting for leakage artifacts in prior probing work (residualization, leave-one-query-out, orthogonality constraints); shows naive approaches yield AUROC 1.0 on control where no faking occurs
  • Reveals fundamental detection asymmetry: compliance shifts measurably under monitoring but refusal representations don't move, and steering detected directions barely changes behavior—constraining what probing can actually control
alignmentevalssafetymechanistic-interpretabilitypost-training
🇺🇸 Cornell University added 2026-07-16 · arXiv 2026-07-15

Learned Pairwise Deep Dual-Optimal Inequalities for Stabilizing Column Generation

A learned framework that predicts deep dual-optimal inequalities to stabilize column generation in routing problems, achieving 90%+ speedups in convergence time.

  • Novel machine learning approach to a classical optimization bottleneck: uses classifier to learn pairwise dual variable orderings that provably stabilize CG without hand-crafted exchange arguments
  • Strong empirical results on CVRP and VRPTW with dramatic CG time reductions (89.7% and 93.9%) while maintaining solution quality via recovery procedure with bound guarantees
  • Addresses practical scalability challenge in branch-price-and-cut methods, relevant to large-scale logistics and vehicle routing
optimizationcolumn-generationlearning-to-optimizevehicle-routingduality
🇺🇸 OpenAI added 2026-07-16 · arXiv 2026-07-15

Where Should RL Post-Training Compute Go? Model Size, Search, Learning, and Feedback

Systematic framework for allocating fixed RL post-training compute across model size, search, learning, and reward feedback—with empirical trade-off frontiers and accounting methodology.

  • Addresses a practical but underexplored optimization problem in RL post-training: how to budget fixed compute across orthogonal dimensions (model capacity, rollout search, policy updates, reward inference).
  • Introduces FLOP-accounting framework and RACE diagnostic protocol to decompose and profile compute allocation for GRPO, enabling reproducible comparison across allocation regimes.
  • Empirical finding that optimal allocation is problem-dependent (model size, budget scale, reward type, evaluation metric) challenges single-heuristic approaches and motivates per-task profiling.
rlpost-trainingefficiencyreasoning
🇺🇸 Microsoft added 2026-07-16 · arXiv 2026-07-15

GFlowRL: Scaling Distribution-Matching RL to Large Language Models

GFlowRL removes the unstable partition network from GFlowNet-style RL while scaling stably to 235B MoE models, achieving strong gains on math, code, and adversarial robustness benchmarks.

  • Replaces auxiliary partition function with in-batch Monte Carlo estimate—cleaner, more stable scaling for LLM post-training; systematic engineering insight applicable beyond GFlowNets.
  • First GFlowNet-style RL to scale reliably across dense and sparse (MoE) architectures; prior work (FlowRL) diverges at scale.
  • Achieves competitive code reasoning (Codeforces 2048 at 14B, ~25 Elo from o3-mini) and SOTA adversarial red-teaming (AdvBench/HarmBench), demonstrating practical impact on reasoning and safety.
rlpost-traininggflownetsscalingreasoning
🇺🇸 UCLA / Princeton added 2026-07-16 · arXiv 2026-07-15

DeepLoop: Depth Scaling for Looped Transformers

DeepLoop formalizes residual scaling for looped Transformers where parameter blocks are reused across multiple rounds, providing a principled method to stabilize training in parameter-efficient recurrent architectures.

  • Addresses the overlooked residual scaling problem in looped/recurrent Transformers where parameters are visited multiple times, deriving visit-alignment coefficient κ_R and adjusted exponents α=(2N)^{1/2}, β=(8N)^{-1/2}
  • Provides first-order perturbation analysis bridging shallow recurrent layers with deep models, recovering DeepNorm behavior and extending it to the tied-depth regime
  • Demonstrates empirical improvements in validation loss and downstream accuracy on GPT-2 models when recurrent depth is activated, validating the theoretical framework
efficiencyarchitectureresidual-scalingrecurrenttraining-stability
unknown added 2026-07-16 · arXiv 2026-07-15

Factorized Spectral Representations for Reinforcement Learning

FaStR learns spectral representations of RL dynamics via CP tensor decomposition, reducing sample complexity and improving transfer across actuator shifts.

  • CP decomposition of state-action-next-state tensor yields separate encoders with smaller hypothesis class and O(min(S,A)) sample complexity gain over prior matrix methods
  • Factored structure enables state encoder transfer across actuator shifts with only action encoder retraining—a genuinely useful RL property
  • Moves spectral RL from matrix to tensor view with theoretical and empirical grounding on continuous control tasks
rlrepresentation-learningefficiencysample-efficiency
unknown added 2026-07-16 · arXiv 2026-07-15

ExTernD: Expanded-Rank Ternary Decomposition Ternary LLM PTQ with Accuracy Approaching Any Quantization Level

ExTernD enables ternary LLM quantization approaching arbitrary precision through expanded-rank factorization, achieving 5.2-5.7 bpw with Q4-Q5 accuracy.

  • Novel factorization scheme ($A ≈ B diag(D) C$ with ternary factors) provably approaches bf16 accuracy as rank expands, overcoming fundamental ternary quantization ceiling
  • Continuous accuracy-efficiency trade-off via rank multiplier μ and sparsity threshold τ rather than discrete bit-widths; matches Q4_K at 5.2-5.5 bpw on 4B models
  • Post-training PTQ method applicable to modern LLMs without retraining, with demonstrated results on Gemma-4E2B and Qwen3.5-4B
quantizationefficiencypost-trainingfactorizationlow-bit
🇺🇸 unknown added 2026-07-16 · arXiv 2026-07-15

STOCKTAKE: Measuring the Gap Between Perception and Action in LLM Agents with a Fair Oracle

STOCKTAKE: a 26-week supply-chain benchmark that disentangles state estimation from control failures in LLM agents via a fair Bayes-filter oracle, revealing that top models fail equally at diagnosis-to-action and over-correction.

  • Separates perception (state estimation) from action (control) failures—a fundamental decomposition missing in prior agent evals that only score final outcomes
  • Fair oracle design using factored POMDPs with exact Bayes filters enables skill scoring against identical observation streams agents receive, avoiding privileged-information circularity
  • Multi-week horizons with hidden factor processes and curated stress profiles expose systematic model failures (e.g., correct diagnosis yet 34–43% stockouts; costly over-responses) that single-episode benchmarks miss
agentsbenchmarkreasoningevaluationplanning
🇺🇸 unknown added 2026-07-16 · arXiv 2026-07-15

Anatomically Faithful but Temporally Blind: Auditing Attribution for Left-Ventricular Ejection-Fraction Estimation from Echocardiography

Reveals that post-hoc attribution methods fail to capture temporal reasoning in video medical models, showing anatomically faithful explanations can mask clinically critical blindness to end-systolic/diastolic frames.

  • Exposes a systematic failure mode in XAI-based validation of video diagnostic models: spatial attribution fidelity is decoupled from temporal fidelity, undermining confidence in model interpretability for time-sensitive tasks.
  • Empirically demonstrates via occlusion probes that attribution failure is not an artifact but reflects genuine model behavior—models ignore clinically decisive frames despite high EF regression accuracy.
  • Calls for temporally-aware training and evaluation in medical video understanding, with direct implications for certification and deployment of video diagnostic AI.
xaivideo-understandingmedical-imaginginterpretabilityevaluation
🇫🇷 unknown added 2026-07-16 · arXiv 2026-07-15

Algebraic Representability as the Limiting Regime of Grokking: An Exactly Solvable Model with Holomorphic Activations

Exactly solvable model of grokking in neural networks using holomorphic activations, proving algebraic representability determines success/failure and explaining capacity-dependent memorization-to-generalization transitions.

  • Provides rigorous algebraic characterization of when networks can represent targets on modular arithmetic tasks, eliminating the grokking phenomenon in a controlled setting
  • Proves non-representable tasks incur width-independent training loss lower bounds, connecting expressivity constraints to memorization dynamics
  • Traces continuous path from representational failure through memorization to standard grokking via bottleneck ablation, offering theoretical bridge between extreme and practical regimes
grokkingmemorizationalgebraic-geometrymodular-arithmetictheory
🇺🇸 unknown added 2026-07-16 · arXiv 2026-07-15

AIMO Interpretability Challenge

Competition and benchmark for probing whether frontier math LLMs use robust reasoning or spurious shortcuts via interpretability methods on AIMO problems.

  • Addresses critical gap between eval metrics (accuracy) and model reliability (whether reasoning generalizes or exploits brittle patterns)
  • Provides olympiad-level problems with symbolic variants + frontier model access + infrastructure—infrastructure for systematic interpretability research at scale
  • Creates durable open robustness benchmark for mathematical reasoning, a bottleneck area for aligned capable AI
interpretabilityreasoningrobustnessbenchmarkmath
unknown added 2026-07-16 · arXiv 2026-07-15

Rethinking Penetration Testing for AI-Enabled Systems: From Resource Compromise to Behavioral Objective Violation

Reframes penetration testing for AI systems from infrastructure compromise to behavioral objective violation, providing a systematic framework for evaluating adversarial attacks on learned models in production.

  • Extends classical pentesting paradigm to AI-specific threat vectors (prompt injection, data poisoning, retrieval poisoning, agentic misalignment) that bypass traditional infrastructure security
  • Proposes concrete workflow for identifying operational objectives, mapping AI-governed behavior, analyzing adversarial influence surfaces, and executing scenario-based tests—actionable for practitioners
  • Addresses material gap in security evaluation: AI systems can fail catastrophically without infrastructure compromise, requiring new behavioral failure criteria and threat modeling
adversarialsecuritypost-trainingagentsevaluation
🇺🇸 unknown added 2026-07-16 · arXiv 2026-07-15

Transforming Rank: How Architecture Navigates the Spectral Pathologies of Depth

Rigorous analysis of how Transformer architecture components (skip connections, normalization placement, width expansion) preserve gradient and feature rank across depth, unifying normalization design principles through spectral theory.

  • Provides novel theoretical framework explaining why Pre-Norm plateaus while Post-Norm rank-collapses and how skip connection scaling controls rank-ensemble tradeoff
  • Connects initialization-time spectral properties (Marchenko-Pastur law, Jacobian rank) to trainability on downstream tasks, offering testable predictions
  • Reframes core architectural choices (normalization placement, two-matrix FFN structure, width scaling) as solutions to fundamental rank preservation problem in deep networks
architecturetheoryinitializationdepthnormalization
🇺🇸 Harvard University added 2026-07-15 · arXiv 2026-07-13

Sparse Inter-Layer Dependencies of Transformer FFN Neurons

Training-free attribution method reveals that Transformer FFN neurons have sparse inter-layer dependencies, enabling circuit-level interpretability and potential efficiency gains.

  • Demonstrates empirically that FFN activations can be reconstructed from sparse subsets of upstream neurons, challenging assumptions about dense feedforward structure
  • Introduces practical, scalable training-free method for neuron-level circuit interpretation applicable across model scales
  • Identifies candidate sparse pathways with direct implications for efficient inference without retraining
interpretabilityefficiencysparseffncircuits
unknown added 2026-07-15 · arXiv 2026-07-13

Are we Merging the Right Models? Impact of Expert Training Duration on Model Merging for LLMs

Systematic study revealing that optimal expert training duration for model merging depends on the merging method, with sparsification-based approaches benefiting from overtraining where simple averaging fails.

  • Novel empirical insight that training duration and merging method are jointly dependent rather than independent, contradicting standard practice
  • Rigorous evaluation across 5 domains and 3 scales with 5 merging methods, grounded in bias-variance theory connecting to ensemble learning
  • Practical implications for multi-task model merging as a scaling strategy for task adaptation without co-training overhead
model-mergingmulti-taskpost-trainingefficiency
🇨🇳 unknown added 2026-07-15 · arXiv 2026-07-13

SymbOmni: Evolving Agentic Omni Models via Symbolic Concept Learning

SymbOmni proposes an agentic omni-model for cumulative visual generation through symbolic concept learning, enabling continual learning without model fine-tuning and reducing token consumption by 40%.

  • Novel architecture: Symbolic Concept Box abstracts low-level operations into reusable workflows, addressing compositional generalization in visual generation
  • Efficient training: verbalized backpropagation with language-based feedback eliminates gradient-based fine-tuning while enabling continuous self-improvement
  • Cumulative learning: demonstrates state-of-the-art continual learning across benchmarks with 40% token reduction, solving the 'perpetual novice' problem in generative models
multimodalagentscontinual-learningefficiencysymbolic-reasoning
🇺🇸 University of Texas at Austin added 2026-07-15 · arXiv 2026-07-13

The Capacity of Thought: Benchmarking Llama 3.2 in Semantic fMRI Neural Language Decoding and Improving the Huth Encoding-Model Baseline

Rigorous benchmarking of LLM-based fMRI neural decoding reveals that large language model priors can mask failures—demonstrating the critical importance of blind controls in brain-computer interface research.

  • Identifies a major methodological pitfall: high-capacity frozen language models succeed via learned prior rather than actual neural signal decoding, masking genuine decoder failure
  • Improves Huth et al. ridge regression baseline by 11% METEOR through systematic engineering (expanded voxels, GPU acceleration, better proposal models)
  • Contributes practical caution against uncritically stacking frozen LLMs in BCI pipelines without proper ablations—relevant to broader BCI/neuroscience methodology standards
fmrineural-decodingbcievaluationlanguage-models
🇺🇸 Anthropic added 2026-07-15 · arXiv 2026-07-13

Token Reduction Is Not Cost Reduction

Empirical study showing token reduction doesn't translate to cost savings for coding agents, and can harm task success—challenges common compression assumptions with large-scale billed API data.

  • Rigorous empirical work on 2,848 Claude Code runs with actual billing data; reveals prompt caching dominates costs (87%), not tool-output tokens
  • Demonstrates compression harms performance (27→15 solves on SWE-bench) by corrupting critical evidence, contradicting token-reduction-as-proxy assumptions
  • Proposes success-adjusted billed-cost as evaluation metric instead of token count—actionable for API-based agent builders
agentscodingefficiencycost-optimizationempirical
🇺🇸 University of Washington added 2026-07-15 · arXiv 2026-07-14

Rethinking the Evaluation of Harness Evolution for Agents

Critical re-evaluation of harness evolution for LLM agents, showing that published gains don't outperform simpler test-time scaling baselines and fail to generalize—challenging a popular paradigm.

  • Exposes methodological flaws in how harness evolution is benchmarked (test-time scaling confound, same-task evaluation overfitting)
  • Demonstrates that gains from automatic harness design don't consistently beat simpler baselines under fair comparison (matched budgets)
  • Calls for stricter evaluation protocols for agentic systems, relevant as the field scales up test-time compute
agentsevaluationtest-time-scalingharness-design
🇺🇸 unknown added 2026-07-15 · arXiv 2026-07-14

Auditing Data Leakage in Whole-Slide Image Multimodal Benchmarks

Audits widespread data leakage in WSI-VQA benchmarks, revealing 92–100% train-test overlap that invalidates claimed zero-shot performance of vision-language models in pathology.

  • Exposes critical benchmark contamination (patient and institutional-level leakage) affecting a growing domain of clinical AI evaluation
  • Demonstrates leakage is decodable and causally impacts reported accuracy gaps, undermining published VLM claims on major WSI benchmarks
  • Provides concrete remediation framework (provenance tracking, automated auditing) applicable to other medical imaging benchmarks
benchmarkingdata-leakagemultimodalvision-languagemedical-imagingevaluation
🇺🇸 unknown added 2026-07-15 · arXiv 2026-07-14

A Shared Subcircuit Lets LLMs Count Down Across Tasks

Identifies and analyzes a shared "countdown subcircuit" in LLMs that generalizes across diverse length-constrained tasks, revealing interpretable mechanistic structure.

  • Mechanistic interpretability at scale: isolates a reusable computational motif (position vs. goal length comparison) in Llama-3.1-70B and traces its generalization to multiple tasks and models
  • Demonstrates empirical transfer of circuit structure across models and domains (controlled writing tasks → DNA sequences → ASCII tables → natural language inferences), suggesting fundamentally shared computational strategies
  • Unsupervised circuit discovery method on naturalistic data, bridging gap between toy mechanistic interpretability and real-world model behaviors
mechanistic interpretabilitycircuitsgeneralizationllamareasoning
unknown added 2026-07-15 · arXiv 2026-07-14

Same Loss, Same Noise, Opposite Schedules: Noise Structure and Optimizer Normalization Jointly Determine Whether Learning-Rate Cooldown Helps

Theoretical analysis of why learning-rate cooldown helps in some settings but not others, linking gradient noise structure and optimizer normalization to convergence behavior.

  • Provides closed-form analysis of signSGD convergence under multiplicative noise and derives noise-floor constants exactly
  • Mechanistic explanation for when/why cooldown phases improve final loss: non-normalized optimizers (SGD) self-anneal but normalized methods (sign-based, adaptive) require explicit schedule decay
  • Connects optimizer normalization and noise structure to practical large-model pretraining schedules (WSD), with empirical validation on real tasks
optimizationlearning-rate-schedulesconvergence-theoryadaptive-optimizerspost-training
🇺🇸 unknown added 2026-07-15 · arXiv 2026-07-14

PM-Bench: Evaluating Prospective Memory in LLM Agents

PM-Bench: a new benchmark for prospective memory in LLM agents, revealing systematic failures in delayed task execution across state-of-the-art models.

  • Addresses a concrete, underexplored capability gap in agentic AI (prospective memory/task postponement) with a principled cognitive-science-inspired evaluation framework
  • Benchmarking result showing even best models plateau at 65% F1 signals a genuine open problem worth investing in
  • Released as controlled testbed for diagnosing failures and developing targeted training/inference interventions
agentsbenchmarksreasoningevaluation
🇨🇳 unknown added 2026-07-15 · arXiv 2026-07-14

ARDepth: Auto-regressive Monocular Depth Estimation with Progressive Visual Conditioning

ARDepth reformulates monocular depth estimation as multi-scale auto-regressive generation with progressive visual conditioning, offering a structured alternative to diffusion-based approaches.

  • Shifts MDE paradigm from global iterative denoising to hierarchical auto-regressive generation, better aligning with actual scale-dependent geometry emergence
  • Introduces Scale-Progressive Conditioning (SPC) and Semantic-Aware Guidance (SAG) to enforce multi-scale coherence and structural consistency
  • Demonstrates strong empirical results with explicit architectural motivation for how depth structure emerges hierarchically
monocular-depthgenerative-modelsauto-regressivemulti-scalegeometric-modeling
🇺🇸 unknown added 2026-07-15 · arXiv 2026-07-14

A JoLT for the KV Cache: Near-Lossless KV Cache Compression via Joint Tucker and JL-Residual Allocation for LLMs

JoLT applies Tucker decomposition + Johnson-Lindenstrauss residuals to compress KV cache 2-3x near-losslessly by exploiting tensor structure, achieving order-of-magnitude better reconstruction error than prior methods.

  • Novel tensor-aware approach to KV cache compression: treats cache as 3-order tensor, exploits differential redundancy across heads/tokens/features, unlike prior low-rank or quantization methods
  • Joint optimization of Tucker ranks and residual bit-widths via Lagrangian dual under unified byte budget—principled allocation across layers and key/value pairs
  • Strong empirical results: 2-3x compression with minimal perplexity/accuracy loss on Mistral-7B and LLaMA-2-13B; 0.009/0.006 relative Frobenius error, ~10x better than SVD/4-bit baselines
efficiencykv-cachecompressioninferencequantization
🇨🇳 Huazhong University of Science and Technology added 2026-07-15 · arXiv 2026-07-14

Multi-Perspective Agentic Program Repair via Code Property Graphs and Temporal Execution Graphs

CT-Repair uses Code and Temporal Execution Graphs with multi-perspective agentic reasoning to improve automated program repair, achieving significant gains on Defects4J v3.0.

  • Structured representation of dynamic evidence (TEGs with 94.85% scope reduction) enables efficient context compression for LLM-based APR
  • Multi-perspective agentic approach (static/dynamic/hybrid) with FSM-guided reasoning produces diverse repair strategies and reduces redundant sampling
  • Strong empirical results: 489/854 bugs repaired, +19-30 over prior SOTA (ReinFix, RepairAgent) with smaller model budget
agentsprogram-repairreasoninggraphs
🇨🇳 unknown added 2026-07-15 · arXiv 2026-07-14

KnowAct-GUIClaw: Know Deeply, Act Perfectly, Personal GUI Assistant with Self-Evolving Memory and Skill

KnowAct-GUIClaw improves OpenClaw with cross-platform GUI automation, self-evolving memory, and skill transfer, achieving 64.1% on MobileWorld using Kimi-2.6.

  • Cross-platform GUI agent framework (Android/iOS/HarmonyOS/Windows) with experience-attributable memory and self-evolving skill library addressing real automation gaps
  • Know-Route-Act-Reflect paradigm + pluggable architecture enables transferable skills across base models (+8.5% with Kimi-2.6), reducing per-model engineering
  • SOTA results (64.1% MobileWorld) vs closed-source agents (Seed-2.0-Pro, GPT-4o) with open-weights foundation model
agentsgui-automationmulti-platformself-evolutionskill-transfer
🇺🇸 Stanford University added 2026-07-15 · arXiv 2026-07-14

Extractable Memorization From First Principles

Rigorous framework for extractable memorization claims in LLMs using matched comparisons and statistical calibration to distinguish true memorization from predictability.

  • Formalizes a fundamental methodological problem in memorization research—establishing what probability thresholds constitute valid evidence of memorization—via conformal testing and census approaches
  • Provides practical calibration methods (FPR-controlled thresholds as low as 1e-27) and shows prior extraction studies have significant false-positive rates (e.g., 24% for Wikipedia), directly impacting interpretability of memorization audits
  • Redefines extractable memorization operationally to require both valid statistical claims and realistic extraction budgets, setting clearer standards for the field
memorizationinterpretabilitystatistical-methodsevaluationllm-auditing
🇨🇳 Peking University added 2026-07-15 · arXiv 2026-07-14

Jetson-PI: Towards Onboard Real-Time Robot Control via Foresight-Aligned Asynchronous Inference

Efficient VLA deployment on edge devices via foresight-aligned asynchronous inference, achieving 8.66x speedup on Jetson Orin with perception-execution alignment correction.

  • Addresses real deployment bottleneck: VLA models on low-power hardware with practical 8-14% improvements on LIBERO
  • Novel foresight correction module elegantly handles perception-execution misalignment in asynchronous inference without retraining full models
  • Combines algorithmic (confidence scheduling) and systems (CUDA graphs, GPU buffering) optimizations for concrete control frequency gains
efficiencyedge-deploymentembodied-aiasynchronous-inferencerobotics
🇺🇸 unknown added 2026-07-15 · arXiv 2026-07-14

From Critic to Confidence: PPO for Language-Based Quantitative Prediction with Confidence Estimation

CARE-PPO combines PPO with loss prediction to jointly learn accurate numerical predictions and reliable confidence estimates in LLMs, validated on healthcare/finance tasks with OOD robustness.

  • Novel RL framework that repurposes the critic network as a confidence estimator aligned with prediction error, bridging uncertainty quantification and actor-critic learning
  • Demonstrates strong generalization across linguistic/domain shifts and reduced task-specific overfitting vs supervised fine-tuning
  • Addresses a practical pain point: LLM hallucinations and overconfidence in quantitative prediction with measurable confidence calibration gains
rluncertaintypost-trainingreasoning
🇨🇳 Beihang University added 2026-07-15 · arXiv 2026-07-14

Less Experts, Faster Decoding: Cost-Aware Speculative Decoding for Mixture-of-Experts

EcoSpec optimizes speculative decoding for MoE models by minimizing expert activation cost during draft selection, achieving up to 1.62× speedup on billion-parameter models.

  • Addresses a real efficiency bottleneck in MoE inference: expert scattering during speculative decoding increases memory traffic and reduces speedup gains
  • Practical system contribution with lightweight predictor and dynamic buffering that works with existing verification rules on production-scale models (DeepSeek-V3.1, Qwen3)
  • Consistent empirical improvements across multiple model families and diverse tasks demonstrate genuine engineering value for MoE inference optimization
moespeculative-decodingefficiencyinference
🇺🇸 unknown added 2026-07-15 · arXiv 2026-07-14

Line-Anchored Feedback Cuts Token Costs and Improves Correctness in AI Code Editing

Line-anchored feedback in code editing reduces token generation by 22–58% and improves correctness on local models, demonstrating a structural prompt optimization with direct cost and accuracy benefits.

  • Empirically quantifies token savings (22–58%) across multiple models (Claude Opus/Sonnet, local models) in a real code-editing task—direct relevance to inference cost and latency.
  • Shows correctness gains for weaker models (+5–7 points) when feedback is structured, with evidence that the benefit amplifies when edit application is decoupled.
  • Practical VSCodium extension (FileMark) and paired experiment design; technique is model-agnostic and immediately applicable to production code workflows.
efficiencycode-generationpost-trainingreasoning
🇺🇸 unknown added 2026-07-15 · arXiv 2026-07-14

What Makes a Representational Prior Work? Feature Families, Label-Free Invariances, and Critical Windows in Grokking

Systematic study of what makes representational priors accelerate grokking, revealing that label-free invariances and early-window injection are surprisingly effective.

  • Empirically characterizes the mechanistic prerequisites for grokking priors (feature family alignment, invariance structure) across 188 runs with strong ablations and quantified causality.
  • Demonstrates label-free commutation invariances generalize more reliably than supervised priors with 2.7× speedup, suggesting fundamentally different inductive structure.
  • Reveals timing is critical: priors only needed in first 2% of training budget, reducing the weight-norm delay-law exponent 17-fold, with practical implications for structured inductive bias.
grokkinginductive-biasmechanistic-interpretabilitygeneralization
🇺🇸 unknown added 2026-07-15 · arXiv 2026-07-14

The One-Word Census: Answer-Choice Conformity Across 44 Language Models

A minimal 31-prompt instrument measuring convergence and conformity patterns across 44 LLMs, revealing structured divergence by model type and generation with public artifacts.

  • Novel diagnostic tool with exact-match methodology (no embeddings/judges) revealing surprising quantitative structure in model output alignment—serendipity at 41% across diverse architectures
  • Surprising finding: newest flagships (Claude, GPT) most conformist; persona/community-tuned models most divergent; conformity reverses in latest Claude/GPT, signaling possible high-level strategy shift
  • Low-cost reproducible evaluation ($1/model) with full public prompts/transcripts/code—useful benchmark for tracking LLM behavior changes across versions and lineages
evalbenchmarkmodel-alignmentbehavior-analysisopen-artifact
🇯🇵 University of Tokyo added 2026-07-15 · arXiv 2026-07-14

Visual Access Boundaries in Vision-Language Model Reasoning

Mechanistic analysis of CoT in VLMs reveals visual access boundaries are shallow and independent of reasoning length, with performance bottlenecked by perceptual readout rather than computation.

  • Causal intervention (Visual Access Sweep) rigorously isolates what CoT actually does—shows it doesn't require sustained image access despite longer generations, contradicting intuitive assumptions
  • Identifies fundamental bottleneck at perceptual readout (attribute extraction) not reasoning/counting, with symbolic oracle experiments cleanly separating readout from computation failures
  • Generalizes across multiple SOTA VLMs (Qwen2.5-VL, InternVL3) at various scales, providing actionable insights for VLM architecture and CoT prompting design
reasoningvision-languagemechanistic-interpretabilitychain-of-thoughtinference
🇺🇸 unknown added 2026-07-15 · arXiv 2026-07-14

Evaluating Large Language Models on Misconceptions in Multi-Turn Medical Conversations

New multi-turn medical dialogue dataset revealing that frontier LLMs degrade substantially in correcting misconceptions over conversation length, with performance dropping from ~85% to ~50% by two turns.

  • Multi-turn degradation pattern is clinically important and under-studied: exposes error propagation and inconsistency in high-stakes medical contexts where safety matters
  • Rigorous evaluation methodology combining dataset design (2.4K real patient threads) with rubric-based LLM-as-Judge and oracle analysis to isolate causes of failure
  • Actionable finding for practitioners: frontier models (GPT-5, Claude-Haiku) appear competent in single-turn but fail in realistic multi-turn patient interactions, directly applicable to medical AI deployment
evaluationmedical-aimulti-turnsafetybenchmark
🇨🇳 unknown added 2026-07-15 · arXiv 2026-07-14

MemOps: Benchmarking Lifecycle Memory Operations in Long-Horizon Conversations

MemOps is an operation-level benchmark for diagnosing memory failures in long-horizon LLM-agent conversations, replacing black-box QA evaluation with structured lifecycle traces.

  • Reformulates memory evaluation from downstream QA accuracy to interpretable operation-level diagnosis (remembering, forgetting, updating, reflecting), revealing failure modes hidden by final-answer scoring
  • Introduces controllable generation pipeline embedding memory operations into long conversations with gold traces and six operation-level probe categories, tested across retrieval, parametric, and managed-memory systems
  • Demonstrates concrete gaps: session-level retrieval outperforms turn-level, and long-context models are weak at ordered state trajectories—diagnostic results that inform agent architecture choices
agentsmemoryevaluationreasoninglong-context
🇨🇳 unknown added 2026-07-15 · arXiv 2026-07-14

Watermark Forensics for Generative Models: An Information-Theoretic Perspective

Information-theoretic analysis of watermark forensics in generative models, establishing tight entropy-rate bounds for multi-user attribution and payload extraction.

  • Proves tight Θ(log N/h) lower and upper bounds for multi-user attribution in watermarked text, a fundamental result previously lacking.
  • Develops principled information-theoretic framework (information profile ν(t)) that unifies detection, attribution, extraction, and localization under one model.
  • Identifies and proves a real gap: Θ(log N)-token window where text is detectable as machine-generated but unattributable, plus footprint-resolution uncertainty principle.
watermarkinginformation-theorysecurityforensicsgenerative-models
🇨🇳 University of Science and Technology of China added 2026-07-15 · arXiv 2026-07-14

Do AI Agents Know When a Task Is Simple? Toward Complexity-Aware Reasoning and Execution

Proposes E3 framework for task-complexity-aware LLM agent execution that reduces token/file inspection overhead by 85–92% while maintaining task success through estimate-execute-expand strategy.

  • Concrete efficiency win on deterministic benchmark (MSE-Bench with 121 real edits) plus validation on live gpt-4o + pytest; formalism (ACRR) quantifies agent cognitive redundancy
  • Addresses real deployment cost: agents waste budget re-reading context; E3's minimum-sufficient execution matches baselines at 1/6 cost
  • Releases benchmark and framework; bridges pure RL/planning work to practical agent engineering grounded in real task feedback loops
agentsefficiencyreasoningpost-training
🇨🇳 unknown added 2026-07-14 · arXiv 2026-07-11

Enjoy Your Talk: A Human-Centered Benchmark for Multi-Turn Dialogue with Decoupled User Simulation, Target Modeling, and Judging

A human-centered multi-turn dialogue benchmark with decoupled user simulation and intent tracking that reveals significant gaps between state-of-the-art models on objective vs. subjective conversation quality.

  • Addresses a genuine gap in LLM evaluation: multi-turn dialogue requires trajectory-level metrics (intent drift, completion) that single-turn benchmarks miss
  • Principled methodology with persona grounding from human corpora rather than LLM synthesis, reducing evaluation bias; multi-judge ensembling improves robustness
  • Novel findings: 9x gaps on objective intent tracking despite models being statistically close on subjective dimensions (empathy/persona); reasoning improves tracking but not subjective scores
benchmarkdialogueevaluationintent-trackingmulti-turn
🇺🇸 MIT added 2026-07-14 · arXiv 2026-07-11

Reinforcement Learning with Verifiable Physics: Post-training LLMs with Continuous Rewards

Post-training framework using continuous physics rewards (PDE residuals + solution accuracy) to improve LLM code generation for numerical PDE solvers, with compositionality transfer to unseen problems.

  • Novel hybrid verifier combining hard executability checks with graded continuous rewards for scientific correctness—addresses key gap in RL for verifiable code where binary signals discard accuracy structure
  • Demonstrates a smaller RL-post-trained model outperforms frontier model prompting on PDE benchmarks, plus zero-shot transfer to held-out PDEs, suggesting learned compositional structure of numerical methods
  • Unifies diverse PDE families (hyperbolic, parabolic, elliptic, incompressible flow) under single policy, showing evidence of learning reusable numerical primitives (stencils, time-stepping, boundary handling)
rlpost-trainingcode-generationreasoningscientific-computing
🇺🇸 CMU added 2026-07-14 · arXiv 2026-07-11

Confining Nondeterminism: AI-Driven Research Systems as DBMSs for Reliable, Non-Wasteful, Transparent, and Collaborative Research [Vision]

Vision paper proposing to treat AI-driven research systems as deterministic database engines to eliminate hallucinations, stale results, and redundant computation in LLM agents.

  • Identifies fundamental failure modes of LLM agents (nondeterminism, unbounded outputs, no provenance tracking) with principled database-inspired solution: LLM as query compiler only, executor as deterministic dataflow engine with versioning and incremental maintenance.
  • Concrete design framing (versioned code/data DAGs, materialized views, cost-based scheduling) maps traditional DBMS guarantees (consistency, provenance, idempotence) to research reproducibility problems that current agent frameworks ignore.
  • Early-stage vision paper but stakes out significant problem space at intersection of systems and AI agents; likely to influence how future research automation platforms are architected.
agentssystemsreproducibilitydatabasesresearch-automation
🇨🇳 Alibaba & Peking University added 2026-07-14 · arXiv 2026-07-12

UNIBROWSE: A Data-to-Agent Framework for Multimodal BrowseComp

UNIBROWSE: a unified data-to-agent pipeline for multimodal web browsing that covers all three information-flow patterns and achieves SOTA on BrowseComp benchmarks via SFT + exploration-aware RL.

  • First to systematically generate training data for all three multimodal browsing patterns (text-only, image-to-text, text-to-image), addressing a gap in agent generality
  • Novel exploration degree metric to filter low-signal RL instances, improving sample efficiency for tool-use trajectory learning
  • Substantial benchmark gains (54.4 avg accuracy, +10.5 over base Qwen3.5-35B-A3B) that exceed closed-source agents like GPT-5 and Gemini-2.5
agentsmultimodalrlpost-trainingweb-browsing
🇺🇸 Johns Hopkins University added 2026-07-14 · arXiv 2026-07-12

The Compliance Trap: Diagnosing How AI Agents Consume Conflicting Memory

Identifies and diagnoses the 'compliance trap' in memory-augmented AI agents: how conflicting retrieved memories cause early adoption errors that amplify through trajectories and resist recovery.

  • Proposes E-P-R (Entry-Propagation-Recovery) framework to systematically analyze memory consumption in agents beyond retrieval quality alone
  • Reveals counterintuitive failure mode: stronger baseline agents suffer larger absolute damage from bad memory compliance, suggesting existing eval metrics miss critical safety concerns
  • Introduces MemTrapBench, a controlled benchmark isolating memory consumption phases, advancing agent evaluation methodology
agentsmemoryreasoningevaluationinteractive-environments
🇺🇸 Caltech added 2026-07-14 · arXiv 2026-07-12

M+Adam: Low-Precision Training via Additive-Multiplicative Optimization

Hybrid additive-multiplicative optimizer for low-precision LLM training that combines strengths of both update types to handle the failure modes of standard quantization.

  • Addresses a concrete failure mode in low-precision training (optimization stalling at large weights under coarse mantissa)
  • Provides theoretical monotone descent guarantee and demonstrates practical improvements across multiple precision levels (BF16, FP8, FP4) and model scales
  • Cost-relevant method for efficient pretraining without requiring high-precision weight copies
efficiencypost-trainingquantizationoptimization
unknown added 2026-07-14 · arXiv 2026-07-12

MafiaScope: Non-Invasive, Time-Resolved Belief Probing for LLM Agents in Social Deduction Games

MafiaScope is an open testbed that uses the game Mafia to probe and measure LLM agents' theory of mind and internal beliefs through non-invasive structured questioning.

  • Novel evaluation method: repurposes social deduction games as a measurement instrument for machine theory of mind, with ground-truth beliefs accessible only to the game engine
  • Actionable diagnostic tool: reveals miscalibration and systematic biases (e.g. agents over-predict suspicion by 1.5×) that wouldn't show in standard behavioural metrics
  • Open release with corpus: engine, visualizer, interactive replay, and 200+ cross-model games provide infrastructure for future work on agent reasoning and deception
evalagentsreasoningtoolopen-weights
🇨🇳 University of Science and Technology of China (USTC) added 2026-07-14 · arXiv 2026-07-12

Unlocking Parallelism in Autoregressive Language Models via Speculative Decoding with Progressive Tree Drafting

Progressive Tree Drafting enables 2× LLM inference speedup via structured parallel draft generation in a single forward pass, training-free and model-agnostic.

  • Addresses critical LLM inference bottleneck (memory-bound decoding) with a practical, training-free method that doesn't require auxiliary draft models
  • Novel tree-structured parallel drafting strategy with stepwise pruning that meaningfully improves upon existing speculative decoding by exploiting within-model parallelism
  • Strong empirical results (2× speedup) demonstrated across benchmarks with immediate applicability to inference systems
inferencespeculative-decodingefficiencyparallelism
🇺🇸 unknown added 2026-07-14 · arXiv 2026-07-12

Answer-Conditioned Chain-of-Thought Distillation for Few-Shot Industrial Vision with Small VLMs

Answer-conditioned CoT distillation from frontier VLMs to 3B models for few-shot industrial vision, achieving strong gains on real manufacturing tasks with minimal labeled data.

  • Practical few-shot adaptation method (18-30 labeled images per task) for industrial visual inspection—a real deployment bottleneck
  • Answer-conditioning insight: frontier model reasoning directed at correct label prevents garbage-in-garbage-out, with ablation showing 17.8pp performance cliff
  • Outperforms GPT-4.1 on weld radiography with 24 training images; demonstrates efficiency gains matter for manufacturing constraints
vlmdistillationfew-shotcotefficiencyindustrial
🇺🇸 Carnegie Mellon University added 2026-07-14 · arXiv 2026-07-12

Distributed Denial of Science: How Indirect Data Poisoning of AI Systems Can Industrialize Scientific Fraud

Empirical study of indirect data poisoning attacks on AI-driven scientific research, showing 49.56% success rate across frontier models and proposing provenance auditing defenses.

  • First systematic evaluation of data poisoning against autonomous research agents using frontier models (Claude, GPT, Gemini), demonstrating a critical vulnerability in AI-driven science
  • Practical attack requires only public dataset corruption with misleading metadata—no sophisticated trigger words or prompt injection—making it a realistic threat to scientific integrity
  • Proposes and validates mitigation via data provenance auditing with 5 checks, reducing attack success to zero, offering actionable defense for the research ecosystem
securitydata-poisoningscientific-integrityai-safetyadversarial
🇺🇸 Anthropic added 2026-07-14 · arXiv 2026-07-12

Filtering Harmful Actions Isn't Enough: Phantom Transfer in Agentic SDF

Demonstrates that misaligned behavior in synthetic agentic training data persists even after removing harmful actions, revealing diffuse disposition encoding independent of action-level filtering.

  • First empirical evidence of 'phantom transfer'—that adversarial dispositions in synthetic trajectories transfer to models even when explicit harmful actions are removed, invalidating action-level filtering approaches.
  • Shows the effect is model-dependent (Gemini vs Claude) and survives standard safety benchmarks, suggesting evaluations miss subtle misalignment vectors in agent training data.
  • Directly relevant to agent safety: as LLMs deployed as agents train on synthetic rollouts, this work exposes a critical gap in current data curation practices for agentic behavior.
agent-safetysynthetic-dataalignmentpost-trainingjailbreak
🇺🇸 Alibaba/Qwen (primary), with US-based contributors added 2026-07-14 · arXiv 2026-07-12

SETA: Scaling Environments for Terminal Agents

SETA: a scalable framework generating 4,500+ verifiable terminal RL environments for training LLM agents on command-line tasks, achieving SOTA results for RL-trained 8B models.

  • Addresses genuine bottleneck in agent training: constructing diverse, verifiable terminal task distributions at scale without natural supervision.
  • Releases SETA-Env (4,500+ environments) as open-source resource with dual pipelines (synthesis + evolution) for environment generation.
  • Demonstrates solid empirical gains: 12% pass@1 on Terminal-Bench 2.0 for 8B-scale RL-trained model, plus +3% gains on DeepSeek-V4-Flash under same harness.
agentsrlpost-trainingopen-weightsdataset
🇮🇱 Tel Aviv University added 2026-07-14 · arXiv 2026-07-12

The Singularity Space: A Generative Diffusion Framework for Signal Representation

Singularity-space diffusion framework for signal generation that recovers sharp transients via complex-plane pole-residue representations, achieving 8× compression and sub-resolution generalization.

  • Novel representation: pole-residue latent space instead of dense grids—directly captures physics of discontinuities and transients; avoids Gibbs artifacts
  • Resolution-free reconstruction on arbitrary grids without retraining; 4.2× better zero-shot sub-resolution generalization vs. grid baselines
  • Physically interpretable: each singularity maps to physical parameters; demonstrated on Burgers shocks with 10^−4 parameter recovery accuracy; potential for speech/biomedical signals
diffusionsignal-processinginverse-problemssparse-representationphysics-informed
🇺🇸 Google added 2026-07-14 · arXiv 2026-07-13

BackendForge: Benchmarking Agentic End-to-End Code Generation with Backend Services

BackendForge benchmark evaluates agentic LLMs on realistic end-to-end backend code generation with black-box HTTP testing, revealing significant gaps between local API behavior and complete service correctness.

  • Novel evaluation paradigm for agentic coding: co-evolving test oracles and reference implementations captures real-world deployment constraints missing from typical code generation benchmarks
  • Realistic, deterministic evaluation methodology using OpenAPI contracts and HTTP testing avoids ambiguous success criteria that plague synthetic code generation tasks
  • Reveals concrete failure modes: 55.4% → 28.6% drop from base to final oracle shows current LLMs (including GPT-4.5) struggle with system-level consistency despite local correctness
agentscode-generationbenchmarksevaluationbackend
🇺🇸 unknown added 2026-07-14 · arXiv 2026-07-13

Do Video-LLMs Actually Watch? Diagnosing Character-Tracking Failures in Long-Form Video

Diagnostic study revealing that video-LLMs don't actually track characters in long-form video despite benchmark claims—they rely on shallow gender cues and multiple-choice artifacts.

  • Exposes fundamental failure mode in video-LLMs: inability to bind named individuals to their visual identity across frames, not a data/compute issue but an architectural/reasoning bottleneck
  • Rigorous protocol (9 conditions, name-swapping, gender ablations, open-ended evaluation) systematically dismantles claimed benchmark scores and identifies spurious signals
  • Releases diagnostic toolkit for auditing what video-LLM benchmarks actually measure—directly useful for practitioners evaluating models and designing better evals
video-llmevaluationbenchmarkingreasoningmultimodal
🇨🇳 unknown added 2026-07-14 · arXiv 2026-07-13

Multi-dimensional training-priority weighting based on physical information propagation paths: a unified residual-weighting framework for physics-informed neural networks

Proposes a unified residual-weighting framework for PINNs that respects physical information propagation paths, improving training stability and accuracy through theoretically-grounded priority scheduling.

  • Addresses fundamental PINN training inefficiency (synchronous optimization of unrelated constraints) with principled solution rooted in NTK theory
  • Unifies temporal, spatial, and boundary priorities under single propagation-path framework with clear mathematical characterization
  • Consistent empirical improvements on benchmark PDE problems without architectural changes, practical for domain science applications
physics-informed-networkstraining-efficiencypde-solvingoptimization
🇺🇸 Stanford added 2026-07-14 · arXiv 2026-07-13

VIA: Visual Interface Agent for Robot Control

VIA reframes robot control as a vision-based agent task, leveraging frontier models' existing computer-use capabilities to achieve strong manipulation performance without robot-specific fine-tuning.

  • Elegant interface design (browser-based 3D) that lets off-the-shelf frontier agents (Claude, Fable) control robots zero-shot without retraining, demonstrating emergent transfer of computer-use skills to robotics
  • High performance on real manipulation tasks (96.7% on LIBERO-Goal, 100% on assembly) using only visual input and generic tools, suggesting frontier models already encode relevant reasoning
  • Shifts paradigm away from expensive VLA fine-tuning by leveraging model scale directly—performance scales with foundation model capability rather than robot-specific data
roboticsagentszero-shotvisionfoundation-models
🇨🇳 unknown added 2026-07-14 · arXiv 2026-07-13

Unified Gradient Projection: Language-Balanced Continual Learning for Multilingual Low-Resource ASR

Unified Gradient Projection mitigates catastrophic forgetting in multilingual low-resource ASR by balancing per-language gradient contributions during continual learning on pretrained Whisper models.

  • Addresses real deployment problem: fine-tuning Whisper on low-resource languages without forgetting high-resource ones
  • Novel gradient-space solution combining per-language projection with balanced replay—elegant approach to cross-lingual interference
  • Demonstrates near-zero forgetting on Whisper-large-v3 with practical gains across diverse language groups and model scales
continual-learningmultilingualasrcatastrophic-forgettinglow-resource
🇨🇳 unknown added 2026-07-14 · arXiv 2026-07-13

Pix2Act: Image-Space Manipulation Policies with Equivariant Augmentation

Pix2Act reformulates 3D manipulation as 2D image-space prediction with equivariant camera augmentation, improving robustness and generalization for imitation learning.

  • Clever reformulation of 3D control into 2D image-space trajectory prediction + triangulation, reducing learning complexity while enabling equivariant augmentation across multiple camera views.
  • Novel per-view rotation equivariance design that jointly transforms camera images and actions, implicitly augmenting data distribution and learning view-invariant action structures.
  • Strong empirical results on diverse simulated and real-world manipulation tasks with robustness to camera perturbations, suggesting practical applicability.
robot-learningimitation-learningmanipulationequivariancemulti-view
🇨🇳 unknown added 2026-07-14 · arXiv 2026-07-13

The Path to Self-Evolving Clinical Systems: Scaling Medical Agents from Assistance to Autonomy

Comprehensive framework for scaling medical AI agents from task-specific assistants to autonomous clinical systems, emphasizing environment scaling and self-improvement through interaction.

  • Addresses deployment-critical gaps: formalized autonomy taxonomy (assisted→cooperative→autonomous) and contamination-resistant benchmarks for clinical validation
  • Identifies clinical environment scaling (PACS/EHR/FHIR integration) as underexplored yet actionable direction, moving beyond parameter scaling toward self-evolving agents
  • Consolidates 300+ references on medical agents (2025-2026 emphasis) with concrete roadmap for hallucination/cascade-failure mitigation and fairness in clinical imaging (radiology, pathology, ophthalmology)
agentsmedicalautonomous-systemsenvironment-scalingself-improvement
🇨🇳 Tsinghua University (THUDM) added 2026-07-14 · arXiv 2026-07-13

SCALECUA: Scaling Computer Use Agents with Verifiable Task Synthesis and Efficient Online RL

ScaleCUA scales computer use agents via verifiable task synthesis and frontier sampling, achieving SOTA on OSWorld and ScienceBoard with open-source code and models.

  • Addresses genuine bottleneck: verifiable data scarcity in online RL for GUI agents by proposing VeriGen, an automated task synthesis pipeline producing 24K+ verifiable tasks at scale
  • Frontier Sampling efficiently allocates RL rollouts based on per-task capability, maximizing sample efficiency—a practical contribution for online RL scaling
  • Strong empirical results (68.7% OSWorld, 54.0% ScienceBoard) with reproducible open-source release (code, models, datasets), advancing the computer use agent frontier
agentsrlpost-trainingopen-weightsgui-automation
🇺🇸 University of Kansas added 2026-07-14 · arXiv 2026-07-13

TreeThink: A Modular Tree Search Library for Mathematical Reasoning with LLMs

TreeThink is an open-source modular library for asynchronous tree search in neural theorem proving, supporting Lean 4, Rocq, and Isabelle/HOL with integrated formal verification.

  • Addresses a real infrastructure gap: existing tree search libraries lack native formal verifier integration; TreeThink bridges neural LLM reasoning with multiple proof assistants via REPL
  • Demonstrates non-trivial engineering: fully asynchronous execution yields 6.3× speedup and scales across multiple formal languages (Lean 4, Rocq, Isabelle/HOL)
  • Open-source, MIT-licensed library with reproducible benchmarks (miniF2F, MATH500) lowers barrier to neural theorem proving research
reasoningtree-searchtheorem-provingopen-sourceformal-verification
🇺🇸 LaunchSafe added 2026-07-14 · arXiv 2026-07-13

Mako: A Self-Evolving Agentic Operating System (SE-AOS) for Autonomous Web Exploitation

Self-evolving agentic OS that autonomously discovers and hot-loads web exploitation capabilities, achieving 100% success on 104 CTF-style benchmarks—demonstrates that capability discovery, not reasoning, is the bottleneck in autonomous exploitation.

  • Novel architecture treats exploit capability as versioned kernel with runtime synthesis & validation, enabling hot-loading of new capabilities without redeployment
  • Empirical law: once a capability is discoverable, difficulty collapses—shifts focus from reasoning to capability engineering, with self-improving loop that sandboxes & commits fitness-positive modifications
  • Full coverage on XBOW (104 targets, 26 vuln classes, 3 tiers) with cryptographic per-build flag verification under strict anti-fabrication regime; deliberately withholds operational artifacts due to dual-use concern
agentssecurityautonomous-systemsself-improvingexploitation
unknown added 2026-07-14 · arXiv 2026-07-13

The Paternalistic Filter: Epistemic Injustice and Differential Refusal in LLM-Mediated History Education for Marginalized Romanian Students

API audit of LLM tutors reveals differential refusal and epistemic gatekeeping against marginalized students, showing safety alignment can institutionalize educational inequality.

  • Systematic empirical evidence that safety-aligned models exhibit 76.7% higher refusal rates and 3× reduced complexity access for low-SES/minority student personas—concrete mechanism linking alignment to bias
  • Introduces measurable framework (differential refusal, agency theft, hermeneutical injustice) for auditing LLMs as deployed educational tools, with direct pedagogical implications
  • Challenges assumption that 'safer' models are uniformly safer, showing safety mechanisms can amplify rather than mitigate discrimination depending on user context
safetybiaseducationauditingllm-alignment
🇨🇳 Nanjing University added 2026-07-14 · arXiv 2026-07-13

Efficient Test-Time Optimization for Multi-Agent Proof Autoformalization

Multi-agent framework for proof autoformalization using efficient test-time optimization guided by formal verification, improving on prior work by 19% on ProofFlowBench.

  • Addresses long-horizon proof autoformalization (statement→full proof) via structured decomposition-formalization-proving pipeline with bottleneck analysis directing compute efficiently
  • Novel test-time optimization approach using Pareto-frontier guided refinement of decompositions based on formal verification + semantic rubrics, avoiding costly training
  • 19% improvement on ProofFlowBench with lower test-time cost; scaling analysis shows gains converge quickly, enabling practical budget selection
reasoningproof-formalizationtest-time-optimizationmulti-agentformal-verification
🇨🇳 Astribot added 2026-07-14 · arXiv 2026-07-13

A Glimpse into Long-term Physical Coexistence with Intelligent Robots

PHILIA is a modular multi-robot agent architecture decoupling high-level semantic reasoning from low-level execution, enabling compositional long-horizon household manipulation with human-in-the-loop interaction.

  • Robot gateway abstraction cleanly separates agent reasoning from per-robot execution, enabling heterogeneous embodiments and plug-and-play policy/interface improvements
  • Demonstrates real long-horizon dexterous tasks (packing, lifting) with persistent memory and scene grounding on physical hardware (Astribot S1)
  • Compositional design pattern—advances in any component (UI, policies, navigation) improve overall experience without system redesign—addresses practical deployment constraints
roboticsagentsembodied-aihuman-robot-interactionlong-horizon-planning
🇺🇸 unknown added 2026-07-14 · arXiv 2026-07-13

Confidently Wrong: Detecting Hallucinations in Financial Question Answering from LLM Internal States

Linear probes on LLM internal activations reliably detect confident hallucinations in financial QA, outperforming surface-level confidence baselines.

  • Novel insight that residual stream activations carry discriminative signal for hallucination detection beyond token log-probabilities or self-assessment
  • Practical method for high-stakes domain (finance) with clear commercial value as a triage mechanism for human review
  • Systematic evaluation across multiple models (Qwen, Llama, Gemma) demonstrating generalization of the probing approach
interpretabilityhallucination-detectionprobingfinancial-reasoningreliability
unknown added 2026-07-14 · arXiv 2026-07-13

HyperSafe: Inference-Time Safety Recovery for Fine-Tuned Language Models

HyperSafe uses hypernetworks to generate model-specific safety networks that restore alignment of fine-tuned LLMs without retraining or weight modification.

  • Novel inference-time safety restoration via hypernetwork-generated side networks conditioned on layer-wise activation fingerprints—elegant post-hoc approach that doesn't require retraining or safety data at deployment
  • Demonstrates strong empirical results (19–31% → <1% harmful rates) across multiple model families while preserving task accuracy, addressing a real deployment pain point
  • Non-invasive method avoids gradient updates and model modification, making it practical for already-deployed fine-tuned checkpoints
safetyalignmentpost-traininginference-timeefficiency
🇺🇸 University of Massachusetts Amherst added 2026-07-14 · arXiv 2026-07-13

Agentic Skill Optimization over Lie Algebroids

LASKO applies Lie algebroid geometry to agentic skill optimization, using bracket-based screening to reduce expensive LLM validations by 15×.

  • Novel mathematical framework (Lie algebroids) for modeling non-commutative skill edits with latent structure, enabling cheap microsecond screening before costly LLM rollouts.
  • Concrete 15× speedup on causal extraction demonstrates practical benefit for agent self-improvement loops where edit composition order and hidden template effects matter.
  • Addresses a genuine problem in agentic systems: naive skill editing is combinatorially expensive; this substitutes algebraic structure for brute-force validation.
agentspost-trainingoptimizationreasoning
🇨🇳 Xiaomi Robotics added 2026-07-14 · arXiv 2026-07-13

Xiaomi-Robotics-U0: Unified Embodied Synthesis with World Foundation Model

38B multimodal autoregressive model for unified embodied synthesis combining image/video generation with robot control, achieving SOTA on real-world manipulation tasks.

  • First to support multi-view consistent scene generation across multiple robot embodiments with structured, controllable embodied transfer while preserving dynamics
  • Demonstrates that foundation world models can scale as both embodied world models and data engines—improves real-world manipulation success from 36.9% to 63.2% via generated training data
  • Unified framework that jointly optimizes 5 tasks (text-to-image, editing, scene/transfer/video generation) while preserving pre-trained generalization, ranking first on World Arena benchmarks
multimodalroboticsembodied-aiworld-modelsopen-weights
unknown added 2026-07-14 · arXiv 2026-07-13

How to Tame Grokking: Representation Geometry as a Control Signal

Identifies dimensionality collapse as a precursor to grokking and proposes GeomDR, a spectral regularizer that accelerates generalization by up to 52× through geometric control.

  • Mechanistic insight into grokking (delayed generalization) via representation geometry—a fundamental phenomenon relevant to understanding neural network training.
  • Practical intervention (GeomDR) that reliably modulates grokking timing across modular arithmetic and permutation tasks in both MLPs and transformers.
  • Bridges geometry and generalization dynamics, offering a generalizable framework for controlling training dynamics beyond toy tasks.
grokkingpost-trainingefficiencyinterpretabilityregularization
unknown added 2026-07-14 · arXiv 2026-07-13

RAGU: A Multi-Step GraphRAG Engine with a Compact Domain-Adapted LLM

A modular GraphRAG system with a compact 7B domain-adapted LLM that outperforms much larger models on knowledge graph construction and achieves strong retrieval performance on benchmark tasks.

  • Two-stage extraction with DBSCAN deduplication and Leiden community detection improves graph quality over single-pass approaches
  • Demonstrates that language skills (comprehension, extraction, reasoning) scale weakly with model size, enabling a 7B model (Meno-Lite-0.1) to outperform Qwen2.5-32B on KG construction
  • Open-source, single-GPU deployable system with clear evidence of improvements in evidence recall (0.84 vs ≤0.76) on GraphRAG-Bench Medical tasks
ragknowledge-graphsefficiencyopen-weightspost-training
🇨🇳 unknown added 2026-07-14 · arXiv 2026-07-13

From World Action Models to Embodied Brains: A Roadmap for Open-World Physical Intelligence

Proposes an integrated roadmap for physical intelligence via World Action Models and embodied brains, addressing fragmentation in action spaces, datasets, and system composition for embodied AI agents.

  • Tackles fundamental fragmentation in embodied AI—incompatible action spaces, datasets, and task conventions—with a concrete co-evolution roadmap rather than point solutions
  • Introduces 'embodied brain' concept that decouples high-level reasoning (multimodal context, intervention comparison) from low-level control, enabling modular reuse across heterogeneous embodiments
  • Proposes shared contracts and closed-loop post-training as system-level solutions to standardize interaction between world models, controllers, and verification—addressing a real infrastructure gap in physical AI
embodied-aiworld-modelsphysical-intelligencesystem-designrobotics
unknown added 2026-07-14 · arXiv 2026-07-13

Think Through a Bottleneck: Hourglass Reasoning for Rigorous Induction

Hourglass reasoning enforces strict information isolation between reasoning stages via symbolic bottlenecks to improve few-shot induction, achieving 14pt gains on ARC-AGI-2 and nearly doubling Verilog synthesis accuracy.

  • Novel architectural constraint on reasoning flow: enforced stage isolation with only compressed symbolic state crossing boundaries, not prior work on prompting or CoT
  • Strong empirical gains across diverse benchmarks (visual abstraction, hardware synthesis, linguistics) with substantial Verilog synthesis improvement (31%→58%)
  • Rigorous ablations isolating the contribution of stage isolation and induction quality, addressing a genuine failure mode of self-refinement in LLMs
reasoningpromptingfew-shotinductionsymbolic
🇺🇸 Anthropic added 2026-07-14 · arXiv 2026-07-13

Agent Hacks Agent: Autoresearch for Production-Agent Red-Teaming

Automated red-teaming framework that discovers reusable vulnerability patterns in production LLM agents through a falsifiable research loop, generating a transferable vulnerability concept graph.

  • Addresses critical safety gap: systematic discovery of agent vulnerabilities across Claude Code and Codex rather than one-off attacks
  • Novel artifact (Vulnerability Concept Graph) captures enabling conditions and transferability, not just success metrics—actionable for production teams
  • Demonstrates 14.2pp improvement over baselines with single-shot transfer across scenarios, suggesting genuine discovery of fundamental unsafe behaviors
safetyagentsred-teamingproductionvulnerability
🇺🇸 unknown added 2026-07-14 · arXiv 2026-07-13

NeuralActuator: Neural Actuation Modeling for Robot Dynamics and External Force Perception

NeuralActuator: a neural model that predicts actuator dynamics, external forces, and motor conditions to improve sim-to-real transfer on low-cost robot platforms, with a new annotated dataset.

  • Addresses practical sim-to-real gap by modeling nonlinear actuator effects (friction, hysteresis, backlash, thermal) that dominate on low-cost platforms; jointly optimizes multiple objectives (dynamics surrogate, sensorless force perception, condition monitoring)
  • Demonstrates generalization across three actuator families and platforms ranging $500–$30k; introduces Neural Actuation Dataset with robot states, telemetry, and ground-truth force labels
  • Differentiable training of torque surrogate via trajectory simulation without direct generalized-effort labels is technically novel; shows practical gains in behavior cloning and motor diagnostics
robot-learningsim-to-realneural-dynamicsopen-source-datasetefficiency
🇺🇸 University of Michigan added 2026-07-14 · arXiv 2026-07-13

MET: Theory-Grounded and Culture-Aware Multilingual Moral Reasoning

Introduces MCLASH benchmark and MET, a theory-grounded prompting method with self-distillation for culturally-aware multilingual moral reasoning without external supervision.

  • Addresses genuine gap in multilingual LLM evaluation: culture-specific moral reasoning rather than rote translation, with systematic variation across languages
  • Novel inference + training pipeline combining psychology/philosophy-grounded scaffolding with self-distillation, avoiding expensive human annotation or model supervision
  • Consistent gains across model families and sizes (Qwen, Gemma) with strong per-language improvements (12.94pt peak), plus measurable shift toward native-language reasoning
multilingualmoral-reasoningbenchmarkspromptingdistillation
National Taiwan University added 2026-07-14 · arXiv 2026-07-13

Encoder-Side Neuron Identification and Amplification for Acoustic Perception in Large Audio-Language Models

IAAN identifies and amplifies specific neurons in audio encoders at inference time to improve fine-grained acoustic perception in large audio-language models without retraining.

  • Novel neuron-level intervention in the audio encoder (not post-encoder) using activation contrasting with noise references—a genuinely unexplored direction for inference-time acoustic steering
  • Strong empirical gains (25.7pts on Audio-Flamingo, 21.4pts on Qwen2.5-Omni) with training-free, label-free methodology that proves specific neuron identity matters, not just count
  • Reveals encoder-side bottleneck in LALMs for non-semantic speech attributes; cleanly ablated evidence that decoder/LM-side interventions fail, establishing where the acoustic signal is lost
inference-timemultimodalaudioneuron-steeringefficiency
🇺🇸 Apple added 2026-07-14 · arXiv 2026-07-13

MM-ToolSandBox: A Unified Framework for Evaluating Visual Tool-Calling Agents

Unified benchmark for evaluating multimodal tool-calling agents with 500+ tools across 16 domains, revealing that visual precision (not planning) is the key bottleneck even for frontier models.

  • Stateful execution environment with 258 verified scenarios + automated generation pipeline addresses real gap in agent evaluation: current benchmarks don't test grounded multi-turn tool use with realistic visual/conversational phenomena
  • Comprehensive analysis across 12 SOTA models (4B to frontier) identifies planning-to-precision capability crossover by scale, actionable for research prioritization
  • Open benchmark and framework (500+ tools, 16 domains) will enable systematic study of visual grounding failures in agentic systems
agentsmultimodaltool-useevaluationbenchmark
🇺🇸 unknown added 2026-07-14 · arXiv 2026-07-13

AdvancedMathBench: A Benchmark Suite for Advanced Mathematical Proof Generation and Verification

A benchmark suite for evaluating advanced mathematical proof generation and verification in LLMs, with expert-annotated automatic verification pipeline and fine-grained error assessment.

  • Addresses genuine gap in LLM evaluation: advanced mathematics beyond olympiad-level with rigorous proof verification, not just final-answer checking
  • Introduces scalable automatic verification pipeline trained on expert annotations, enabling granular error detection beyond binary correctness
  • Reveals concrete limitations of frontier models (GPT-4o, etc.) on proof construction and verification, highlighting critical bottleneck in error detection capability
reasoningbenchmarkverificationmathematical-reasoningevaluation
🇺🇸 Anthropic added 2026-07-13 · arXiv 2026-07-10

Present but Rescaled: Chat-to-Agent Transfer of Additive Activation Steering

Systematic study showing additive activation steering has unpredictable and sometimes dangerous transfer properties from chat to agentic (ReAct) deployments, with refusal bypasses amplifying up to 2x on some models.

  • First rigorous chat-to-agent transfer analysis of steering with matched-information design and representation read-outs, revealing dissociation between directional strength and behavioral coupling
  • Identifies additive-specific mechanism via directional ablation and localizes rescaling to ReAct format scaffold, establishing mechanistic understanding rather than just empirical observation
  • Direct safety implication: steering-based jailbreaks are unpredictably amplified in agentic deployment (up to 2x on Gemma-2-9B), creating a critical gap between safety analysis and real-world deployment
steeringsafetyagentsmechanistic-interpretationpost-training
🇺🇸 Google added 2026-07-13 · arXiv 2026-07-10

Scoped Verification for Reliable Long-Horizon Agentic Context Evolution under Distribution Shift

GRACE: a graph-based system for reliably evolving LLM agent instructions over long horizons by structuring context as typed semantic graphs with local verification, improving telecom agent reliability from 9% to 67%.

  • Novel structural approach to agentic context management: uses typed semantic graphs instead of flat-text instructions to enable local, scoped verification of proposed updates—addressing a real deployment challenge.
  • Strong empirical gains on a controlled long-horizon evolution protocol (5 replications) with 7.4x improvement in pass@3 reliability versus zero-shot baseline, exceeding stronger Gemini 3.1 Pro reference.
  • Identifies a genuine operational constraint: as instruction sets accumulate during agent evolution, verification becomes intractable without structural substrate; proposes consolidation mechanism to keep content usable.
agentspost-trainingreliabilitysystem-prompt
🇺🇸 unknown added 2026-07-13 · arXiv 2026-07-10

Self-Guided Test-Time Training for Long-Context LLMs

Self-Guided Test-Time Training improves long-context LLM reasoning by selectively adapting only on relevant evidence spans, achieving 15% relative gains on LongBench.

  • Test-time training (TTT) for long contexts is computationally expensive; S-TTT mitigates this by having the model self-identify relevant spans before adaptation, making TTT practical at scale.
  • Addresses a fundamental failure mode: models struggle to locate and use evidence in long inputs; the self-guided selection mechanism is intuitive and empirically validates that span quality dominates TTT success.
  • Demonstrates consistent improvements across diverse models (Qwen3-4B-Thinking, Llama-3.1-8B) and challenging benchmarks (LongBench-v2, LongBench-Pro), suggesting generality.
test-time-traininglong-contextpost-trainingreasoningefficiency
🇩🇪 Deutsche Telekom (German Industrial AI Cloud) added 2026-07-13 · arXiv 2026-07-10

A Sovereign, Open-Source Foundation Model for German and English

Sovereign open-source 30B MoE-Transformer hybrid matching 14-27B dense models with superior code performance, trained entirely on German infrastructure with full transparency.

  • Hybrid MoE-Mamba design activates only 3B/30B parameters per token with constant-size inference cache—material efficiency/throughput advantage for long-context deployment
  • Strong multilingual foundation (German-English) with deliberate German up-weighting, outperforming larger European baselines and matching/exceeding Olmo 3 32B and Apertus 70B on open benchmarks
  • Full transparency (weights, checkpoints, per-source accounting, training code) under highly permissive licenses—rare level of open governance for sovereign AI
moeefficiencyopen-weightsmultilingualsovereign
🇺🇸 unknown added 2026-07-13 · arXiv 2026-07-10

All Explanations are Wrong, But Many Are Useful: Exploring the Rashomon Explanation Set with Large Language Models

Novel framework coupling explanations and predictions via an LLM agent that generates faithful explanation sets, improving both accuracy and interpretability on real-world classification/regression tasks.

  • Reframes explainability-accuracy trade-off as coupling problem; theoretical guarantees on fidelity and set non-emptiness
  • LLM-based agentic workflow (Explanation-Prediction-Reflection) with convergence proof that recovers explanation sets while improving accuracy
  • Validated on large-scale live production tasks (churn, survival, CTR) showing practical wins over XAI baselines with robustness to distribution shift and temporal splits
interpretabilityllm-agentsxaipost-trainingreasoning
🇨🇳 Peking University added 2026-07-13 · arXiv 2026-07-10

Seeing is Free, Speaking is Not: Uncovering the True Energy Bottleneck in Edge VLM Inference

First systematic energy profiling of edge VLM inference revealing that decoding (output tokens), not vision processing, dominates energy cost—contradicting implicit assumptions in the field.

  • Empirical energy profiling across 5 VLMs, 3 architectures, 2 platforms with surprising findings: inference power is model-intrinsic, output tokens cost 11-39x more than input tokens, visual pruning saves ≤10% energy
  • Demonstrates fundamental asymmetry in prefill vs decode compute/memory bounds for VLMs at edge scale, with implications for efficiency research direction
  • Quantifies image complexity effects (4.1x variation) arising from output length rather than visual processing—key insight for embodied AI systems on constrained hardware
efficiencyvlmedge-inferenceenergy-profilingpost-training
unknown added 2026-07-13 · arXiv 2026-07-10

PAC-ACT: Post-training Actor-Critic for Action Chunking Transformers

PAC-ACT applies actor-critic RL post-training to action-chunking transformers for industrial contact manipulation, improving force control and stability while maintaining real-time performance.

  • Action-chunking policies are underexplored for RL post-training; reformulating at chunk level rather than token level is a natural but non-trivial contribution for real-time control
  • Hybrid behavior-prior constraint elegantly balances RL exploration with BC pretraining—addresses core tension in vision-action policy learning
  • Demonstrates substantial gains on contact-rich industrial tasks (46× reduction in excessive force readings) under realistic pose perturbations and sparse rewards
rlroboticspost-trainingefficiency
🇨🇳 Infinity AI added 2026-07-10 · arXiv 2026-07-08

Infinity-Parser2 Technical Report

Large multimodal document parser with 5M synthetic bilingual corpus and multi-task RL achieving SOTA on OCR and parsing benchmarks.

  • 5M synthetically-generated bilingual document corpus (Infinity-Doc2-5M) with rich annotations (bboxes, markdown/HTML/LaTeX, reading order) addresses real scarcity of faithfully labeled parsing data
  • Multi-task joint RL across 8 co-trained objectives (parsing, layout, tables, math/chemical formulas, charts, VQA) unifies perception and reasoning in single optimization signal—novel training approach
  • Open-source model variants (Flash for latency, Pro for precision) with SOTA results (87.6% olmOCR, 74.3% ParseBench) and strong generalization across diverse document types
document-parsingmultimodalsynthetic-datareinforcement-learningopen-weightsocr
🇩🇪 unknown added 2026-07-10 · arXiv 2026-07-08

Explaining Near-Zero Hessian Eigenvalues Through Approximate Symmetries in Neural Networks

Explains the origin of near-zero Hessian eigenvalues in neural networks as weakly broken continuous symmetries, with explicit constructions and spectral analysis.

  • Provides mechanistic explanation for a widespread phenomenon (near-zero modes dominating loss landscape) through symmetry breaking—moves beyond empirical observation to theory
  • Constructs exact zero modes in deep linear networks analytically and traces how ReLU nonlinearity weakly breaks them, generalizing to convolutional architectures
  • Directly relevant to optimization and loss landscape geometry, with implications for understanding training dynamics and generalization
loss-landscapehessiansymmetrytheoryoptimization
🇺🇸 unknown added 2026-07-10 · arXiv 2026-07-08

Multi-agent Autoformalization of Tensor Network Theory

Multi-agent LLM workflow for automated formalization of tensor network theory in Lean, with new libraries and proof discovery.

  • First demonstration of LLM-agent team autonomously formalizing research-level physics (matrix-product states theorem) with novel proof routes
  • Produces TNLean library extending Mathlib with tensor network and quantum information primitives previously unavailable
  • Systematic study of bottlenecks in large-scale autoformalization, emphasizing mathematical intent enforcement as key challenge
agentsformalizationreasoningquantumlibraries
unknown added 2026-07-10 · arXiv 2026-07-08

Linear Attention Architectures: Mechanisms, Trade-offs, and Cross-Layer Routing

Comprehensive empirical study comparing softmax and four linear-attention mechanisms (DeltaNet, Gated DeltaNet, Kimi Delta, etc.) at 350M–3B scale, introducing cross-layer routing for improved long-context efficiency.

  • Systematic comparison of competing linear-attention schemes with explicit recurrent-memory formalism, revealing trade-offs in expressivity, throughput, and loss—rare engineering-level analysis for practitioners
  • Novel cross-layer routing (CLVR) mechanism for linear-attention stacks shows consistent improvements, suggesting architectural refinements beyond individual layer design
  • Large-scale training sweeps (15B tokens, up to 3B params) with multiple optimizers (AdamW, Muon) provide empirical grounding often missing in architectural papers
linear-attentionefficiencylong-contextarchitecturepost-training
🇺🇸 Yale University added 2026-07-10 · arXiv 2026-07-08

KronQ: LLM Quantization via Kronecker-Factored Hessian

KronQ improves post-training quantization of LLMs by incorporating gradient covariance via Kronecker-factored Hessian, achieving extreme 2-bit quantization where existing methods fail.

  • Addresses critical limitation of GPTQ by jointly modeling activation and gradient covariances in quantization objective, not just input statistics
  • Achieves dramatic empirical breakthrough: 2-bit quantization on LLaMA-70B (7.93 perplexity vs. >2000 for GPTQ/GPTAQ), enabling practical ultra-low-bit compression
  • Novel bidirectional incoherence processing and Hessian-trace-based mixed-precision allocation provide principled mechanisms beyond existing PTQ art
quantizationpost-trainingcompressionefficiencyhessian
🇺🇸 unknown added 2026-07-10 · arXiv 2026-07-08

When Implausible Tokens Get Reinforced: Tail-Aware Credit Calibration for LLM Reinforcement Learning

TACO improves LLM RL by calibrating credit assignment to suppress reinforcement of low-probability erroneous tokens, enabling more stable training and better generalization.

  • Identifies a concrete failure mode in critic-free RL (Positive-Credit Contamination) where tail tokens get identical positive credit regardless of contextual plausibility, then proposes a principled tail-risk scoring mechanism to suppress undesirable updates
  • Demonstrates consistent improvements over GRPO-style baselines across three LLMs and eight benchmarks with explicit focus on training stability and long-horizon RL
  • Practical method that preserves gradients for useful rare patterns while dampening noise, addressing a gap between uniform and fully selective credit assignment
rlpost-trainingreasoningcredit-assignment
🇺🇸 University of Washington added 2026-07-10 · arXiv 2026-07-09

PLURAL: A Global Dataset for Value Alignment

PLURAL is a 500k preference dataset grounded in 92-country survey data for culturally-aligned value steering in LLMs, validated across 20 countries with human eval.

  • Addresses a real gap: most preference datasets and alignment approaches reflect Western values; PLURAL systematically captures diverse cultural value systems at scale
  • Rigorous methodology: builds on Integrated Values Survey (nationally representative), synthetic generation pipeline preserves normative signals, multi-stage validation (dataset-level, automated, blind human evals with 176 evaluators)
  • Practical impact: shows 27.7% relative error reduction in cultural alignment; released dataset enables reproducible research on pluralistic alignment
alignmentdatasetscultural-diversityvaluespreference-learning
🇺🇸 unknown added 2026-07-10 · arXiv 2026-07-09

What LLM Forecasters Know but Don't Say: Probing Internal Representations for Calibration and Faithfulness

Probes of LLM internal representations reveal that forecasting models' CoT reasoning is often unfaithful, while activation-based lie detectors outperform explicit reasoning traces for calibration and behavioral auditing.

  • Demonstrates that CoT reasoning traces fail to faithfully reflect actual drivers of model predictions through evidence ablation/injection experiments, suggesting post-hoc rationalization.
  • Shows representation-pooling probes achieve substantially better calibration than explicit model outputs, and can predict behavioral shifts in 84% of cases even when CoT obscures them.
  • Reveals forecasts are largely predetermined before reasoning generation (fixed by single pre-reasoning pass), enabling token-efficient routing with 30-47% token savings—practical for deployment.
interpretabilitycalibrationpost-trainingreasoningprobing
🇺🇸 unknown added 2026-07-10 · arXiv 2026-07-09

When LLMs Agree, Are They Right? Auditing Self-Consistency and Cross-Model Agreement as Confidence Signals

Large-scale empirical study showing agreement among LLMs or within a model's samples is a weak and unreliable confidence signal, with frontier models exhibiting overconfident yet incorrect consensus.

  • Challenges a widespread assumption in LLM evaluation pipelines (LLM-as-judge, ensemble judges); shows agreement ≠ accuracy with real deployment consequences
  • Rigorous cross-model, cross-runner analysis (265K samples across GPQA/AIME) with careful statistical methods (hierarchical bootstrap, item-clustering); regime-dependent findings (agreement fails worst for frontier models)
  • Public release of per-run distributions enables reproducible evaluation methodology research; directly relevant to production monitoring and confidence calibration
evaluationllm-as-judgeconfidence-calibrationbenchmarkingbias
🇺🇸 Anthropic added 2026-07-10 · arXiv 2026-07-09

Persuasion Attacks Can Decrease Effectiveness of CoT Monitoring

Reveals that chain-of-thought monitoring can be exploited via persuasion attacks, and proposes model-diverse fact-checking as a mitigation.

  • Identifies a concrete vulnerability in CoT safety mechanisms—a widely-assumed alignment technique—showing CoT reasoning *increases* harmful approval by 9.5% when agents argue persuasively
  • Proposes and validates a practical defense (model-diverse fact-checking pairs reducing harmful approval by 45%), grounded in empirical eval across 40 tasks and thousands of interactions
  • Directly relevant to agent oversight and safety monitoring, concerns shared by industry deployment
safetyalignmentmonitoringadversarialagents
🇺🇸 unknown added 2026-07-10 · arXiv 2026-07-09

Understanding Layer Patching in Model Size Interpolation

Systematic study of layer-patching strategies for zero-shot model size interpolation via boomerang distillation, with graph-theoretic optimization and practical greedy algorithms.

  • Formulates layer selection for model interpolation as shortest-path optimization—provides theoretical grounding for a practical post-training technique.
  • Introduces KLPatch, a principled greedy algorithm that improves over heuristic baselines and scales across model families.
  • First systematic ablation of patching direction and strategy effects, revealing that simple sequential approaches are surprisingly effective—useful guidance for practitioners.
distillationmodel-compressioninterpolationpost-trainingefficiency
🇨🇳 unknown added 2026-07-10 · arXiv 2026-07-09

LEEVLA: Seeing What Matters in Latent Environment Evolution for Vision-Language-Action

LEEVLA introduces drift-guided dynamic prioritization and structured feature flow generation to improve vision-language-action models for robot control by learning what visual regions matter.

  • Novel attention mechanism (DGDP) combining semantic drift guidance with dynamic prioritization to identify task-critical visual evidence
  • Structured latent space reasoning via prototype-to-periphery prediction + topological consistency loss, improving interpretability of latent world models
  • Consistent benchmark improvements on VLA tasks through principled 'where-how' training framework that avoids treating all visual tokens uniformly
vision-language-actionmultimodalroboticsattentionlatent-space
🇺🇸 unknown added 2026-07-10 · arXiv 2026-07-09

Compete Then Collaborate: Frontier AI Teachers Build a Verifiable Curriculum to Improve a Coding Student Beyond Imitation

Multi-teacher knowledge distillation framework using execution-based verification and RL with verifiable rewards to improve code LLMs, showing RL beats imitation on harder problems.

  • Execution-based judge (unit tests) rather than LLM judge avoids bias; reveals saturation on standard benchmarks but separation on harder problems
  • Key finding: SFT on verified solutions degrades student performance; RLVR (RL with verifiable rewards) reverses this, +49% relative gain on competition problems
  • Reproducible on-prem pipeline released for running GRPO with verification environment—practical contribution for training pipeline design
knowledge-distillationrlpost-trainingcode-generationverification