Claude Code v2.1.181 introduces support for setting config settings via prompt syntax like /config thinking=false, adds sandbox Apple Events support on macOS, and improves streaming, auto-retry, and subagent behavior. It also fixes numerous bugs related to startup, file handling, clipboard, and UI responsiveness across platforms.
A new dataset, IFLLM, collects mouse trajectories and eye gazing data from users interacting with LLMs. It shows that implicit feedback significantly improves LLM alignment, boosting text-based reward model accuracy from 55% to 64% and nearly tripling response quality improvements after DPO training on eight LLMs.
H-RePlan introduces a hierarchical replanning framework that separates device-local strategy recovery from global orchestrator replanning. It outperforms existing baselines by achieving higher completion and instruction adherence, with reduced token cost, through scope-aware recovery in multi-device agent systems.
UAV-VLN-FOV isolates the see-and-reach stage for precise evaluation of UAV navigation. 3DG-VLN enhances visual grounding and spatial alignment using dynamic 3D direction cues, achieving a 13.82% success rate improvement over baselines and validated in real-world trials.
This work shows that Lean can serve as a symbolic process oracle, providing fine-grained, verified feedback during reinforcement learning. By parsing proof attempts into tactic sequences and using Lean's elaboration to mark sound steps and first failures, the system generates dense, type-theoretic reward signals. Experiments demonstrate tactic-level supervision outperforms outcome-only methods on benchmarks like MiniF2F and ProofNet, highlighting Lean's role as both evaluator and training reward source.
MACR introduces a multi-agent reasoning framework to resolve knowledge conflicts in LLM inference by jointly assessing internal and external knowledge. It uses semantic entropy to measure confidence and employs three specialized agents to induce rules, detect conflicts, and resolve inconsistencies across contexts. Empirical results show MACR outperforms state-of-the-art methods and provides interpretable conflict resolutions.
CRAX introduces a high-fidelity, accelerated safety benchmark for reinforcement learning using MuJoCo XLA. It achieves up to 100x speedups over CPU-based benchmarks via vectorization and hardware acceleration, featuring six environment suites and three agent-specific tasks across three difficulty levels. Evaluation of six safe RL methods shows no single approach dominates, highlighting trade-offs between performance and safety, with curriculum learning and safety transfer improving results.
StreamKL introduces a fused GPU primitive that eliminates quadratic memory usage in attention distillation by streaming query-key tiles through on-chip SRAM. It achieves up to 43x speedup in forward and 14x in backward passes, reducing extra HBM footprint from O(N_QN_K) to O(1), enabling long-context distillation on a single GPU.
VIMPO introduces a critic-free policy optimization method that derives a policy-implied value function from KL-regularized reinforcement learning. It enables verifiable reward incorporation without training a critic and outperforms GRPO on mathematical benchmarks, especially under noisy rewards.
Benchmarks using fixed-shape checks miss real bugs in LLM-generated GPU kernels. A controlled corpus of 24 kernels, including 9 buggy variants with transcription errors, reveals that an op-schema-aware oracle detects all failures and passes all correct controls, with identical results across five GPU architectures.
CRAX introduces a high-fidelity, fast safety benchmark for reinforcement learning using MuJoCo XLA. It achieves up to 100x speedups over CPU-based benchmarks via vectorization and hardware acceleration, featuring six environment suites and three agent-specific tasks across three difficulty levels. Evaluation of six safe RL methods shows no single approach dominates, highlighting trade-offs between performance and safety, with curriculum learning and safety transfer improving results.
A study evaluates four AI review systems across six language models, finding OpenAIReview with GPT-5.5 achieves 83.0% accuracy in matching paper quality to external signals and detects 71.6% of injected errors. Real user feedback shows positive sentiment, with a 1.44-to-1 vote ratio, though false positives and minor nitpicks remain common.
Sevra, a serving-layer controller, selectively verifies answers to improve accuracy and reduce token usage. On \mathfive, it achieves 76.3% accuracy with 26.8% fewer post-generation tokens and halved harmful flips, while on \gsm it verifies only 3.0% of examples, boosting accuracy to 94.5% and cutting verification tokens by 91.2%. The study shows that initial solve length and explicit control needs determine optimal verification strategy.
JAMER introduces JamSet and JamBench, the first project-level game code dataset and benchmark on a professional game engine. Built from 8,133 verified Game Jam projects, it enables deterministic evaluation and reveals a capability cliff in AI models as project scale increases, with runtime pass rates dropping from 80.4% to 5.7%.
A new framework defines when single-neuron interventions coherently control model behaviors without output collapse. The control window, based on alignment and norm ratios, predicts behavior triggers and collapse ceilings using forward pass data, with high accuracy on held-out neurons. On refusal, control is typed: coherent bypass occurs without actionable content, while genuine actionable reach appears only in specific cases and at later rollout stages.
AtomMem introduces a memory system that stores high-value atomic facts from long-form interactions. It uses hierarchical event structures and temporal profiles to capture coherent episodic contexts and track evolving user attributes, enabling stable and efficient memory evolution. Experiments on the LoCoMo benchmark show AtomMem achieves state-of-the-art performance in reasoning tasks.
REDACT introduces a systematically controlled multilingual benchmark for personally identifiable information detection, featuring 51 entity types, 4,127 surface-form patterns, and 25 languages. It evaluates five detectors across 1,000 records, revealing that rule-based models fail on high-stakes data while LLMs perform better, especially in high-sensitivity categories. A reference-free LLM assessment confirms sensitivity-tier assignment as the most challenging evaluation axis.
GEMS enables training-free superposition of multiple semantic directions in LLMs by addressing distributional deviation and directional interference through geometric constraints. On GSM8K, it maintains 98% accuracy with three non-mathematical directions, while unconstrained addition drops to 4%; on Wikitext-2, it increases PPL by only 2.2%.
LLM agents commonly select higher-privilege tools despite sufficient lower-privilege alternatives. This over-privileged behavior is amplified by transient tool failures and does not reliably improve with general safety alignment. A new privilege-aware post-training defense reduces unnecessary high-privilege tool use while maintaining agent capabilities.
STAGE is a pipeline that generates text-to-JSON training data by using LLMs to synthesize reports and JSON schemas, validated against underlying spreadsheets. Evaluations on STAGE-Eval show it improves Qwen3-4B exact match from 31.37% to 74.27% and value accuracy from 45.46% to 90.69%.