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.
PragReST is a self-supervised framework that enhances large language models' pragmatic reasoning by generating counterfactual reasoning traces and training via supervised fine-tuning and reinforcement learning. It outperforms baseline models on four pragmatic benchmarks, improving Qwen3-8B and Qwen3-14B by 5.37% and 5-5.50% accuracy respectively, and maintains strong performance on general-knowledge and mathematical reasoning tasks.
A new study introduces VETO, a benchmark of 2,032 BBQ-derived contrastive pairs, to quantify misfired alignment in large language models. It defines the Misfired Alignment Rate (MAR) and finds that all benchmarked LLMs exhibit MARs between 4.7% and 18.9%, while human participants achieve 0%. The research shows alignment cues can amplify these failures, with evidence suppression occurring in late layers of models and emerging after instruction training.
A new method called output vector editing minimally modifies MLP neurons' output vectors to suppress memorized sequences in large language models, achieving up to 87.9% suppression in OLMo-7B. This approach outperforms zeroing neuron activations by a factor of 2.7 and works across four models from 36-7B parameters, with success rates scaling with model size and showing consistent performance across architectures.
SenFlow introduces a novel method for detecting AI-generated text in hybrid documents by modeling inter-sentence dependencies. It achieves state-of-the-art performance on MOSAIC, a benchmark of 16,000 documents from PubMed and XSum, with a +4.15 pp Macro-F1 gain on cross-domain transfer. SenFlow reveals that AI-generated content still exhibits generator-dependent sentence-length patterns, exploitable by sentence-level detectors despite perplexity filtering.
Decoupled Search Grounding (DSG) separates search functionality from reasoning models, enabling vendor-agnostic, tunable, and reusable search grounding. DSG achieves near-native accuracy on SimpleQA with 91% lower search cost and 99.4% warm-cache hit rate, while reducing latency by 68% and preserving concise output contracts.
SC-GRPO uses per-token KL divergence from self-conditioned trajectories to weight gradients in reinforcement learning. It outperforms GRPO by 8.1% and DAPO by 5.9% across math, code, and agentic tasks, with superior out-of-distribution performance and better results than OPD.
WorldLines introduces a project-driven benchmark for long-horizon embodied household assistance, capturing extended household traces with dialogues, actions, and state changes. It enables evidence-linked samples for Memory QA and Embodied Task Planning, and proposes ObsMem, an observer-grounded memory framework that supports visibility-aware memories and state-aware decisions. Experiments highlight challenges in partial observability and memory translation, with ObsMem providing a stronger reference architecture for such settings.
SGCD introduces an iterative framework to improve GUI agents by addressing supervision gaps in off-trajectory states. It extracts skills from both successful and failed rollouts, using them to guide policy continuations that are mixed with expert trajectories. On OSWorld-Verified, SGCD boosts success rates of three base models from low-30\% to over 50\%.
Decoupled Search Grounding (DSG) separates search functionality from reasoning models, enabling vendor-agnostic, tunable, and reusable search grounding. DSG achieves near-native accuracy on SimpleQA with 91% lower search cost and 99.4% warm-cache hit rate, while reducing latency by 68% and preserving concise output contracts.
TRAP evaluates how well models complete tasks using private data without leaking it. Across 22 models, all show non-trivial privacy leakage, with instruction-following ability linked to higher leakage. Structural private field isolation prevents leakage by replacing private fields with hash keys, maintaining task accuracy without sacrificing privacy.
Spotlight enables DiT RL post-training by leveraging idle spot GPUs, reducing costs by 1.4-6.4× while achieving superior image quality. It uses stale model weights in exploration and reconfigures sequence parallelism in real time, allowing efficient GPU utilization without breaking training pipelines.
AdsMind is a closed-loop multi-agent system that uses machine learning force fields and feedback to correct errors in adsorption configuration searches on catalyst surfaces. It achieves 100% and 98.8% success rates on AA20 and OCD-GMAE62 benchmarks, reduces energy dispersion by 14-fold compared to baselines, and maintains correct adsorption-energy signs in DFT validation, outperforming open-loop LLM agents.
NoiseTilt introduces NTRK, a reward-guided diffusion sampler that injects reward gradients via the noise term without altering the reverse kernel. By using a whitening operator, NTRK safely biases noise toward high reward, preserving sample quality while maintaining strong guidance. On aesthetic generation, NTRK achieves superior reward performance with 25 NFEs, reducing compute by 20× compared to state-of-the-art baselines.
Handlebars' triple-brace interpolation fails to protect against structural role injection, as HTML escaping only neutralizes angle-bracket delimiters. It leaves colon and Markdown hash delimiters intact, enabling attackers to hijack model behavior. The default escaping provides no protection for most role delimiter schemes and cannot replace a clear separation of instructions and data.
Handlebars' triple-brace interpolation fails to protect against structural role injection, as HTML escaping only neutralizes angle-bracket delimiters. It leaves colon and Markdown hash delimiters intact, enabling attackers to hijack model turns. The default escaping provides no protection for most role delimiter families and cannot replace a structural separation of instructions and data.
A study reveals that even neutral prompts trigger region-specific responses in large language models due to user metadata. Location leakage increases by up to 793 times in some models, and using 'Unknown' instead of location metadata still causes significant bias, indicating the user profile frame itself acts as a conditioning signal.
TAC, the first agentic benchmark for implicit animal welfare, tests AI agents' ability to avoid animal exploitation in travel booking scenarios. All seven frontier models score below 64%, with the best at 53%, and even minor prompt improvements yield only modest gains. An audit finds no signs of evaluation awareness, indicating performance gaps stem from lack of true welfare reasoning, not prompt recognition.
A red-team study of Anthropic's Fable 5 and Opus 4.8 models reveals both are vulnerable to adaptive iterative attacks, with Opus 4.8 breached on 11.5% of intents and Fable 5 on 6.1%. Despite robust defenses, both models generated 1,620 and 702 panel-confirmed harmful completions across all harm categories, automatically and efficiently under automated attack.
RubricsTree introduces a hierarchical taxonomy of over 100 clinically-verifiable Boolean rubrics, evolved from 4,000 real user queries via human-in-the-loop curation. It enables scalable, expert-aligned evaluation of personal health agents by dynamically routing queries to relevant rubrics and outperforms baseline methods in alignment, context sensitivity, and model performance gains of up to 66% on HealthBench.