A new study introduces ACE, an evaluation framework that controls for accuracy differences when comparing LLM calibration, revealing that many previously observed calibration advantages vanish and rankings can reverse.

This paper proposes a framework using generative AI agents as behavioral engines for black-box auditing of personalization algorithms. In a case study on X after the 2024 U.S. election with 1,120 agents, they find that the algorithmic feed amplifies toxic, polarizing, political, and right-leaning content compared to the chronological feed, with amplification varying by user ideology. Counterfactual analysis shows demographic signals affect content delivery in persona-dependent ways.

The Indi-RomCoM benchmark covers seven instruction-following tasks, four Indic languages, and three code-mixing intensity levels to systematically evaluate LLMs on Romanized Code-Mixed instructions. Results show LLMs consistently underperform, with performance degrading as code-mixing density increases; reasoning tasks degrade less than detection tasks.

A single LLM rewrite using false-positive and false-negative cases can match manually tuned skill descriptions, reducing engineering effort by 32x while maintaining routing accuracy.

Probabilistic downscaling, crucial for atmospheric science and climate modeling, often uses a mean-residual framework. However, in real-world applications, this approach frequently yields biased and under-dispersive ensembles. The root cause is identified as residual target misspecification: the residual distribution during training differs systematically from that required at test time due to downscaling bias. To address this, ReMatch (Residual Distribution Matching) aligns the training residual distribution to the test-time regime via optimal transport in a low-dimensional PCA space. Experiments on synthetic benchmarks and a real-world HRRR-ERA5 wind field downscaling task show ReMatch substantially reduces under-dispersion, improves calibration, and outperforms strong baselines.

This paper introduces Depth-wise Gradient Augmentation, a paradigm that transforms optimizer updates along depth to leverage inter-layer structure. A simple instantiation, Gradient Smoothing with a window operator, consistently improves optimization and generalization across diverse tasks (LM pretraining, RL post-training, diffusion, ViT) without architectural changes. It promotes structured representation evolution, interpreted as depth-wise preconditioning.

This paper develops a first-principles reduced-order model of GRPO training dynamics, subsuming the empirical single-exponential saturation law as its overdamped limit and adding an inertial term to capture the slow-start phase. It yields predictions tied to independently measurable quantities such as group-size invariance, a sharp stability threshold in the refresh interval, and an overdamped-to-oscillatory transition. The closed-form trajectory fits training reward with R² ≥ 0.91 across three models and two group sizes, and the predicted group-size invariance holds on both the reward curve and out-of-distribution transfer to eight math benchmarks. Additionally, the model furnishes diagnostics that separate failure modes conflated by the reward curve alone, such as reward hacking, advantage degeneracy, policy concentration, and dynamical instability.

This paper introduces process sidecars, a two-coefficient edit family for revoking learned state in language models after safety training, proving second-order accuracy over naive methods and demonstrating improvements across three models.

ReactionAtlas introduces a machine learning framework that autonomously builds chemical reaction networks from a small set of seed molecules without hand-crafted rules. Using a generative model and a DFT-trained machine learned force field, it discovers approximately 47,000 reactions among 12,000 compounds, achieving near-DFT accuracy for transition states. This enables new insights into prebiotic chemistry, particularly the formose cycle.

Hierarchical Global Attention (HGA) is a drop-in replacement for dense causal attention in pretrained long-context transformers, enabling 64K token context on a single RTX 5090 without retraining or calibration, with minimal quality loss.

This paper reveals that deterministic few-step generation fails on continuous text latents due to geometric constraints: smooth deterministic maps cannot resolve discrete branch choices before sharp categorical readouts. Diagnostics DABI and CCI quantify the sharpness gap between text and image decoders. Two escape mechanisms are identified: categorical commitment (autoregressive) and stochastic re-injection.

Large language models (LLMs) excel at many tasks but often lack structural consistency in their solutions. A new approach called MetaFlow treats workflow generation as a meta-learning problem, training LLMs to compose solution strategies using a two-stage process: supervised fine-tuning on synthetic workflows, followed by reinforcement learning with verifiable rewards. MetaFlow achieves strong performance on in-domain tasks and demonstrates remarkable zero-shot generalization to out-of-domain tasks and novel operator sets.

This study introduces the NHANES Accelerometry Cardiometabolic Benchmark using data from 1,381 adults (2003-2006). It evaluates ridge regression, XGBoost, and TabPFN v2 for predicting HbA1c, triglycerides, and CRP from accelerometry and lifestyle data. TabPFN v2 performs best for HbA1c and CRP, while triglycerides remain largely unpredictable. Conformal prediction shows marginal coverage targets are met, but subgroup inequities exist.

A new competitive optimization framework, MCO-PDE, discovers shared partial differential equations from multiple datasets, overcoming limitations of single-dataset methods by dynamically assessing data credibility via soft-competitive weighting, recovering canonical equations with high accuracy from as few as 50 observations per dataset.

AgRefactor is an LLM-based multi-agent workflow for refactoring software into HLS-compatible code. It features a self-evolving memory system and integrates automated tools, outperforming state-of-the-art on 9 out of 11 benchmarks with up to 6.51x speedup.

This paper formalizes the LLM Jury under the Huber contamination model and shows that PoLL incurs unbounded bias under any positive contamination if a single judge fails in a biased, LLM-typical way. By framing jury consensus as robust mean estimation, the authors propose RoPoLL using the geometric median as aggregation, achieving optimal breakdown point 1/2. Experiments across 13 judges, three benchmarks, and four corruption regimes show RoPoLL dominates PoLL on every biased corruption type, with a 3-judge committee at 38B outperforming Mistral-Large-3 (675B) by 1.31x under 30% bimodal-random corruption.

This paper introduces HASTE, a hierarchical multi-agent system that organizes cross-competition knowledge into three scope tiers (global, domain, competition-specific), with LLM-driven abstraction between tiers. On the MLE-Bench Lite benchmark, HASTE achieves a 77.3% medal rate using Claude Sonnet 4.6 at 12h per competition. Warm starts use 52% fewer refinement iterations, suggesting that better knowledge organization can partly substitute for model strength and compute budget.

A new study explores multi-agent deliberation methods for legal reasoning using LLMs, introducing two novel frameworks inspired by courtroom procedures and legal argumentation. Experiments show comparable overall performance to single models but significantly distinct answers, with multi-agent approaches excelling in tasks requiring critical thinking from multiple perspectives.

Traditional agents assume users have well-formed preferences, but often users lack domain knowledge. This paper proposes CoPref model and CoShop benchmark, finding that top agents achieve only 56% accuracy after 5 turns, failing to expand users' understanding of their own needs.

This study proposes LearnStop, a hidden-state-free checkpoint stopper for reasoning language models. Experiments show that learned multi-feature stopping outperforms scalar exits on free-form math tasks, but scalar rules are competitive on multiple-choice and very hard settings. The main finding is that learned stopping is useful when many questions become correct before full budget but lack a reliable scalar signal.

Large language models (LLMs) in multi-turn conversations should update beliefs as evidence accumulates, yet evaluations often ignore this process. BayesBench introduces a suite of simulation environments with three progressively complex tasks (Bayesian estimation, prediction, and latent-framed prediction) to assess how closely LLM belief updates match a rational Bayesian reasoner. Across seven LLMs (3B-70B), scaling improves latent inference and evidence accumulation, but gains do not reliably transfer to downstream prediction, revealing a gap between inferring latent structure and rationally updating beliefs.

This experimental study investigates AI-driven discovery of simulation models using natural language queries. It examines data representation, transformer-based embeddings, and retrieval strategies, finding that data representation matters, open-source embeddings perform well, and reranking is crucial for complex queries. The work provides a baseline for AI-driven model composability and interoperability.

This paper introduces Contrastive Reflection, an iterative framework for optimizing prompts in agentic information retrieval workflows. By analyzing retrieval or reasoning traces, it identifies error-anchored behavioral slices and contrasts them with nearby successful examples, prompting a Teacher LLM to propose targeted edits. On HotpotQA, exact-match accuracy improved from 51.4% to 60.4%, outperforming failure-only and random-evidence variants, and comparable to MIPROv2 (59.4%) and GEPA (57.0%). The framework emphasizes interpretability and validation-driven prompt repair.

A study introduces a controlled student-teacher protocol to separate genuine feedback-driven improvement from gains due to repeated attempts or resampling. Across multiple benchmarks, self-feedback adds little beyond unguided self-refinement, while strong external teachers yield larger feedback-specific gains. The bottleneck is the student's ability to act on feedback.

Context.dev offers a unified API to scrape, enrich, and extract data from the web, simplifying data collection for developers.

A research team at KAIST has developed an AI model called BehaVERT that treats animal body movements like words in a language, enabling it to autonomously identify core social deficits in a mouse model of autism. The model outperformed existing systems on benchmarks and provides interpretable reasoning, marking a step toward general-purpose behavior analysis tools.

AWS emphasizes its commitment to securely providing frontier AI models through Bedrock, collaborating with Anthropic on guardrails to balance powerful capabilities against misuse risks.

In this op-ed, Emily M. Bender and Nanna Inie argue against anthropomorphizing language in discussions about AI, providing categories of common phrases and offering alternatives such as 'probabilistic automation' for 'artificial intelligence'. They encourage readers to adopt precise language to foster clearer understanding of technology.

Anthropic released Claude Sonnet 5 today as its new default mid-tier model, with a 1M-token context window and promotional pricing. Third-party benchmarks show solid improvements over Sonnet 4.6, but some users express disappointment due to higher per-task costs and the absence of Fable 5. Fable/Mythos 5 was later re-approved after government collaboration.

The US commerce department has lifted export controls on Anthropic's Fable and Mythos AI models, less than three weeks after the company was ordered to suspend access over national security risks. Anthropic says access will be restored tomorrow.