Model releases, architecture work and prompting research on large language models — from frontier-lab announcements to the arXiv papers behind them.
arXiv:2609.27822v1 Announce Type: cross Abstract: A common multi-agent design asks agents to report confidence and lets the highest-scoring agent speak next, implicitly using one scalar both to route...
arXiv:2603.08166v2 Announce Type: replace Abstract: Automated Drug Combination Extraction (DCE) from large-scale biomedical literature is crucial for advancing precision medicine and pharmacological...
arXiv:2603.09222v2 Announce Type: replace Abstract: Efficient context compression is critical for retrieval-augmented question answering in resource-constrained settings, where long retrieved context...
arXiv:2603.16932v2 Announce Type: replace-cross Abstract: Vision-language models (VLMs) typically process images at a native high-resolution, forcing a trade-off between accuracy and computational ef...
The paper introduces Reinforcement Learning with Decomposed Subtasks (RLDS), a method that splits trajectory rewards into per‑subtask shares before policy updates, replacing the scalar advantage used in Group Relative Policy Optimization (GRPO). RLDS employs Subtask‑Decomposed Advantage Estimation (SDAE) to compute group‑relative advantages and distribute credit to tokens based on subtask importance, focusing on steps where a reflection marks a subtask as consequential. Experiments on four benchmarks—FrozenLake, HotpotQA, ScienceWorld, and DeepResearch—show that RLDS improves performance on high‑heterogeneity tasks (ScienceWorld and FrozenLake) and is more compute‑efficient than scalar GRPO for long rollouts.
NV-Reason-CT is a generative vision‑language model designed for chest and abdominal CT analysis that preserves native 3D visual encoding and incorporates radiologist‑guided reasoning. The system couples a 3D vision transformer with a language model, feeding all visual tokens and their 3D coordinates directly into language decoding to maintain volumetric spatial information. Trained on a curated corpus of about 550,000 multimodal instruction examples, the model supports abnormality classification, report generation, and interactive reasoning, achieving strong performance on CT benchmarks and reducing expert interpretation time by 50%.
arXiv:2605.29313v2 Announce Type: replace Abstract: LLM multi-agent systems often coordinate through natural-language dialogue or loosely structured shared memory, making intermediate state difficult...
arXiv:2604.08602v2 Announce Type: replace-cross Abstract: Server-based screening tools impose subscription costs, while open-source alternatives require coding skills, and full-text screening has rem...
Researchers use synthetic survey respondents generated by large language models as substitutes for human samples, but current validation methods often compare them to human surveys in ways that may not reflect real-world consequential behaviour. The authors propose a new validation framework that requires explicit statements of how well synthetic data correspond to human behaviour, specifies which diagnostics are addressed, and demands subgroup-level validity claims to avoid misrepresentation. The framework operationalises distributional, procedural, and recognition justice dimensions and introduces within-persona counterfactual experiments, illustrated with a case study on electric vehicle charging tariffs and concluded with a reporting checklist for researchers.
arXiv:2609.27901v1 Announce Type: new Abstract: Video is a rich representation of a physical event, capturing appearance, geometry, motion, and temporal evolution. Other modalities, such as 3D body m...
arXiv:2609.28187v1 Announce Type: new Abstract: Self-supervised vision transformers trained with DINO-style objectives exhibit striking emergent semantic representation quality across visual tasks, y...
The paper introduces R3Con, a framework that applies principled design rules—derived from cognitive relevance theory—to construct effective context representations for large-scale reasoning tasks. By systematically organizing vast, heterogeneous information into manageable structures, R3Con outperforms nine state‑of‑the‑art baselines on two large‑document reasoning benchmarks, achieving up to 20 percentage points higher accuracy. Notably, R3Con enables smaller models (4B–9B) to surpass larger 35B models and even Claude Code, demonstrating that principled context representation can reduce dependence on model scale.
PASTABench introduces a benchmark of 1,139 multi-turn trajectories to evaluate proactive safety monitoring in large language models. It formalizes three dimensions of intervention—whether, when, and what risk—to address gaps in step-level isolation and post-hoc trajectory assessment. The study finds that proactive intervention is largely unsolved, with the best model achieving only 40.74% optimal-timing interventions, and reveals that smaller models’ safety scores are often driven by lexical overfitting rather than true risk comprehension.
The paper introduces FWBench, a benchmark for assessing how language models choose and employ time‑series forecasts under budget constraints. Using 1,251 electricity and cycle‑hire cases, the study evaluates agents that select models, histories, and horizons to submit capacities that minimize a loss‑cost objective. Experiments with both hosted and local configurations—including small language models and TSFMs—show that GPT‑6 Astra can achieve superior performance by selectively buying inexpensive short‑horizon forecasts, using only 2.5% of the budget.
The paper investigates how small language models (SLMs) perform in knowledge graph question answering (KGQA) when evaluated on the reasoning paths they take, rather than just the final answer. Using the THESEUS navigation and traceability framework, the authors test frozen, off‑the‑shelf SLMs as local action policies that choose graph actions and decide when to stop, without any task‑specific training or free‑form answer generation. By measuring both Hits@1 and Path Edit Distance (PED) across the Kinship and MQuAKE‑ST datasets, the study finds that models vary significantly in both answer accuracy and path fidelity, and that prompting can either help or hurt navigation depending on the model. "whyItMatters":"The results show that evaluating SLMs solely on endpoint accuracy can be misleading, highlighting the need to assess reasoning path fidelity in KGQA tasks."
The paper "Query Implied Generative Engine Optimization" introduces QI‑GEO, a method that infers user intent directly from documents to enhance visibility in Generative Search Engines. By approximating a document’s intent space, QI‑GEO identifies missing yet relevant content, improving objective scores by up to 15.9% and subjective scores by up to 17.6% on GEO‑Bench datasets. The approach yields nearly twice as many citation gains as losses, demonstrating that document‑derived intent approximations can boost content visibility without explicit query inputs.
The paper studies matrices built from block‑diagonal factors interleaved with fixed permutations, a structured family useful in deep learning for balancing expressivity and efficiency. By applying Riemannian geometry, the authors determine when this class forms a smooth manifold and develop Riemannian tools for the orthogonal two‑factor case. They propose efficient algorithms that use automatic differentiation, allow parameter sharing, and avoid dense matrix construction, testing them on matrix approximation and fine‑tuning large language models, while also exploring properties of factorizations with more factors.
The paper "When Context Misleads: In-context Learning with Jurisdiction in Large Language Models" introduces FakeContextBench, a benchmark of pseudoscientific claims across seven domains to test large language models’ ability to judge whether contextual information should dictate the answer. Experiments show that both commercial and open‑source models struggle with context authority and that standard ICL fine‑tuning can worsen susceptibility to misleading context. The authors propose Jurisdiction In‑Context Learning (J‑ICL), a post‑training framework that adds context validation to the objective, which improves benchmark scores and reality accuracy by significant margins over baseline models.
The paper introduces Anchored Neighborhood Optimization (ANO), a new policy‑optimization method that directly designs a smooth, bounded gain field for the probability‑ratio surrogate objective. ANO anchors the identity map at a ratio of one, peaks at a specified trust‑region boundary, and limits the influence of extreme off‑policy samples while providing a bounded, redescending pull on outliers. Empirical results show ANO consistently outperforms existing methods on Atari and MuJoCo benchmarks, and it remains robust under aggressive learning‑rate settings.
The paper explores using large language models (LLMs) to generate pairwise preferences between slates for synthetic A/B testing of slate recommendation systems. It introduces a validation protocol that checks how well these synthetic preferences align with traditional RecSys metrics and satisfy preference axioms, and examines how LLM pre‑training and configuration influence preference articulation. By combining the synthetic preferences with a generalized Rao‑Kupper model, the authors show that LLM‑based A/B testing can recover stable ranking orderings across different utility weightings, offering a cost‑effective screening step before conducting live experiments.