MS-GLA introduces Multi-Scale Gated Linear Attention, distributing attention heads across multiple temporal resolutions to address representational bottlenecks in GLA Transformers. Coarser resolutions pool longer token spans for long‑range dependencies, while finer heads capture local syntax, and a learnable fusion layer recombines outputs dynamically. Experiments on language modeling, recall‑intensive tasks, and long‑context generalization show MS‑GLA outperforms GLA with up to 18.9% improvement on recall tasks and 9.5% lower perplexity on language modeling benchmarks.
By Prasoon Dev, Anirudh Sankar, Vasudeva Varma
The paper reports that all 1,325 theorems from the VeriContest benchmark—comprising 1,007 competitive‑programming problems in Rust with Verus specifications—were successfully proved using Rust‑Prover, a Lean 4‑based verifier. The authors translated each Verus specification and Rust program into Lean, then used Lean’s kernel to confirm the proofs, achieving a 70% first‑iteration success rate and completing 1,259 proofs in under 32 hours on Claude Opus 5.5. The translated Lean programs matched the original Rust outputs on 21,413 benchmark test cases, and the study also compared performance across multiple large‑language‑model settings, noting that higher reasoning effort increased cost without improving proof counts.
By Traian Serbanuta, Jun Xu, Andrei Stefanescu, Cosmin Radoi
The paper introduces DCOCO (DFlash Contextual Corrector), a lightweight mechanism that restores intra-block dependencies in speculative decoding by evaluating multiple candidate tokens for contextual compatibility across adjacent draft positions. DCOCO maintains parallel drafting while enabling contextually informed token selection at low additional cost. Across multiple verifier backbones and benchmarks, DCOCO achieves the highest average throughput, delivering speedups of up to 6.6× over vanilla decoding and outperforming competing methods by 6–23% under concurrent serving loads.
By Roy Eisenstadt, Harel Lidar, Tal Gonen, Omer Luxembourg, Itamar Zimerman, Lior Wolf, Eli David, Ilan Naiman
The paper investigates reflective prompt optimization (GEPA) for a multi‑agent deep‑research language‑model, focusing on a topic‑disjoint train/validation/test split to avoid memorization. Across three random seeds, the optimized prompts improve held‑out citation reward from 0.702 to 0.781, with a 95% confidence interval of [0.051, 0.107]. The study shows that the gain is driven by decomposed per‑dimension feedback rather than a single scalar reward, and that the improvement persists even with an out‑of‑family verifier and a hand‑written version of the discovered recipe.
By Elias Lumer, Sahil Sen, Vamse Kumar Subbiah, Kevin Paul
GeopoliticalQA is a new benchmark for retrieval‑augmented question answering over 415 open‑source geopolitical and strategic intelligence reports, containing 5,525 document‑grounded question‑answer pairs that cover factual, inferential, comparative, and risk‑oriented queries. The authors propose a document‑aware generation pipeline that profiles reports before generating questions, and they validate dataset quality through expert annotation and an LLM‑as‑judge protocol. Experiments on the benchmark show that while retrieval performance is similar across models at the document level, chunk‑level retrieval is not driven by embedding‑model scale, and answer generation remains difficult even with gold evidence, especially for inferential and risk‑oriented questions.
By Veronika Solopova, Ivan Havlytskyi, Viktoria Skorik, Ostap Vykhopen
MADS (Multi-Agent Dialog Simulation) is a synthetic data generation pipeline that creates structured, diverse dialogue trajectories, including adversarial conversations that reflect realistic failure scenarios. The resulting MADS-Bench benchmark covers six domains, 55 personas, and 1095 conversations, with about 66% featuring failure-prone situations. Models trained with MADS are evaluated using a two-stage approach combining supervised fine-tuning and reinforcement learning with verifiable rewards to assess robustness.
By Arkajyoti Chakraborty, Adam Zernik, Ishika Kulkarni, Mudit Arora, Hugo Lin, Justin Chiu, Alessandro Di Bari, Neha Gupta, Andreas Stolcke
The paper introduces DistressBench, a benchmark comprising 718 English clinician-authored conversations covering 23 self-harm subcategories, each evaluated against a weighted rubric across seven crisis-support dimensions. Models are scored by the fraction of criteria they satisfy, revealing that the best models meet 88.4% of weighted criteria while clinicians achieve 98.7%. The study highlights gaps in crisis response, notably that 35.3% of conversations where crisis is recognized fail to route users to human help, underscoring the need to evaluate models on supportive care rather than mere refusal.
By Drew Rein, Patrick Oathout, Vishal Kumar, Udari Madhushani Sehwag
The paper introduces SEEDS, a preview-guided evidence admission framework that allows search agents to selectively admit retrieved documents into their reasoning process. SEEDS exposes lightweight previews of candidates and lets the agent decide which and how many to fully inspect based on its current reasoning state, enabling arbitrary subset selection rather than fixed or rank-based policies. Experiments on nine knowledge-intensive QA benchmarks show that SEEDS improves search effectiveness and reduces full-evidence consumption compared to traditional retrieval prefixes.
By Jiajin Liu, Chenxi Xu, Dongzhe Fan, Yuanfu Sun, Qiaoyu Tan
The paper explores whether language models encode components that represent narrativity. By examining GPT‑2‑Small, Gemma‑2‑2B, and Llama‑3.1‑8B, the authors identify specific attention heads, MLP neurons, and sparse autoencoder features whose activations correlate with narrative properties. Ablating these components raises next‑token prediction loss more for high‑narrativity texts and diminishes narrativity in open‑ended generations, with the strongest effects seen in story completion and narrative question answering tasks.
By David Yiyang Liu, Aditya Joshi, Paul Dawson, Andrew Piper
Evo$^2$ is a framework that jointly evolves a language model’s weights and its harness—the set of prompts, tools, and feedback mechanisms—using a dual-role pipeline where a single policy acts as both a Proposer and a Reflector. The training pipeline EvoTIDE combines supervised fine-tuning with interleaved GRPO to refine both roles based on harness edits and execution outcomes. Experiments on coding benchmarks show Evo$^2$ outperforms existing self‑evolution methods, achieving higher scores on Polyglot and SWE‑bench Verified, and its evolved harnesses transfer across tasks and models.
By Chejian Xu, Shujian Zhang, Wenxuan Zhou, John Lambert, Beck Chen, Bo Li, Andrew Hard, Rajiv Mathews
CliQ‑RRG is a two‑stage framework for generating radiology reports that aligns visual and textual data using disease‑aware embeddings and prior‑guided attention to capture multi‑view dependencies. The second stage retrieves graph‑grounded clinical evidence and employs a large language model to produce concise, interpretable question‑answer style reports with diagnostic summaries. Experiments on two chest X‑ray benchmarks show that CliQ‑RRG outperforms previous methods on both lexical and clinical metrics, producing accurate and clinically coherent reports.
By Chandravardhan Singh Raghaw, Shahid Shafi Dar, Arnav Nishith Sharma, Shrey Modi, Nagendra Kumar
SharP is a text‑generation framework that uses both hard and soft prompts to guide large language models in creating task‑specific synthetic news samples for low‑resource fake news detection. The method combines structured factual constraints with learned semantic patterns, and employs a boundary‑aware strategy that focuses on classifier‑uncertain areas to refine decision boundaries. Dual objective optimization balances semantic alignment with source data and classifier feedback, aiming to improve robustness and generalization when labeled data are scarce.
By Ziyu Luo, Wei Li, Haoping Huang, Kecheng Liu, Min Gao
BLiMPPI is a multilingual benchmark comprising over 12,000 minimal pairs for negative polarity items (NPIs) across 33 languages. It is constructed by extracting NPIs from Universal Dependencies treebanks, labeling their licensing contexts, and generating minimal pair perturbations. The study evaluates 12 multilingual and 2 monolingual language models, revealing that multilingual models perform poorly on low‑resource languages and on pairs that cannot be solved by simple n‑gram patterns, while also exploring how NPI typology, language family, licensing variability, and NPI‑licenser distance affect performance.
By Arzu Burcu G\"uven, Anna Rogers, Leonie Weissweiler, Joakim Nivre, Rob van der Goot
OccludeBench is a manually curated benchmark that tests vision‑language models (VLMs) on their ability to recover fully occluded objects from remaining visual evidence. The benchmark distinguishes direct cases—requiring recovery from a specific physical cue such as a shadow or reflection—from inferential cases, which rely on broader scene context. Across five VLMs, accuracy averages 42.7 % on direct cases versus 62.7 % on inferential cases, a 20‑point gap, while humans achieve 63.4 % and 68.5 % respectively with no significant gap. The study introduces OccludeTrace, a four‑stage diagnostic that isolates the failure point, revealing that models struggle most when recovering the relevant object from the full image projection. Interventions such as inpaint‑then‑confront reduce the gap to a statistically non‑significant difference for models with image‑generation capabilities.
By Jasin Cekinmez, Addison J. Wu, Ryo Mitsuhashi, Linrong Cai, Xingyu Fu, Zhuang Liu
TokenSwap introduces a method to benchmark and reduce the modality gap in multimodal large language models (MLLMs) by interleaving semantically aligned images with text tokens. By transforming text-only benchmarks like MMLU into image‑interleaved versions (TokenSwap‑Bench), the authors reveal a pervasive performance drop—averaging 19.6%—across 42 MLLMs, with reasoning models showing smaller gaps. Training with TokenSwap mitigates this gap while maintaining strong text‑only and vision‑language performance.
By Andong Hua, Colton Bishop, Igor Mordatch, Arian Hosseini, Jindong Gu, Aleksandra Faust, Rebecca Roelofs, Yao Qin
M3-DuplexBench is a new benchmark designed to evaluate full‑duplex spoken dialogue systems (FDSDSs) that can listen while speaking. It covers multi‑turn interactions in both English and Japanese across casual conversation and question‑answering domains, jointly assessing interaction timing and response content. The benchmark also offers a reference‑conditioned setting to test sensitivity to dialogue history and trajectory mismatch.
By Ryo Fukuda, Atsushi Ando, Hiroki Kanagawa, Takatomo Kano, Marc Delcroix, Naohiro Tawara, Yuya Chiba
The paper revisits large‑language‑model (LLM) based automatic speech recognition (ASR) from an entropy‑allocation perspective, introducing three metrics to analyze how entropy reduction is distributed between the speech encoder and the LLM. It proposes a multi‑stage training strategy that addresses entropy‑allocation inefficiencies, including a redesigned pretraining phase to reduce the speech‑text modality gap and an iterative asynchronous fine‑tuning stage to maintain functional decoupling and limit encoder drift. Experiments on Mandarin and English benchmarks demonstrate competitive performance with state‑of‑the‑art models using only 2.3 B parameters, while also reducing hallucinations through the decoupling‑oriented design.
By Yuan Xie, Jiaqi Song, Guang Qiu, Xianliang Wang, Ming Lei, Jie Gao, Jie Wu
The paper investigates whether continual learning regularization is necessary for financial speech recognition. Using Whisper‑small adapted to the Earnings22 benchmark, the authors find that all tested methods actually improve performance on new financial data rather than causing catastrophic forgetting, indicating that task complementarity makes regularization counterproductive. When regularization is required, a layer‑wise mini‑KFAC approach yields lower word error rates and more stable optimization than diagonal EWC.
By Thai Quoc On, Homayoon Beigi
The paper investigates whether language models can be fine‑tuned to produce outputs that better match intended probability distributions, even when the valid outputs cannot be enumerated in advance. Two calibration fine‑tuning methods—soft‑target and hard‑target—are evaluated across 12 models from four families, showing substantial improvements in structured‑sampling fidelity on unseen distribution families and parameter settings. The gains also transfer to broader stochastic generation tasks, expanding first‑token support and improving NoveltyBench scores, though some downstream capabilities such as arithmetic reasoning may be slightly affected.
By Davide Baldelli, Sruthi Kuriakose, Maryam Hashemzadeh, Amal Zouaq, Sarath Chandar
The paper introduces Trajectory-Aware Policy Optimization (TAPO), a new framework for reinforcement learning that addresses the state distribution shift caused by long-horizon autoregressive rollouts in large language models. TAPO derives a state-distribution correction from the off-policy policy gradient, integrates it into the RL objective, and uses a geometric-mean estimator for variance reduction along with length-aware dynamic clipping to stabilize updates. Experiments on eight competition-level benchmarks show TAPO outperforms existing LLM RL methods, achieving up to 48% improvement on mathematical reasoning and 12% on algorithmic coding tasks while enhancing training stability.
By Zhicheng Cai, Weinan Dai, Yuyan Zhou, Wei-Ying Ma, Ya-Qin Zhang, Hao Zhou