The paper introduces Labadain-MedQA, a Tetun medical question‑answering dataset created by translating MedAlpaca’s Medical Flashcards and WikiDoc Patient Information into Tetun with Gemini 2.5 Flash, yielding 39,077 QA pairs. Translation quality was evaluated on 50 random pairs, showing human translations scored higher than LLM‑generated ones for medical accuracy and fluency (4.43/5 vs. 4.02/5, p < 0.001). Fine‑tuning Gemma and Llama models with parameter‑efficient methods on this dataset improved performance on Tetun‑translated USMLE multiple‑choice questions by up to 10.79% relative accuracy.
By Gabriel de Jesus, S\'ergio Nunes
The paper introduces TTT‑Consolidate, a framework that leverages multi‑agent reasoning trajectories as adaptation data to improve subsequent reasoning. It consolidates these trajectories into parametric representations via test‑time training and then generates a new solution conditioned on the consolidated information, with meta‑learning used to optimize the consolidation process. Using the newly created OpenParallelThinking corpus of 12,810 math problems and trajectories from six LLMs, TTT‑Consolidate outperforms existing aggregation baselines by up to 20 percentage points on challenging math and STEM benchmarks, especially as model size grows.
By Haotian Wu, Zeming Chen, Hao Zhao, Antoine Bosselut
EEG-HealthLLM introduces a hierarchical adaptation framework that unifies heterogeneous EEG health assessments using large language models. Task‑specific EEG adapters preserve unique spatio‑temporal dynamics, while a shared LoRA module in the LLM backbone captures common high‑level assessment semantics. The new EEG‑HealthBench benchmark reorganizes five EEG datasets into five health‑assessment tasks, and experiments show EEG‑HealthLLM achieves the highest average balanced accuracy and macro‑F1, improving performance while reducing adapter size by 80%.
By Ruidong Ma, Jisheng Dang, Ronghao Lin, Bimei Wang, Jingwen Zhao, Wenjun Ma, Wei-Shi Zheng, Meng Wang, Qi Tian, Tat-Seng Chua
Dual-Axis Policy Entailment (DAPE) is a benchmark that evaluates ordered pairs of policy directives on two dimensions: whether they govern the same policy domain and whether the later directive continues, broadens, narrows, reverses, or shifts to a different domain. The benchmark includes 1,000 expert-annotated pairs from Chinese government policy and a calibration protocol for annotation consistency. Models tested on DAPE perform poorly zero‑shot, but a fine‑tuned Qwen3‑14B model reaches the human ceiling in agreement.
By Charles Chang (Duke Kunshan University), Yuen Yuen Ang (Johns Hopkins University), Bolun Sun (Northwestern University)
The paper introduces Multi-Level Reasoning (MLR), a two-level stochastic framework that separates long chain-of-thought generation into a high-level planner producing structured step descriptors and a low-level executor generating detailed reasoning. MLR replaces single-policy reinforcement learning with an iterative Step‑DPO preference optimization that uses Twisted Sequential Monte Carlo for scalable, stepwise supervision, improving training stability and accuracy. Experiments on math, science, and logic benchmarks show that MLR outperforms both supervised fine‑tuning and strong RL baselines while maintaining robustness over extended reasoning horizons.
By Siheng Xiong, Ali Payani, Faramarz Fekri
Saidzoda‑Tajik‑E5‑Large is a 560 M‑parameter Tajik text embedder fine‑tuned from multilingual‑e5‑large‑instruct. The authors generated 235,140 question‑passage pairs from Tajik laws, tax documents, Wikipedia and web text, trained for one epoch with hard negatives, and mixed in English and Russian pairs to preserve multilingual performance. On unseen Tajik law and tax passages the model retrieves the correct answer in the top ten for 93.6 % and 94.5 % of questions, respectively, outperforming the base model and matching larger models on similar tasks.
By Tohir Saidzoda, Nekruz Maxudov, Bedil Karimov, Abubakr Abdullozoda, Amin Jalilov, Jonibek Kishvarov, Rahmonali Sharipov, Mariyamkhon Shodyeva
Depth pruning removes entire Transformer blocks to shrink and speed up language models, but the resulting damage is uneven across capabilities. In a study of four 1.2–1.7 B instruction‑tuned models, function calling and mathematical reasoning collapsed in every case, while other skills varied in order of degradation. Fine‑tuning a short ‘healing’ run can restore some abilities, with the choice of which blocks to retrain and the allocation of a limited token budget influencing recovery outcomes.
By Piyush Dhoka, Vedant Pandhare, Varun Inamdar, Aadarsh Pathre, Pavitha Nooji
The paper investigates gated knowledge distillation (KD) for low‑resource Bangla generation, testing two gating strategies—EWAD and CHAD—to selectively trust teacher signals. Across six Bangla tasks and multiple student models, neither gate improves over plain KD, and KD generally underperforms cross‑entropy fine‑tuning on gold outputs. The primary advantage of KD is regularisation, with teacher‑free label smoothing narrowing the performance gap, though gating only modestly alters this effect.
By Dipto Sumit, Zobeida Zaman Chowdhury, Syeda Fabiayee Jahan, Farig Sadeque
The paper investigates how large language models (LLMs) tend to reuse a small set of character names in fiction, a phenomenon termed ‘name slop’. It evaluates eight prompt‑level strategies and introduces a new fine‑tuning‑free method called speculative naming, which uses a SlopTracker and a draft‑verify‑reject loop to reduce name concentration. Across 20 LLMs and 44,782 stories, speculative naming cuts the Slop Ratio by 95 % while preserving name naturalness and within‑story collision rates.
By Alejandro Mozo Quesada (Universidad de M\'alaga, Spain), Sergio G\'alvez Rojas (Universidad de M\'alaga, Spain), Francisco L\'opez Valverde (Universidad de M\'alaga, Spain)
LegoEnv introduces a new paradigm for creating agentic environments by recombining reusable atoms—intent, rule, data, and server—rather than building each environment from scratch. The system automatically evolves these atoms when an environment fails to provide reliable feedback, leading to a significant increase in solvable environments from 71.6% to 88.3% after one evolution round. Experiments on AutomationBench show that training on LegoEnv environments doubles the performance of large language models, with SFT boosting Qwen3.6-27B from 16.0% to 32.0% and RL improving Qwen3.8-27B from 37.6% to 40.3%.
whyItMatters":"LegoEnv’s modular, self‑evolving approach enables scalable, generalizable environment construction that directly enhances the training and performance of LLM agents on diverse tasks."
By Zhuosong Jiang, Chengxu Wu, Junqi Liu, Yujie Yi, Siqin Wang, Zhixuan Huang, Jinqian Jiang, Lijun Xiao, Jianxun Li, Yongyang Pan, Jieran Wang, Shiqi Shen, Fei Liu, Xinyu Lin
The paper investigates whether large language models can master "impossible languages"—languages that violate typical structural, cognitive, or logical constraints. By fine‑tuning billion‑parameter models on such languages, the authors find that large models perform comparably on impossible and possible languages, whereas smaller million‑parameter models degrade significantly, especially as semantic complexity rises. The study introduces semantic coherence as a metric and shows that scaling shifts models’ focus from grammatical rules to semantic understanding, reducing sensitivity to low‑level perturbations while preserving high‑level semantic performance.
By Yaning Jia, Chunhui Zhang, Chiyu Ma, Zhongyu Ouyang, Ivory Yang, Soroush Vosoughi
The paper introduces GI-TraceBench, a benchmark of 10,282 gastrointestinal oncology questions that combines provenance question answering, knowledge discovery, and hypothesis evaluation using shared Paper, Claim, and Evidence Span records. It proposes FindLink, a two‑stage framework that first grounds premises with explicit study context and support status, then structures these premises into support plans to generate evidence‑conditioned answers. On a 1,545‑question test set, FindLink raises the balanced family macro score from 67.95 to 87.25 compared to direct supervised fine‑tuning, demonstrating the importance of both premise identification and connection assessment in scientific reasoning.
By Jingxuan Wei, Song Tan, Xiaohan Liu, Chao Sang, Changwei Wang, Linzhuang Sun, Bihui Yu, Fuqiang Wang, Junnan Zhu
The paper investigates how a language model’s first answer to a yes/no question influences its subsequent answer to a different question about the same entity, a phenomenon termed self‑commitment. By fixing the initial answer and measuring changes in the second across 3,600 question pairs for 22 open‑weight models (0.5B–72B parameters), the authors find that each answer writes a commitment direction into the residual stream, and the similarity of these directions predicts answer coupling. They show that ablating a shared direction reduces coupling significantly, while the reading circuit for this direction varies with model size, instruction tuning, and prompts, and that sparing the attention sink during ablation preserves low perplexity cost in most models.
By Marcus Armstrong
The paper introduces SHIFT, a method for unlearning in large language models that avoids explicit forgetting targets by shrinking a fine‑tuned model toward its pre‑fine‑tuning weights and briefly relearning the retain set. In the linearized regime, this shrink‑and‑heal lookahead moves the model a fixed fraction toward the retrained model, and SHIFT extrapolates these per‑token changes into distillation targets for iterative refinement. Evaluated against eleven baselines under the OpenUnlearning protocol, SHIFT achieves the highest aggregate score across all TOFU settings, aligns 3.7× closer to the oracle than the next best method, and preserves the most retain knowledge on MUSE, maintaining proximity to the oracle even under larger optimization budgets and relearning attacks.
By Hoang-Chau Luong, Lingwei Chen
The paper introduces Zep Agent Skills, a production service that autonomously compiles and curates procedural memory for large language model agents by converting verified execution traces into versioned skills. It records each task attempt as a trace, uses a compiler with Reflector and Curator roles to transform verified traces into skills, and applies admission gates and search ranking based on verified outcomes. The open Agent Skills format, originally from Anthropic, allows any host to load compiled skills, and evaluation on several benchmarks shows significant performance gains for weaker executor models without additional training.
By Daniel Chalef, Peter Evans, Pavlo Paliychuk, Mehul Patel, Preston Rasmussen, Jack Ryan
PixelBench is a benchmark that evaluates large language models on their ability to generate and understand images pixel by pixel, without relying on learned image tokenizers or decoders. The benchmark includes text-to-pixel (T2P) and image-to-pixel (I2P) generation tasks across multiple resolutions, as well as a pixel-understanding evaluation. Models produce an RGB palette followed by a grid of indices that map each pixel to a palette color, and the study shows that strong models can generate recognizable images from text and outperform specialized pixelization algorithms at low resolutions, with performance correlating with pixel-understanding ability and improving up to a point before declining at high resolutions. Additionally, Pixel Reconstruction Tuning (PRT) augments visual instruction tuning with caption‑palette‑grid trajectories, and post‑training Molmo2‑4B with PRT data improves performance on 16 visual‑centric benchmarks, demonstrating the utility of explicit pixel reconstruction as auxiliary supervision.
By Haoquan Zhang, Feiyang Liu, Juanxi Tian, Wanhua Li, Haodong Duan, Weiyang Liu
The paper examines how different preprocessing steps in fine‑tuning data influence the performance and fairness of a generative large language model recommender system. Using a controlled pipeline that includes supervised fine‑tuning, catalog‑constrained decoding, and prompt augmentation, the authors evaluate three preprocessing choices—item‑title representation, demographic balancing, and auxiliary item‑knowledge enrichment—on the MovieLens‑1M dataset. Results show that catalog constraints eliminate out‑of‑catalog items, prompt augmentation improves consistency and utility, and that demographic balancing and knowledge enrichment can reduce utility gaps for certain groups while preserving overall performance, though gender balancing may lower utility without improving gender indicators.
By Furkan Yasir Goksu, Yaren Yilmaz, Sule Gunduz Oguducu
DR Pixel introduces a page‑native deep research agent that uses rendered document page images as the sole unit for retrieval and reasoning, bypassing OCR. The open‑source system operates on 1.06 M pages and, without training, outperforms OCR‑based baselines on GPT‑5.4, Gemini‑3.5‑Flash, and Qwen3.6‑27B while requiring fewer tool calls. Training with counterfactual credit reweighting further boosts retrieval recall by ~15 points and accuracy by >10 points on the DRPixelBench benchmark, and brings the model within 1.8 points of an oracle on the ViDoRe‑V3 out‑of‑distribution test.
By Tiansheng Hu, Yichuan Wang, James V. Talwar, Nicholas Tomlin, Chen Zhao, Sewon Min
The study examines whether large language models (LLMs) show syntactic convergence with human users by analyzing turn‑adjacent reuse of context‑free grammar (CFG) rules in 1,901 matched positions across sixteen open‑weight Llama and Gemma models. All models exhibited greater CFG‑rule overlap with the preceding human turn than with unrelated primes, especially for lower‑frequency rules, and instruction‑tuned models showed even higher overlap with the actual prime than the human responses they replaced. However, instruction tuning also increased overlap with unrelated primes, reduced the actual‑versus‑random difference, and lowered conditional rule‑reuse odds when controlling for rule‑set size, while consistently producing responses with higher lexical and semantic similarity to the preceding turn.
By Zandi Eberstadt
HerHealthGPT is an evidence‑grounded conversational AI designed for women's reproductive health. It uses retrieval‑augmented generation, context‑aware clarification, and source validation to avoid premature conclusions and to ask follow‑up questions when user queries are incomplete. Evaluation on 540 messages and a user study with 30 adults showed improved accuracy and a high preference for the retrieval‑enabled version.
By Hassan Saeed Hassan Albattra, Mazen Mohammed Bahgat, Rahatara Ferdousi, Hana Essam Sayed Ahmed Amrya, Mariam Mousa