Chronicle introduces a method called cut‑point replay to make regression testing of large language model (LLM) agents reproducible. It records an agent’s run at non‑deterministic boundaries as immutable envelopes and then replays selected boundaries while executing the rest live, enabling continuous‑integration tests that detect faulty code changes. Benchmarks show minimal overhead, perfect bit‑stability, and effective detection of unsafe actions in a mutation study.
By Tisha Chawla, Susheem Koul
oMeBench is a large-scale, expert-curated benchmark designed to evaluate large language models (LLMs) on organic mechanism reasoning. It contains over 10,000 annotated mechanistic steps, including reaction type labels, intermediate structures, and difficulty ratings, and introduces the oMeS scoring framework to assess logical consistency and chemical structural similarity. Evaluation shows that while current LLMs display promising chemical intuition, they often fail to produce correct and consistent multi-step reasoning, though prompting and fine-tuning can bring smaller models up to the level of closed‑source frontier models.
By Ruiling Xu, Yifan Zhang
The paper introduces NeuroCognition, a benchmark based on three neuropsychological tests—Raven's Progressive Matrices, Spatial Working Memory, and the Wisconsin Card Sorting Test—to evaluate foundational cognitive abilities in large language models (LLMs). It finds that while LLMs excel on text tasks, their performance drops on image-based and more complex tasks, and they fail different parts of the same tasks compared to humans. NeuroCognition correlates with standard general-capability benchmarks yet measures distinct cognitive skills, highlighting where LLMs align with or diverge from human-like intelligence.
By Faiz Ghifari Haznitrama, Faeyza Rishad Ardi, Alice Oh
FOCAL is a framework that aligns fine-grained ECG waveform segments with specific report tags using Optimal Transport, addressing the lack of localized representation in prior methods. It introduces a semantic similarity matrix to mitigate false negatives when reports share diagnoses, and a coarse‑to‑fine enrichment pipeline that employs Large Language Models to recover missing waveform semantics while filtering hallucinations. Experiments on six datasets show FOCAL achieves state‑of‑the‑art zero‑shot prediction and linear probing performance.
By Haitao Li, Che Liu, Zhengyao Ding, Ziyi Liu, Wenqi Shao, Zhengxing Huang
The paper adapts the compact PP‑OCRv6 recognizer for historical text recognition and compares it to a conventional CRNN across various training regimes, including generalized pretraining, domain‑specific training, corpus‑level fine‑tuning, and manuscript‑specific few‑shot adaptation on multilingual Latin and Arabic scripts. While PP‑OCRv6 does not always beat the CRNN when trained from scratch, heterogeneous pretraining significantly improves its generalization. Additionally, fine‑tuned PP‑OCRv6 can surpass a large vision‑language model (Qwen3.5‑based Medusa) that is specifically tailored for historical Latin‑script handwriting recognition.
By Benjamin Kiessling (ALMAnaCH)
MTF‑Net is a Multi‑Modal Temporal Feature Fusion Network that jointly models kinematic, appearance, and contextual cues for pedestrian intention prediction. It fuses four modalities—bounding‑box dynamics, human pose keypoints, local context, and scene‑level semantics—within a recurrent framework enhanced by gated linear units (GLUs) and an attention‑guided fusion head. Evaluations on the PIE and JAAD benchmarks show that MTF‑Net outperforms recent transformer‑ and graph‑based models, achieving up to 0.95 AUC on PIE and 0.94 AUC on JAAD while maintaining real‑time performance.
By Md Mahfuzur Rahman, Pengzhan Zhou, A. F. M. Abdun Noor, Md Imam Ahasan, Md Mustafizur Rahman, Fang Qu
The paper introduces BATON, a dual‑axis policy optimization framework for large language model agents that addresses two key aspects of reinforcement learning: intra‑trajectory feedback attribution and inter‑trajectory objective aggregation. BATON combines Bayesian Feedback Attribution, which builds a feedback‑conditioned posterior over sampled actions, with Trajectory Mass Normalization, which ensures equal optimization weight for complete trajectories. Experiments on ALFWorld, WebShop, and SearchQA demonstrate that each axis independently improves performance and that their combination yields the best results across different model scales.
By Yingxuan Zhuang, Binhe Yu, Jingxiao Yang, Ruopei Sun, Ziting Li, Cheng Tan, Xuhong Zhang, Jianwei Yin, Jintao Chen
The paper critiques the common practice of evaluating depth usage in depth‑recurrent language models by truncating depth during inference and measuring performance decline. It argues that this method conflates three distinct effects—fewer block applications, reduced computation, and an out‑of‑distribution readout—yet is usually interpreted as measuring only the second. To address this, the authors introduce the Depth Control Protocol (DCP), a suite of positive and negative controls that isolate each factor, along with a training intervention to confirm causality, specifically tailored for depth‑wise weight‑sharing architectures.
By Ha Van Dau, Thanh Tung Khuat, Nguyen Thanh Dung
arXiv:2609. 20519v1 Announce Type: new Abstract: As coding agents move from supervised code completion to unattended, around-the-clock exploration, their work expands from isolated predictions into long trajectories of reasoning, tool use, and feedback.
By Haozhe Liu, Tian Ye, Sensen Gao, Qihang Cao, Yitong Li, Mingchen Zhuge, Duomin Wang, Ruihua Zhang, Ping Luo, Jiawang Bian, Lei Zhu, Ligeng Zhu, Enze Xie, Song Han
The paper analyzes the environmental footprint of machine learning model training, focusing on large language models and their hardware. It finds that energy use and environmental impacts have risen exponentially over the past decade, even when employing carbon‑efficient electricity and more efficient hardware. The study argues that optimization strategies alone cannot curb these impacts due to a rebound effect, and stresses the need to evaluate hardware life‑cycle impacts and integrate environmental metrics into NLP research practices.
By Cl\'ement Morand (STL), Anne-Laure Ligozat (ENSIIE, LISN, STL), Aur\'elie N\'ev\'eol (STL, LISN)
The paper introduces an adaptive Mixture-of-Experts (MoE) framework for time series forecasting that incorporates expert-specific losses to give each expert a direct learning signal independent of gating weights. The overall objective combines base forecasting loss with these expert losses, encouraging experts to specialize on different temporal segments. A partial online learning strategy is added for efficient incremental updates, and experiments on economic, tourism, and energy datasets show the method outperforms state‑of‑the‑art neural models and foundation models, with ablation studies confirming the benefit of expert loss integration.
By Btissame El Mahtout, Florian Ziel
FakeSpotter is a new tool that estimates the viral misinformation risk of textual content by measuring structural fingerprints of misinformation instead of directly judging truthfulness. It operates across linguistic, narrative, logical, and critical‑thinking dimensions, using repeated large language model assessments and domain‑specific logistic regression classifiers for both short and long texts. In a labeled corpus of 764 texts, FakeSpotter achieved macro F1 scores of 0.788 for short texts and 0.793 for long texts, and its interpretive layer offers explainable outputs such as feature‑based scores, signal agreement, and a caution index for social listening.
By Giovanni Spitale, Federico Germani
The paper introduces D-Quant, a KV cache quantization framework that addresses the memory bottleneck of large language models by using a drift mechanism to convert entropy-coded representations into fixed-size bitstreams. This approach leverages the non-uniform distribution of KV cache values—after rotation and normalization, they approximate a normal distribution—allowing entropy coding to assign shorter codewords to frequent symbols while maintaining regular memory layouts suitable for parallel attention kernels. D-Quant thus aims to reduce memory footprint and bandwidth usage without sacrificing performance.
By Yi Su, Hong Liu, Guanghua Yu, Jianchen Zhu
The paper introduces pre‑trained models for extracting variant‑phenotype relations from biomedical text, focusing on the SNPPhenA corpus. Fine‑tuning small BERT‑based models, especially DeBERTa, achieves performance close to the current state‑of‑the‑art. Moreover, careful fine‑tuning of Google’s Gemini Pro 1.0 surpasses existing benchmarks on both sentence‑level and abstract‑level relation extraction tasks.
By Claudiu Creanga, Liviu P. Dinu, Daniela Gifu
The paper introduces SALR, a schema‑anchored latent reasoning approach for generating logical forms in knowledge‑base question answering. SALR delays explicit schema commitments by generating continuous thoughts in hidden states and aligns these thoughts with a codebook of KB schema elements, guided by an alignment objective derived from gold logical forms. Experiments on GrailQA and WebQSP demonstrate that SALR consistently outperforms strong baselines, notably improving compositional question performance by 2.86 F1 points over TIARA.
By Guangze Gao, Zixuan Li, Sikui Zhang, Chunfeng Yuan, Wenjuan Li, Bing Li, Xiaolong Jin, Weiming Hu
The paper introduces StratCBT, a new dataset of 9,688 psychological counseling sessions with 256K utterances, each counselor response aligned to one of eight Cognitive Behavioral Therapy (CBT) strategies. It was created by modeling clients’ negative thoughts and generating high‑quality conversations through self‑chat, using realistic sessions for guidance. Experiments show that strategy‑aligned generation improves professional and effective counseling when evaluated with large language model‑simulated clients.
By Zimu Wang, Yiwen Jiang, Xiangyu Zhao, Yaling Shen, Jiahe Liu, Stephanie Fong, Maxmartwell H Cheng, Guilherme C Oliveira, Anh Nguyen, Robert Desimone, Barnaby Nelson, Dominic Dwyer, Zongyuan Ge
The paper investigates why Word-in-Context (WiC) remains difficult for language models, suggesting that the lack of an explicit sense inventory contributes to the challenge. By evaluating open LLMs on both WiC and traditional Word Sense Disambiguation (WSD) tasks, the authors find that providing candidate senses—akin to WSD—consistently improves WiC performance. Human evaluation indicates that many WiC errors stem from label ambiguity or mismatched sense boundaries, with models often over‑discriminating senses and making overly fine‑grained distinctions.
By Yi Zhou, Kiamehr Rezaee, Danushka Bollegala, Mohammad Taher Pilehvar, Jose Camacho-Collados
The paper investigates how privileged information—such as a teacher’s full solution or reasoning trace—affects on‑policy self‑distillation (OPSD) in language models. Using the AMPLE‑Math benchmark, the authors compare distillation with and without extra teacher views, finding that reference‑free distillation explains most gains for Qwen3‑1.7B, while additional references provide modest benefits, especially for polished solutions. The study also shows that the impact of privileged data depends on the student’s training regime and that altering token‑level supervision can leave student behavior largely unchanged.
By XiuYu Zhang, Wei Chow, Junfeng Fang, Zhenkai Liang, Tat-Seng Chua
The paper introduces On‑Demand Attention (ODA), a decoding strategy that lets pretrained language models decide when to use global attention based on a lightweight recall head. ODA keeps the original model weights unchanged, only training the recall head, and can be implemented with GPU‑side conditional execution to reduce global reads. Experiments on Qwen, Gemma, and hybrid‑attention models show that ODA largely recovers performance lost by local attention while cutting the number of global attention operations.
By Haibo Feng, Ruiqi Liang, Hanyang Peng, Shiqi Yu
The study investigates how limited reading capacity and claim wording influence consensus outcomes in language‑model networks. By modeling message capacity as the number of messages an agent reads, the authors show that when agents read fewer than about 6.4 messages on average, a wrong consensus becomes unreachable. However, the wording of a claim—its inherent threshold—can override this effect, leading to incorrect consensus even when most agents start correct.
By Makoto Fukushima