The paper investigates how large language models encode and use relational information among tokens across transformer layers. By analyzing activations from prompts that require inferring relationships among three cyclic tokens (months, hours, weekdays, musical notes), the authors find a consistent layerwise progression: intermediate layers capture pairwise relationships, while later layers encode the full three‑token relationship to predict the next token. They also identify geometrically structured token relationships that do not influence prediction, and show that constraining models to use only causally relevant joint representations improves next‑token accuracy.
By Gurbir Arora, Toni J. B. Liu, Jiajun Bao, Rapha\"el Sarfati, Christopher J. Earls
The paper benchmarks vision‑language models (VLMs) on two key tasks in connectomics: synapse detection (presence and polarity) and proofreading (split and merge errors). It evaluates 21 models (19 open, 2 closed) under zero‑shot, few‑shot, and LoRA fine‑tuning, comparing them to specialist models on datasets built from public resources. While most VLMs perform at chance in zero‑shot settings, LoRA fine‑tuning with a few thousand labels brings open models to specialist performance, and adapted VLMs outperform specialists on unseen species for merge‑error detection.
By Yicong Li, Junjie Wang, Leander Lauenburg, Ella Hugie, Alexandra Irger, Wanhua Li, Donglai Wei, Hanspeter Pfister
The paper introduces a controlled evaluation to disentangle answer coverage, repeatable task advantages, and gains from pre‑execution selection in large‑language‑model (LLM) harnesses. On 386 MATH‑500 tasks, eight generated harnesses and a baseline with nine identical copies were compared over three executions each, revealing that identical programs provide a 2.16‑point repeat‑averaged oracle headroom while generated programs show more repeatable score patterns but mainly expose persistent weaknesses. The study concludes that coverage and repeatability alone cannot justify claims of useful specialization and proposes an evaluation standard for harness diversity that requires task advantages to persist across executions and improve on additional fixed‑program executions under matched inference budgets.
By Ziyang Xu, Haitian Zhong, Hao Zhou, Hao Qin, Chenhan Jin, Te Qi, Shengze Xu, Tieyong Zeng
The study evaluates multi‑agent debate (MAD) in small language models, testing whether cognitive diversity—via personas, sampling temperature, or model identity—drives performance gains. Across 23 models, five tasks, and over 5,500 runs, MAD consistently outperforms single‑model inference but, when matched for generation budget, it ties or falls behind self‑consistency sampling, with persona prompting actually reducing accuracy. The authors find that MAD’s benefits largely stem from the first answer exchange and that many reported gains are due to ensemble‑sampling effects rather than true diversity, highlighting the need for budget‑matched, contamination‑checked baselines.
whyItMatters":"The findings clarify that MAD’s perceived advantages may be overestimated and that future debate mechanisms must be evaluated against rigorous, budget‑matched baselines to ensure genuine performance improvements."
By Leonardo Ferreira, Gardenia Liu, Kaden Zheng
The paper investigates how scaling a team of small language‑model agents affects performance across different orchestration architectures. By testing eight architectures on five short‑answer benchmarks and an executable‑code benchmark, it finds that team scaling yields large gains on arithmetic word‑problem tasks but only modest improvements on multiple‑choice and code generation tasks, with no single architecture dominating all tasks. The authors explain these patterns using a generate‑transform decomposition that separates coverage and transformation effects, showing that arithmetic tasks benefit from both coverage and critic‑guided transformation, while other tasks are limited by saturation or poor conversion.
By Blaz Bertalanic, Carolina Fortuna
The paper introduces the Distributed‑Evidence Paradox, where long‑running LLM agents compress past interactions into persistent memories that may not be fully supported by the interaction history. It defines three key requirements—evidence scope, compositional validity, and admission reliability—and proposes DerivAudit, a framework that checks whether a memory is truly supported by the available history. Experiments on two memory corpora show that expanding the evidence base can recover support for many memories, yet many remain unsupported, and broader evidence alone does not guarantee reliable admission.
By Hongjun Liu, Chen Zhao
The paper introduces REST (REpresentation‑Supervised Thoughts), a new training objective for latent recursive language‑model systems that supplements cross‑entropy loss with differentiable penalties enforcing causality, minimality, separability, and stability of internal thought representations. By applying REST to both single‑agent and multi‑agent setups without changing architectures or adding inference parameters, the authors achieve up to 7.5 percentage‑point gains in accuracy across seven diverse benchmarks and a 30 % improvement in convergence to the final answer. The method also yields more informative latent thoughts, improving interpretability of agent communication.
By Fahd Seddik, Fatemeh Fard
The paper introduces OTROPE, a likelihood‑free method for off‑policy evaluation of large language models (LLMs) that uses optimal transport to align labeled samples from a behavior model with unlabeled samples from a target model in a semantic space. OTROPE corrects human‑labeled residuals with proxy predictors, achieving a doubly robust evaluation without requiring behavior‑policy modeling or density‑ratio estimation. The authors provide theoretical guarantees for consistency and convergence, and demonstrate through synthetic and real LLM tasks that OTROPE outperforms existing baselines and can elevate weaker evaluators to match or exceed stronger ones.
By Liner Xiang, Wenbo Zhang, Hengrui Cai
The study examines how four large language models (Gemma-3-4b-it, Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct, and GPT-5-nano) exhibit repetition effects in a social media simulation. Using a two‑phase within‑context design, researchers collected 336,000 ratings on truth, importance, sentiment, and interest for 100 statements across 10 feed variants and 3 replications. Results show distinct patterns: Gemma-3 displays a genuine Illusory Truth Effect, Qwen2.5 shows a mere exposure effect, GPT-5-nano shows no truth boost and mild skepticism, and Llama-3.1 shows a small truth boost with reduced evaluative dimensions.
By Azza Bouleimen, Nicol\`o Pagan, Anik\'o Hann\'ak
HyperZip introduces an efficient data compression framework that uses diffusion-based large language models (dLLMs) with Multi-Token Prediction to speed up compression. It addresses the trade‑off between throughput and compression rate by employing a hypernetwork that generates data‑specific updates from a context representation, allowing the dLLM to adapt to target data without costly fine‑tuning. Experiments show HyperZip outperforms state‑of‑the‑art baselines in both compression rate and speed.
By Thai Nguyen, Khang Tran, NhatHai Phan
The paper introduces Persona Dosing, a method that uses an activation‑steering coefficient to control the intensity of a language model’s persona traits. By conditioning a FLAS controller on trait descriptions and calibrating its flow time against measured trait expression, the approach can adjust trait intensity without requiring paired training data. Experiments on Llama‑3.1‑8B, Qwen3‑8B, and Gemma‑3‑4B show significant increases in core‑trait expression and low targeting errors across multiple traits.
By Zehao Jin, Junran Wang, Ruixuan Deng, Jiahao Chen, Jingyuan Zhang, Yuxuan Zhang, Xinjie Shen
The paper introduces PhenoAIR, a reliability‑aware multi‑agent framework for predicting mechanisms of action (MOA) from Cell Painting morphological profiles. It treats retrieved neighbors as uncertain evidence, calibrating their reliability based on source, phenotype stability, and mechanism confusion, and refines predictions through controller‑guided evidence evaluation. PhenoAIR is benchmarked on JUMP Cell Painting data and outperforms both representation‑matching and LLM‑based baselines across controlled, realistic, and open‑world settings.
By Jiayuan Chen, Botao Yu, Tianyu Liu, Thai-Hoang Pham, Meng Wu, Ping Zhang
The paper proposes PAIR, a method that treats reasoning paths of large language models as phase‑structured trajectories within each question. By sampling multiple trajectories per question, aligning them to shared relative phases, and comparing successful versus unsuccessful paths only within the same phase, PAIR isolates path‑quality signals from question‑level variation. Experiments show that standard correctness probes lose predictive power under this within‑question evaluation, while PAIR improves trajectory ranking, Best‑of‑N selection, and enables phase‑wise steering of generation outcomes.
By Zhenghao He, Guangzhi Xiong, Sanchit Sinha, Bohan Liu, Wenqian Ye, Aidong Zhang
MAADBench is a refreshable benchmark for anomaly detection in multi‑agent systems powered by large language models. It addresses the challenge of keeping benchmarks current by sampling and coupling generative tasks, generating trace data under configurable LLM backbones, and automatically providing deterministic step‑level labels. The authors evaluated 25 anomaly‑detection methods on 5,200 labeled traces, finding that existing approaches depend heavily on supervision, struggle with subtle MAS‑specific anomalies, and lack robustness across different LLM backbones.
By Lei Ma, Dennis Hofmann, Haowen Xu, Joshua DeOliveira, Peter VanNostrand, Lei Cao, Elke Rundensteiner
The paper introduces BRIDGE, a bilevel optimization framework that jointly trains a large language model (LLM) and a retriever for agentic reinforcement learning (ARL). It demonstrates that adapting the retriever before the policy yields better rewards, and that BRIDGE outperforms existing methods on seven open‑domain QA benchmarks and medical QA tasks, achieving significant gains in accuracy and reasoning quality.
By Quan Xiao, Mingda Liu, Gaowen Liu, Katsuki Fujisawa, Tianyi Chen
MemEvo presents an automated approach to discover memory mechanisms for streaming video understanding. It frames memory design as a search over executable programs expressed in a lightweight domain-specific language, and uses a pretrained large language model to generate and refine candidate memory programs. The framework evaluates each candidate deterministically while keeping the underlying vision-language model frozen, ultimately producing a training-free, bounded-memory mechanism that shows strong performance on StreamingBench and OVO-Bench.
By Guohong Liu, Jialei Ye, Shanhui Zhao, Yunxin Liu, Yuanchun Li
Code4Scene is a benchmark that evaluates coding agents on constructing and editing 3D scenes in Unreal Engine. It tests agents on two tasks: construction, where they must build a scene from open‑ended language, and editing, where they must recover a target scene from reference images while preserving everything else. The benchmark measures task fulfillment, artifact integrity, and physical validity, revealing that construction and editing performance are correlated but not interchangeable, with agents struggling most with spatial composition and precise edits.
By Xiaokang Ye, Siddhant Hitesh Mantri, Zimeng Chen, Edward Zhang, Zhaoxu Zheng, Yuanheng Li, Yizhao Chen, Tianyang Huang, Lianhui Qin
The paper explores using large language models (LLMs) to replace traditional compiler backends, a process termed AI lowering. An LLM agent translates Triton kernels directly into NVIDIA PTX, achieving 0.83x–3.34x the performance of autotuned Triton on a variety of GPUs and ML kernels. The study also extends a PTX verifier to support modern GPU features, highlighting the potential for AI compilers to reduce engineering effort for new hardware.
By Fran\c{c}ois Costa, Charly Castes, Thomas Bourgeat, Azalia Mirhoseini
The paper investigates how large language model agents that use tools respond to changes in plan priorities versus default plan removal, a phenomenon termed the "default trap." Experiments across 3,200 decision windows on Retail, Airline, and AgentDojo tasks show that switching priorities strongly redirects model choices, while removing a default plan yields weaker responsiveness. Additional studies reveal that the order of account lists and the presence of extra text significantly influence default target selection and priority effects, with overall task success varying from -19.4 to +8.3 points relative to no plan.
By Xueqi Li, Jingjie Ning, Yibo Kong
The paper introduces Robust Weighted Profile-Loss Change Point Detection (RWCP), a method for locating authorship transitions in mixed human‑LLM documents using detector scores with varying reliability. RWCP combines capped reliability weights, Huber profile gains, and a narrowest‑over‑threshold search to express the population gap as a merge cost, enabling recovery of change points without a closed‑form nonlinear center. Experiments on five benchmark families show that core RWCP reduces WindowDiff by 17.6% compared to weighted change‑point detection, and an extended variant RWCP‑R further improves performance, especially for isolated changes.
By Wan Tian, Zhongyi Li, Yawen Li, Rui Zhang, Yijie Peng, Fuzhen Zhuang