The paper investigates whether complex communication topologies are necessary for effective multi‑agent debate (MAD) among large language models. It demonstrates that a simple random-without-replacement routing policy—where each agent debates with two newly sampled peers each round—consistently improves the accuracy‑cost trade‑off in sparse MAD setups. Additionally, the study shows that lightweight deliberation stopping can further reduce inference costs without sacrificing accuracy.
By Boxuan Wang, Zhuoyun Li, Xiaowei Huang, Yi Dong
The paper investigates how long‑horizon language model agents encode memory‑management signals before taking actions. By examining hidden states just prior to each action, the authors find that the model already signals the need for compression and recall, independent of context length or interaction progress, and that these signals vary across model depth. They propose the Preaction Memory with Evidence Retrieval (PaMER) framework, which uses state‑guided compression and selective evidence retrieval to reduce context consumption while preserving task performance.
By Mingxuan Wang, Guorun Yao, Fei Luo, Yinglong Guo, Chao Ning, Bo Wang, Hongyue Chen, Yanbiao Ma, Jungong Han
StateComp introduces a method for long‑horizon agents to decide when to compress historical interactions based on the current agent state, rather than relying on fixed windows or periodic schedules. The framework uses a two‑stage annotation process to create KEEP and READY labels, trains an imbalance‑aware router on frozen language model representations, and groups adjacent READY interactions into compact summaries. Experiments on WorkBuddyBench show that StateComp cuts agent and summarization tokens by 52.27% and speeds up representation extraction 12.67‑fold while preserving task performance.
By Mingxuan Wang, Hongyue Chen, Yinglong Guo, Fei Luo, Chao Ning, Bo Wang, Guorun Yao, Yanbiao Ma, Jungong Han
Omni-Decision is an omni-modal agent that uses evidence-ledger planning to manage noisy multimodal observations. It replaces the growing dialogue history with a compact evidence ledger that tracks missing, confirmed, and conflicting evidence, allowing the planner to operate on a streamlined context. The system achieves state‑of‑the‑art accuracy on OmniGAIA and WorldSense while operating at a fraction of the cost of larger models.
By Ming Ma, Yi Zhu, Yiran Zhong, Feida Zhu, Yuhao Wang, Junhan Shi, Lingrui Mei, Tianming Yang, Steven Hoi
The paper introduces the concept of an agent’s "taste"—its ability to make effective long‑horizon decisions—and presents Taste‑Bench, a new benchmark that automatically generates decision‑fork questions from agent trajectories. Taste‑Bench evaluates models on choosing the best path without seeing future outcomes, revealing that top models answer only about 60% of questions correctly and that later‑appearing evidence makes forks harder. The authors also demonstrate that training a student model to mimic a teacher’s judgment improves decision quality and overall success on held‑out software engineering tasks.
By Wenbo Pan, Zhichao Liu, Shujie Liu, Jingying Zeng, Chin-Yew Lin, Xianfeng Tang, Yan Lu, Qi He, Xiaohua Jia
The paper "Optimizing watermarks for large language models" discusses the growing importance of watermarks for generative LLMs amid concerns about misuse. It presents a systematic multi‑objective optimization framework to balance watermark identifiability with the impact on text quality. The authors identify Pareto‑optimal solutions for a broad class of robust, efficient watermarks that outperform the current default watermark.
By Bram Wouters
The article investigates how the textual description of a shared state influences the collective behavior of language‑model agents. By testing 507,112 responses across different model families on a circular coordination task, the authors show that varying the state description (e.g., numerical summaries vs. histograms) can alter whether agents align, split, or fail to coordinate. The study demonstrates that the way a shared state is described is an integral part of the interaction rule that determines collective order.
By Takahiro Ezaki, Naoto Imura, Katsuhiro Nishinari
The paper presents empirical scaling laws for autoregressive language models, linking prediction loss to model size, data size, and compute, and investigates their theoretical basis using a teacher–student linear RNN framework. In this tractable setting, a stable latent linear RNN generates trajectories while a sketched linear recurrent student is trained via full‑batch WSD gradient descent on next‑token prediction. The study derives explicit approximation, optimization, and statistical scaling laws that depend on the sketch dimension, number of trajectories, and trajectory length, revealing how different power‑law exponents for innovation and initialization covariances affect the rates and crossovers between regimes.
By Ziyan Chen, Zhongzhu Zhou, Peilin Liu, Ding-Xuan Zhou
The paper introduces FRESH, a Failure-aware Retrieval framework that uses Experience-Structured Heterogeneous graphs to transform past successes and failures into structured external memory for tool‑using agents. By explicitly modeling dependencies among tasks, actions, errors, repairs, and execution conditions, FRESH enables frozen language models to reuse reliable strategies, avoid recurring failures, and make safer decisions in stateful tool interactions. Experiments on τ‑Bench and AppWorld with multiple open‑source models demonstrate that FRESH consistently improves task success and tool‑use reliability compared to no‑memory agents and other memory‑based baselines.
By Jiaxing Li, Lei Song, Rui Dong, Youyong Kong
The paper investigates how to combine shared global key‑value (KV) caches with layer‑specific local history in decoder‑only Transformer language models. By separating historical content from the input source used to form it, the authors show that adding local history can reduce held‑out test perplexity by about 1.4% compared to a current‑token local branch, while also demonstrating benefits in capacity, entry‑count, and training‑compute controls. Experiments on a 126M‑parameter model with 2K context reveal that local history remains valuable even when adjacent layers share local inputs, and that a sufficient suffix schedule can reduce upper‑layer construction work without losing cache completeness.
By Xinglang Xian
Rolling Conformal Prediction (rolling‑CP) is a distribution‑free predictive inference method designed for sequential model training. It calibrates each incoming observation against the current predictor and incorporates it into future training, eliminating the need for data splitting. For exchangeable data, rolling‑CP guarantees marginal coverage with a universal factor‑two bound, and for i.i.d. streams it provides high‑probability training‑conditional validity over time, improving to the target level under stability conditions.
By Chen Cheng, Ruiting Liang, Rina Foygel Barber
arXiv:2609. 27008v1 Announce Type: cross Abstract: We study the long-time behavior of Wasserstein gradient flows for interaction energies \[ \mathsf E[\mu] = \frac12\iint_{M\times M}K(x,y)\,\mathrm d\mu(x)\,\mathrm d\mu(y) \] on a closed manifold $M$.
By Zhengjiang Lin, Philippe Rigollet
The paper introduces EduBehaviors, a framework that uses large language models to identify observable behaviors in educational dialogues and then trains a classifier to predict pedagogical constructs. By measuring repeated behaviors, the approach provides interpretable and scalable annotations, achieving macro‑F1 scores of 0.673 and Cohen’s kappa of 0.688 on the TalkMoves dataset. The authors also release the EduBehaviors Toolkit, enabling researchers to apply the framework to their own data.
By Julian Bernado, Ana Trindade Ribeiro, Xander Beberman, Susanna Loeb
The paper investigates how attention dynamics evolve across recurrent depth in language models, finding that attention support stabilizes early while hidden states and outputs take longer. It proposes WISE, a training‑free method that uses full attention in early steps and then reuses the discovered sparse working set for later steps, preserving performance on multi‑hop QA tasks. Experiments show that WISE maintains quality up to 2K context, offers measurable speedups, and highlights the importance of recurrent discovery of attention support.
By Ke Wan, Chen Chen
The paper introduces FFM-CP, a framework that fuses multiple pathology vision‑language foundation models for few‑shot learning. It aligns heterogeneous representations with an Orthogonal Procrustes transformation, then uses a unified graph to refine support‑image features and class prototypes across backbones. Experiments on six histopathology datasets show that FFM‑CP outperforms the best single adapted model in 50 of 54 few‑shot comparisons.
By Anh-Tien Nguyen, Trung DQ. Dang, Nghiem Tuong Diep, Bui Ngoc Han Nguyen, Tan-Ha Mai, Miriam Cindy Maurer, Phuong Hoa Nguyen, Thi Thuy Uyen Nguyen, Youngjun Park, Daniel Sonntag, Duy Minh Ho Nguyen, Anne-Christin Hauschild
The paper introduces RALCT, a lightweight Convolutional Transformer that combines randomized audio augmentations, MFCCs, and log‑mel spectrograms to extract robust features for environmental sound recognition. With only about 310,000 parameters, RALCT achieves state‑of‑the‑art accuracy—over 93% on UrbanSound8K, peaking at 94.56%—making it suitable for deployment on mobile devices. The authors also develop a mobile app that integrates the model to provide real‑time safety alerts for hearing‑impaired users.
By Julia Huang
ChronosAttack is a delay‑only scheduling attack that manipulates the timing of authentic tool responses in large language model agents, without altering the content of those responses. By introducing bounded delays, the attack can reorder identical evidence and shift the agents’ final decisions. Experiments on GPT‑5.6 Sol, Gemini 3.6 Flash, DeepSeek V4 Flash, and Claude Sonnet 4.6 show that the attack can cause strong targeted shifts in some models, large opposite shifts in others, and that a single scheduling inversion can dramatically change outcomes. Defenses based on synchronization and order‑consistency reduce the attacker’s influence over observation order.
By Arash Vashagh
The paper examines LLM-based recommendation rerankers that are often evaluated under an oracle protocol, which guarantees the ground-truth item is present in the scored set. Across Amazon datasets, this protocol overestimates realistic NDCG@10 by 92–95% because realistic retrieval only covers 2–19% of relevant items at K=100, creating a recall ceiling that limits any closed-candidate reranker's top‑k NDCG. The authors find that various optimisation strategies—including prompt engineering, model scaling, sequential models, supervised neural rerankers, LoRA fine‑tuning, hybrid retrieval, score‑aware prompting, and LLM+CF fusion—do not significantly improve over a collaborative‑filtering baseline under realistic retrieval, and they propose a Recall‑Aware Evaluation Protocol (RAEP) to better assess rerankers in low‑recall regimes.
By Zhaohui Wang
The paper introduces a method to prune six layers from the encoder of OpenAI’s Whisper ASR model, reducing the encoder stack by 18.5% without requiring custom inference code. Layers are selected based on their minimal impact on Word Error Rate when removed. After pruning, the model’s WER rises from 18.2% to 21.9%, but distillation with unlabeled monolingual speech data lowers it to 20.1%.
"whyItMatters":"The approach offers a straightforward way to accelerate Whisper inference by simplifying the encoder while maintaining acceptable accuracy, and the released code and model enable immediate adoption by the community."
By Rasmus Aagaard, Nicki Skafte Detlefsen
The paper introduces PIECE, a Parameter Importance-Driven Continual Learning method that selectively updates only 0.1% of core parameters to preserve general abilities while learning new domain knowledge. PIECE employs two importance estimators—PIECE‑F using Fisher Information and PIECE‑S combining gradient and curvature information—to guide updates. Experiments on three language models and two multimodal models demonstrate that PIECE maintains general capabilities and achieves state‑of‑the‑art continual learning performance without accessing prior training data or adding parameter overhead.
By Lingxiang Wang, Hainan Zhang, Zhiming Zheng