The study investigates how different message formats affect the fidelity of information as it passes through multiple LLM agent relays. Using a controlled testbed, the authors encode twelve atomic facts in five formats (free natural language, precision‑instructed NL, JSON, triples, key‑value) across six hops and evaluate recall against programmatic ground truth. Results show that strong relays maintain near‑lossless recall for all formats, while weaker relays exhibit significant format‑dependent recall loss, and that any injected error is faithfully propagated across all formats without causing collateral damage.
By Sicheng Zeng
AntiGrounding is a visual action-selection framework that turns short robot trajectories into both executable motion plans and rendered prompts for vision‑language model evaluation. After filtering for feasibility, each trajectory is scored on safety, task alignment, efficiency, and physical plausibility using structured multi‑view visual question answering, and the best trajectories are refined and validated by a digital twin before real‑world execution. In eight real‑world manipulation tasks, the system achieved a 71.25% success rate with a single GPT‑6 Astra evaluator, outperforming baseline methods.
By Wenbo Li, Yiteng Chen, Wenhao Li, Qingyao Wu
The paper introduces the first watermark designed specifically for diffusion language models (DLMs), which generate tokens in arbitrary order unlike traditional autoregressive models. It overcomes the challenge of missing prior tokens by applying the watermark in expectation over the context and promoting tokens that strengthen the watermark when used as context. Experiments show a >99% true positive rate with minimal quality loss and comparable robustness to existing autoregressive watermarks.
By Thibaud Gloaguen, Robin Staab, Nikola Jovanovi\'c, Martin Vechev
BuildBench introduces a realistic benchmark for evaluating large language model agents on the task of compiling open‑source software (OSS). It includes diverse OSS projects that lack clear build instructions, have undocumented dependencies, and may require source patching or script modification. The authors also present OSS‑BUILD‑AGENT, a baseline LLM‑based agent that retrieves build instructions effectively and achieves state‑of‑the‑art performance on the benchmark.
By Zehua Zhang, Ati Priya Bajaj, Divij Handa, Siyu Liu, Arvind S Raj, Hongkai Chen, Hulin Wang, Yibo Liu, Zion Leonahenahe Basque, Souradip Nath, Vishal Juneja, Nikhil Chapre, Tiffany Bao, Yan Shoshitaishvili, Adam Doup\'e, Chitta Baral, Ruoyu Wang
MAS-Shield is a defense framework for Large Language Model–based Multi-Agent Systems that uses a coarse‑to‑fine filtering pipeline. It first selects critical agents, then applies lightweight auditing to most cases, and finally escalates only suspicious signals to a heavyweight committee. Experiments show a 92.5% recovery rate against adversarial attacks and a latency reduction of over 70% compared to existing methods.
By Kaixiang Wang, Zhaojiacheng Zhou, Bunyod Suvonov, Jiong Lou, Zihan Wang, Yuxiang Zheng, Yidan Lin, Wutong Zhang, Xianghan Kong, Chentao Wu, Jie Li
The paper presents a framework for translating natural‑language queries into Kusto Query Language (KQL) using small language models (SLMs). It introduces lightweight retrieval, error‑aware prompting, LoRA fine‑tuning with rationale distillation, and a two‑stage architecture that pairs an SLM drafter with a low‑cost LLM judge. Evaluations on Microsoft’s NL2KQL Defender dataset show the two‑stage approach achieving high syntax and schema‑valid accuracy while dramatically reducing cost compared to larger LLM baselines.
By Saleha Muzammil, Rahul Reddy, Vishal Kamalakrishnan, Hadi Ahmadi, Wajih Ul Hassan
The paper introduces VLM-CAD, a workflow that uses Vision Language Models (VLMs) for analog circuit sizing while mitigating spatial blindness and logical hallucinations. It incorporates a neuro‑symbolic parsing module, Image2Net, to convert schematics into topological graphs and JSON, and an Explainable Trust Region Bayesian Optimization method, ExTuRBO, to guide design decisions with sensitivity evidence. Experiments on 12 sizing tasks across six circuits and four technology platforms show a Strict Pass@1 of 23.3% and a Relaxed Pass@1 of 91.7%.
By Guanyuan Pan, Shuai Wang, Yugui Lin, Tiansheng Zhou, Pietro Li\`o, Zhenxin Zhao, Yaqi Wang
This study investigates whether large language models (LLMs) can reliably answer scientific questions and how susceptible they are to manipulation by fringe scientific material. The authors modified custom LLMs to prioritize knowledge from selected fringe papers on the Fine Structure Constant and Gravitational Waves, then compared their responses with those of domain experts and standard LLMs. The altered models produced fluent, convincing answers that contradicted scientific consensus and were difficult for non-experts to detect as misleading, demonstrating that LLMs are vulnerable to manipulation and cannot replace expert judgment.
By Harry Collins, Hartmut Grote, Paul Newbury, Patrick Sutton, Simon Thorne
The paper introduces a systematic method for comparing different formalizations of the same legal provision by analyzing their inferences on individual cases. It matches formalizations at the node level, derives shared interfaces, and uses a SAT solver to identify edge cases where any two formalizations disagree. The authors apply this approach to ten EU provisions formalized by nine advanced LLMs, finding that behavioral divergence is largely uncorrelated with structural agreement and that the resulting edge cases expose distinct types of disagreement, some reflecting real legal controversies.
By Julius Vernie, Matthias Grabmair
The paper investigates whether key‑value (KV) cache eviction strategies should vary across Transformer layers. By combining existing eviction methods in different layer configurations and profiling their performance, the authors find that heterogeneous, layer‑wise routing consistently outperforms homogeneous policies on LongBench tasks. Even with a fixed set of methods, the placement of each method strongly influences overall quality, and a single well‑chosen route surpasses all nine standalone baselines across multiple cache budgets.
By Chao Fei, Kaihua Liang, Hanzhi Hu, Hongcheng Guo, Jian Weng, Marco Canini, Panos Kalnis
The paper presents an AI-based method that models human behavior by assigning a language model a vector of trait intensities—called a type vector—and asking it to predict actions in various settings. By adjusting traits such as Altruism, Risk Aversion, Fairness, and Trust, the authors fit the model to 119,147 decisions from 78,657 subjects across 35 countries and 10 economic games, finding that three dimensions (Risk Aversion, Strategic Sophistication, and Trust) closely match human choices. The resulting type vectors cluster into fewer than a dozen groups and can predict behavior in new games with different rules, demonstrating the method’s generalizability and interpretability.
By Matthew O. Jackson, Benjamin S. Manning, Yutong Xie, Walter Yuan, Qiaozhu Mei
The paper introduces M2Tok, a Multi-head Multi-codebook Action Tokenizer that reduces reconstruction loss for continuous action signals by decomposing latent features into multiple heads and assigning independent codebooks to each. This design expands representational expressivity, leading to lower reconstruction error and higher success rates in Vision‑Language‑Action models evaluated on RoboTwin, Simpler‑Env, and zero‑shot real‑world tasks. Ablation studies confirm the effectiveness of both multi‑head and multi‑codebook mechanisms.
By Chunpu Xu, Zhixuan Liang, Yuhao Zhang, Chi-Min Chan, Jessie Wang, Yang Xiao, Mengkang Hu, Xiaokang Yang, Yao Mu
The paper introduces Stiefel Attention, which constrains the query and key projection matrices of transformers to the Stiefel manifold and optimizes them with a Riemannian Adam variant. It demonstrates that this approach yields steepest‑descent updates, is well‑conditioned, and preserves learned attention geometry during weight decay. Empirical results show significant accuracy gains on modular arithmetic grokking and CIFAR‑10 patches, with the improvement attributed to a step‑scale‑free update rule rather than equivariance or projector changes.
By Rub\'en Dar\'io Guerrero
The paper introduces KAINN, a hybrid neural‑mechanistic model that augments the Agriculture‑informed Neural Network with domain knowledge on fertilizer diffusion, soil respiration, and water‑filled porosity to predict nitrous oxide emissions from agriculture. Experiments across CNN, LSTM, and Transformer architectures show that KAINN achieves lower root mean square error, lower mean absolute error, and higher R-squared values compared to purely data‑driven models and the original AINN. The learned interfaces exhibit smoother, more physically consistent parameter trajectories with reduced uncertainty.
By Ci Lin, Futong Li, Rose Chong-Wu, Tet Yeap, Iluju Kiringa
AURA: Adaptive Uncertainty-Routed Analysis for Email Threat Detection is a multimodal system that evaluates both email content and embedded URLs to detect spam and phishing. It uses a two-layer approach: first, a URL classifier estimates prediction uncertainty, and only messages with high uncertainty are passed to a fine-tuned transformer encoder for deeper semantic analysis. Evaluated on eight diverse training corpora and two real-world datasets covering a decade of attacks, AURA achieves a macro F1-score of 0.9858 in-distribution and maintains scores above 0.94 on the NazPhish-Eval and GuenterTrap-Eval datasets, demonstrating strong generalization to new attack scenarios.
By Omran Berjawi, Walid fahs, Rida Khatoun
The paper presents a modality‑agnostic, hierarchical Transformer framework for assessing cognitive workload using heterogeneous biosignals. In a pilot study, the authors evaluated all 31 combinations of five modalities (ECG, EDA, RESP, SpO₂, EEG) across three tasks (IQ, MATH, GAME) and found that EEG alone performed best, while adding more modalities did not consistently improve results. The full five‑modality model achieved the highest average accuracy (73.02% on IQ, 68.08% overall) and reduced model size by about 50% compared to late‑fusion approaches.
By Stefanos Gkikas, Christian Arzate Cruz, Calvin Joseph, Giorgos Giannakakis, Raul Fernandez Rojas
The study evaluates zero‑shot essay scoring using sub‑3B open language models that run locally on a single 8 GB consumer GPU. Four instruction‑tuned models (Qwen2.5‑0.5B, 1.5B, 3B and SmolLM2‑1.7B) were tested on all eight ASAP‑AES prompts, comparing rubric‑decomposed versus holistic prompting, different aggregation methods, and trait‑mapping strategies. Results show rubric‑decomposed prompting consistently outperforms holistic prompting, trait‑mapping is sensitive to calibration, and longer essays reduce error, yet the best local configuration (macro QWK 0.388) still falls short of human agreement and a length‑only baseline, suggesting these models are best suited for formative, human‑supervised feedback.
By Nguyen Dung Son, Dang Quang Minh, Nguyen Huu Loi, Truong Viet Vu, Nguyen Thai Anh
The paper compares the parallelism capabilities of three diffusion large language model paradigms—masked, uniform, and Gaussian diffusion. It proves that uniform and Gaussian diffusion can sample with a number of forward passes scaling with the dual total correlation of the distribution, potentially much less than the context length, whereas masked diffusion may require more passes. The study establishes a provable separation in parallelism, showing that masked diffusion’s critical windows are asymptotically narrower than those of the other two approaches.
By Sitan Chen, Liye Wang
The paper introduces a method called score centering to address the training‑inference mismatch (TIM) that destabilizes reinforcement learning for large language models. By adding an additive correction term that cancels drift between training and inference engines, score centering stabilizes RL and can match or surpass importance‑sampling techniques, especially as model size and mismatch severity increase. The approach also composes with importance sampling, yielding further performance gains in staleness experiments.
By Martin Marek, Max Ryabinin
The paper introduces ALIBI, a semantic cover story attack that injects a small, non-executed read‑only section into compiled binaries to mislead large language model (LLM) malware analyzers. By embedding a coherent but false security narrative, ALIBI can cause LLMs such as Gemini 2.5 Pro, GPT‑5.5 Pro, and Claude Opus 4.7 to downgrade or flip the verdicts of malicious samples. The attack also transfers to ELF binaries, and even a verification‑guided defense prompt only partially mitigates the effect, leaving a significant portion of malicious samples classified as benign.
By Hyeongjun Choi, Wonyoung Jung, Haehoon Seo, Sungyup Nam