MIRAGE is a controlled study that examines how multimodal personal agents use historical evidence when conversation state changes. The study keeps evidence, questions, and scoring constant while varying only the conversation state, then checks if agents can determine answerability, recover the correct source, and answer from it. Results across seven multimodal backbones show distinct failure regimes before and after compaction, heavy reliance on context continuity by open-weight models, and mixed effects of retrieval pressure on source attribution.
By Yu Liu, Wenxiao Zhang, Cheng Hu, Cong Cao, Fangfang Yuan, Xinyu Wang, Jin B. Hong, Yanbing Liu
The paper introduces Comparison-based Preference Optimization (ComPO), a zeroth-order method that aligns large language models with human preferences using comparison oracles instead of direct differentiable loss optimization. It provides theoretical convergence guarantees for both offline and online variants under smoothness, gradient sparsity, and oracle compatibility assumptions, and establishes performance bounds under local coverage and in-distribution reward accuracy. Experiments on several LLMs (Mistral, Llama, Gemma-2, Qwen3, Gemma-3) show that ComPO outperforms existing direct alignment methods, achieving higher length-controlled win rates and diagnostics that suggest mitigation of likelihood displacement.
By Peter Chen, Xi Chen, Wotao Yin, Tianyi Lin
This survey reviews recent advances in converting non‑invasive EEG signals into images, text, and audio using generative AI techniques such as GANs, VAEs, transformers, and diffusion models. It summarizes datasets, feature‑encoding methods, evaluation metrics, and key challenges, noting that EEG‑to‑image models mainly use encoder‑decoder architectures, EEG‑to‑text leverages transformer language models, and EEG‑to‑audio maps signals to mel‑spectrograms for vocoder synthesis. The paper highlights the limitations of small, heterogeneous datasets, poor cross‑subject generalization, and the lack of standardized benchmarks, while providing open‑source resources to support reproducible research.
By Shreya Shukla, Jose Torres, Akshaj Murhekar, Christina Liu, Abhijit Mishra, Jacek Gwizdka, Shounak Roychowdhury
EvoUndo is a framework that enables large language model agents to self‑evolve—modifying prompts, tools, and execution harnesses—while ensuring that these changes can be reliably reversed across different states. The study evaluates EvoUndo on 600 unseen one‑shot self‑evolution tasks, finding that 197 capability‑improving mutations fail recoverability checks. By extending the recovery language and adding exact state‑address diagnostics, the framework recovers up to 191 out of 197 failures, demonstrating that robust self‑evolution requires co‑designing verification, grounding, witness semantics, and recovery expressivity.
By Tanmay Sah, Dolly Sah, Harshul Jain, Tanya Sah
The paper investigates how Large Language Models can be used to approximate domain expert priors for Bayesian Networks by extracting probabilistic knowledge about real‑world events. Experiments on eighty publicly available networks across domains such as healthcare and finance show that LLM‑derived conditional probabilities outperform random, uniform, and next‑token baselines. The authors also demonstrate that these LLM‑generated priors can refine data‑driven distributions, especially when data is scarce, and provide the first comprehensive baseline for evaluating LLM performance in probabilistic knowledge extraction.
By Aliakbar Nafar, Kristen Brent Venable, Zijun Cui, Parisa Kordjamshidi
The study examines how language models (LMs) alter the expressed certainty of statements when rewriting text, a process termed certainty distortion. Using an LM‑based metric aligned with human judgments, the authors find that up to 75% of LM outputs exhibit such distortion, with most models more likely to inflate certainty than reduce it. Repeated paraphrasing can amplify this effect, especially in medical contexts, and while prompt interventions help, they do not fully eliminate the bias.
By Catarina G Belem, Shang Wu, Hongyu Yao, Mark Steyvers, Sameer Singh, Padhraic Smyth
EvolveTrade is a self‑evolving framework that treats the system prompt of a tool‑using LLM trading agent as a text‑parameterized policy. After each update interval, a Policy Agent revises this policy using accumulated decision traces and portfolio feedback while keeping the backbone LLM fixed, allowing the agent to refine its information‑acquisition and portfolio‑construction procedures over time. Experiments across multiple market regimes and two LLM backbones show that EvolveTrade often improves Sharpe Ratio and Cumulative Return over fixed‑policy baselines, with behavioral analyses indicating increased code‑mediated analysis and regime‑relevant computations.
By Sehee Kim, Yumin Choi, Minki Kang, Sung Ju Hwang
GraphEcho is a benchmark that examines how large language model agents navigate graph paths and handle evidence redundancy. It tests whether agents treat repeated encounters as additional corroboration by varying path counts and evidential origins while keeping evidence content constant. The study finds that redundant paths increase repeated walks, and that provenance-aware post‑training can reduce revisits but may limit source diversity, revealing a gap between efficient exploration and effective evidence use.
By Sikun Wang, Yixi Zhou, Lei Fan, Fan Zhang
The paper evaluates three approaches for emotion recognition in conversation— a low‑cost stacked ensemble, an off‑the‑shelf LLM prompt, and a confidence‑gated hybrid that escalates only uncertain ensemble predictions to the LLM. Across three datasets (IEMOCAP, MELD, CMU‑MOSI), the hybrid consistently outperforms each pure system, achieving higher weighted F1 scores while routing most traffic through the inexpensive ensemble. This results in significant cost savings (≈$10‑85 per million utterances) and provides an interpretable escalation signal tied to emotion or sentiment shifts.
By Sai Babu Udayagiri, Arjun Chouhan, Ravisekhar Kanagala, Trishala Pavagada
The paper presents a mechanistic analysis of how large language models solve grade‑school math word problems. It identifies a four‑stage sequential pipeline—Schema Abstraction, Operation Planning, Operand Binding, and Computation—each represented in distinct layer bands. The study shows that inserting an irrelevant clause disrupts the Operation Planning stage, pinpointing the cause of failure to specific attention heads.
By Zhongdi Qu, Carla P. Gomes
The paper presents a Pareto atlas of LLM inference optimizations, mapping cost, quality, and latency trade‑offs for Qwen2.5‑7B‑Instruct on L4, A100, and H100 GPUs. Using 54 measured configurations and a calibrated simulator, it identifies 18 of 36 setups on the Pareto frontier, showing that combined methods outperform single ones. Quality tests reveal that AWQ 4bit and FP8 weights offer significant latency reductions while largely preserving accuracy, but naive FP8 KV caching fails to answer any questions correctly.
By Srikanta Datta Tumkur, Jay Iyer, Mehar Simhadri, Sai Pavan Kumar, Sai Kapil Kumar, Ramesh Nampelly
The paper proposes an epidemic model to explain how a multi‑agent system can shift from a single accidental deviation to a collective loss of control. It identifies accidental mutation, contagion through communication, and recovery as key mechanisms, and demonstrates that unsafe trajectories can spread rapidly among agents, leading to high harm rates in injected scenarios. The study also highlights the importance of auditing communication paths and strengthening both prevention and recovery measures to mitigate such risks.
By Xiangfan Wu, Zonghao Ying, Huiyu Wu, Xing Zheng, Huangsheng Cheng, Xiaorong Shi, Jing Guo
WetRobo is a reproducible robot kit designed to enable wet‑lab researchers to delegate tasks to coding agents without teleoperation or neural‑network training. The kit includes a robot arm, essential lab equipment, pre‑recorded teleoperation demos, and a skill file, allowing a coding agent to observe the lab, write, and execute programs using external tools. Experiments with OpenAI Codex on tasks such as lifting a Petri dish lid, removing a bottle cap, and opening an incubator door demonstrated successful performance in two different laboratories, outperforming a fine‑tuned vision‑language‑action policy that failed to transfer.
By Yuna Oikawa, Kei Endo, Takanori Uzawa, Yunzhe Zhang, Manan Anjaria, Lerrel Pinto, Sherry Yang, Koji Tsuda
ContrAgent is a contract‑based framework that provides symbolic temporal supervision for large language model agents. It records an agent’s tool‑call sequence as a trace of checkable predicates and formalizes desired behaviors with assume‑guarantee contracts expressed in linear temporal logic over finite traces (LTLf). Each contract is compiled into a deterministic finite automaton that both gates actions online and evaluates recorded traces offline, enabling deterministic, reproducible verdicts and significantly lower per‑call latency compared to existing LLM‑judge and rule‑based guardrail baselines.
By Yifeng Xiao, Pierluigi Nuzzo
BENCHCOMPASS is a new payment‑domain benchmark that transforms typed evidence packs into scenario‑grounded tasks, applies LLM‑based quality checks, generates attack variants, and reserves final item admission for domain experts. It includes an expert‑reviewed Pro benchmark covering payment knowledge, context‑grounded scenario reasoning, and attacked open robustness, plus a lower‑assurance Normal pool. Across 16 model variants, BENCHCOMPASS reveals distinct failure modes—missing payment knowledge, incomplete reasoning, and failure to reject invalid workflows—while the best model scores 89.6% on open context‑grounded reasoning and 81.7% under attacked inputs.
"whyItMatters":"The benchmark provides a structured way to isolate and evaluate specific weaknesses in LLMs for payment operations, a critical financial infrastructure where rules change rapidly and decisions depend on complex contextual factors."
By Sijie Dong, Wei Ren, Xuanwei Hu, Jiawei Luo, Zifan Wang, Xiaoyun Feng, Hui Cai, Lyuxin Xue, Peng Lu, Jianshe Li, Xin Zhang, Wei Wu
The paper introduces "cunning questions"—non‑safety prompts that contain misleading premises or subtle inconsistencies—to train large language models (LLMs) to scrutinize underlying intent and assumptions. Experiments show that incorporating these questions improves robustness against out‑of‑distribution jailbreak attacks and enhances subsequent safety fine‑tuning, achieving a new state‑of‑the‑art reduction in mean ASR from 17.40% to 15.05% across nine backbone–benchmark combinations. The authors argue that this training fosters vigilance, enabling models to prioritize safety judgments before engaging in harmful planning.
By Youjia Wang, Lin Xu, Yang Sun, Yuxiao Lu, Chengfang Fang, Jie Shi
The paper evaluates how well current Large Language Models can translate natural language goals, written by video game testers, into well‑formed PDDL targets for classical planning. Using a carefully designed prompt template, six state‑of‑the‑art LLMs were tested on correctness, speed, and error tendencies with real‑world benchmarks. All models achieved high correctness (>92%), with Gemini 2.5 Flash reaching 96% accuracy and the fewest false positives, while GPT‑4.1 was the fastest, yet differences in performance and occasional failures due to ambiguity and domain limits remain.
By Tomas Balyo, Lukas Chrpa, G. Michael Youngblood
CERA-MoA introduces a co-evolving router and independently learning agents for Mixture-of-Agents systems, allowing routing strategies to adapt to evolving agent capabilities. The framework uses a predictive familiarity estimator based on mid-layer hidden states to evaluate semantic competence, and a cumulative-threshold adaptive routing mechanism to activate a minimal agent subset. Experiments across multiple domains show that CERA-MoA outperforms static-agent routing and fixed-workflow fine-tuning baselines.
By Jiaxuan Jiang, Liyuan He, Zhixuan Fang
The paper introduces Gauntlet, a framework that lets large language models autonomously build game-playing agents from a bare contract—just a game description, raw observation/action interface, and an empty policy file. In a single session, the model experiments with the game, compiles a standalone controller, and the resulting program is evaluated on held‑out instances without further model calls. The authors demonstrate that these compiled agents can win full‑scale games such as StarCraft II and Civilization, marking the first time a language‑agent system has achieved standalone victory in such complex titles.
By Joey Xiao, Haonan Huang
The Flag Game is a toy model designed to study how AI agents form collective beliefs. In the game, each agent sees only a private crop of a hidden country flag and can share beliefs with peers, leading to complex phenomena such as non‑monotonic performance scaling, accuracy gains from social awareness, and polarization that degrades performance at large population sizes. The authors introduce social circuit attribution to identify key agents and views, and develop a statistical mechanical theory to explain collective belief collapse and polarization in larger populations.
By Elizabeth Pavlova, Hidenori Tanaka