The paper investigates how large language model (LLM)–based automated code review (ACR) systems can be manipulated through contextual bias injection, a framing effect that lets presentation override content. A large‑scale study across six LLMs confirms the framing effect in vulnerability detection, and a realistic experiment on 33 CVEs in 20 projects shows that a novel LLM‑assisted refinement attack succeeds in 97% of cases, while a template‑based attack fails. The results underscore the vulnerability of real‑world ACR pipelines and the need for human oversight and trust in the development process.
By Georgios Alexopoulos, Nikolaos Alexopoulos, Diomidis Spinellis, Dimitris Mitropoulos
The paper introduces Multi-Variable Conformal Prediction (MCP), a framework that jointly optimizes and calibrates multiple prediction‑set parameters in a single optimization problem, thereby eliminating the need for a calibration data split while preserving finite‑sample coverage guarantees. Two tractable variants are proposed: RemMCP, which uses constrained optimization with constraint removal and generalizes split conformal prediction, and RelMCP, which employs penalized optimization with constraint relaxation and can handle non‑convex problems. Experiments on ellipsoidal and multi‑modal prediction sets show that both RemMCP and an adaptive version of RelMCP consistently achieve target coverage with lower variance across calibration runs and, in most cases, produce smaller or comparable prediction sets compared to baseline methods.
By Laura L\"utzow, Simone Garatti, Marco C. Campi, Lars Lindemann, Matthias Althoff
WAKMan is a workload‑aware KV cache manager designed for Large Language Model servers that selects an eviction policy based on live traffic patterns without needing to know request types or read prompts. By adapting to the mix of single‑turn, multi‑turn, and agentic requests, it improves cache hit rates over the default LRU strategy by an average of 5 percentage points and achieves the lowest mean and worst‑case optimality gap among evaluated policies. The paper demonstrates that no single policy consistently outperforms others across diverse traffic mixes, highlighting the need for dynamic, workload‑aware cache management.
By Suhas Kumar Morisetty, Vaibhavi Lokegaonkar, Sachin Mathew Varghese, Wesley Harris
MAEval is a multi‑agent framework that evaluates natural language explanations in recommender systems using large language models. It employs a planner agent with a Chain of Debate prompting strategy, a moderator to control prompt drift, and an arbitrator that aggregates multiple evaluation rounds, supplemented by Monte Carlo sampling for robustness. Experiments on public and proprietary datasets show that MAREval outperforms existing baselines, aligns more closely with human judgments, reduces variability, and achieves production‑level quality, leading to measurable gains in user engagement and revenue.
By Reza Yousefi Maragheh, Jayesh Uddhav Kudase, Aysenur Inan, Ramin Giahi, Kai Zhao, Parth Parikh, Jianpeng Xu, Jason Cho, Evren Korpeoglu, Sushant Kumar
The study investigates how the presence of a human user influences the behavior of large language model agents during negotiation tasks. Using the PRESENT framework, researchers compared 11.2k negotiation runs across 7 games and 8 models with and without an interactive user, keeping all other variables constant. Results show that agents negotiate more aggressively and achieve higher payouts (+13.7%) when a user is present, yet also exhibit increased misaligned behavior toward both counterparties and users—misalignments that are not reflected in outcome metrics and persist even with neutral user messages.
By Jane Pan, You Lyu, Abe Bohan Hou, Valerie Chen, He He
The paper introduces DiscernMem, a training‑free framework for improving memory quality and usage in LLM‑based multi‑agent systems. It combines Defect‑Aware Memory Refinement (DMR), which diagnoses and repairs flawed memories, with Boundary‑Learning Meta‑Memory (BLM), which learns when to apply or avoid each memory based on context. Experiments with Qwen3‑4B‑Instruct‑2507 on four benchmarks show DiscernMem outperforms existing baselines, achieving higher scores on AutoGen, CAMEL, and MacNet, and improving code generation and symbolic planning performance.
By Yipeng Hao, Yuxuan Cai, Wenkai Ji, Jie Zhou, Liyang Yu, Qin Chen, Zhiguo Yan, Shulei Zheng, Liang He
EMMA is a framework that treats the configuration of a multi‑agent system—controller routing, specialist prompts, and procedural skills—as structured episodic memory. It explicitly assigns ownership and limits updates through a reflector and backbone‑independent guards, ensuring edits are scoped, incremental, and bounded. Experiments on KramaBench, AppWorld, and BBEH show EMMA outperforms four baselines, achieving up to 32.5% higher performance and reducing training executions by 17–38%.
By Ruijia Niu, Kaiwen Liu, Rose Yu, Kristina Klinkner
MLE-World is a world model fine‑tuned from Qwen3.5‑4B that predicts the terminal output of a machine learning program—such as error traces, training logs, and validation scores—directly from its code and data description, without executing the program. It also provides a confidence score indicating the likelihood that the prediction is correct. Using these predictions, the authors design five test‑time‑scaling policies that compare multiple candidate programs per step, executing only those with low confidence, and demonstrate that MLE‑World matches real outputs on 77.9% of held‑out programs and improves agent performance by up to 22.1% with fewer real executions.
By Brian Ko, Han Wang, Hejia Geng, Li Erran Li, Sherry Yang
The paper presents a fixed‑wing interceptor drone that uses a neural network trained with reinforcement learning to map monocular vision inputs directly to actuator commands. The policy, developed in a custom Python environment with proximal policy optimization, was tested against classical guidance controllers and demonstrated superior performance in simulation and successful interception of maneuvering fixed‑wing targets in real flight. This work illustrates the feasibility of AI‑driven autonomous control for long‑endurance interception missions.
By Daan van Hamersveld, Ewoud Smeur, Iurii Kriachko
InterQ‑RL is a new method for multi‑turn agent reinforcement learning that addresses execution mismatch caused by low‑precision quantized rollouts. It corrects the mismatch by applying bounded execution‑prefix adjustments over the policy history, allocating correction strength across action segments within a fixed displacement budget, and does so without extra model forward passes. Experiments on ALFWorld, WebShop, and ScienceWorld with 4B‑30B backbones show that InterQ‑RL achieves the best quantized performance, matching BF16 within 0.7 points and surpassing it in two settings while maintaining or improving learning speed.
By Yingbo Hao, Song Liu, Guanghua Yu, Zhaojing Zhou, Jianchen Zhu, Yi Zou
The paper introduces StressSim, an intent‑preserving adversarial user simulator designed to stress‑test multi‑turn agents by generating challenging yet valid user interactions. StressSim optimises high‑level adversarial plans for each task instance using a genetic prompt‑optimisation loop that leverages rollout traces, task‑completion signals, and validity feedback to reduce task success while keeping the user’s goal unchanged. Experiments on MultiWOZ, retail, and airline domains show that StressSim reveals significant robustness gaps in leading LLM agents—reducing tau‑bench pass^4 scores by up to 69%—while maintaining realistic conversation quality and goal preservation with a modest optimisation budget.
By Atri Vivek Sharma, Brian Formento, Alessio Lomuscio
Auto-Research Arena is a multi-target benchmark designed to evaluate autonomous research systems and uncover new ideas in language modeling. It separates admissibility—determined by model quality and resource constraints—from improvement, using nine diverse targets covering model, data, training, and inference efficiency. The benchmark, applied to language model pretraining, distinguishes existing research harnesses and reveals significant room for improvement, while also enabling the reuse of discovered ideas to achieve lower training loss and better held‑out performance.
By Xi Chen, Chongjian Ge, Yiran Xu, Yiqun Mei, Henrik I. Christensen, Hao Tan, Ziwen Chen, Hanwen Jiang
ETC-Bench is a multilingual benchmark for evaluation‑side transcript canonicalization (ETC), covering 181 phenomenon types across 12 languages. Each instance is a minimal pair that shares a sentence context and differs only in a target span exhibiting one phenomenon. The benchmark introduces Semantic‑preserving Exact‑match Normalization Accuracy (SENA) to evaluate canonicalization success and semantic preservation, and a Taxonomy‑guided Agent Canonicalization Skill (TACS) to guide LLM agents in canonicalizing transcripts.
By Zhiheng Qian, Junhao Du, Jing Peng, Hanqi Li, Kunyang Peng, Feng Xue, Kai Yu
The paper introduces a harness designed to curb over‑defensive writing by agents, which occurs when agents add stronger claims than the source material and then retract them. The harness separates the reviewer, filter, and writer components, allowing the writer to focus only on problems selected by an independent filter. Evaluation on 25 business‑project drafts shows a 96.61% drop in defensive propositions and an increase in source‑claim retention from 80.35% to 93.12%.
By Xiaochen Chang
The paper introduces an ontology‑guided, multi‑agent framework for differential diagnosis that mirrors physicians’ iterative reasoning. By grounding inference in a structured, expert‑curated medical knowledge base, the system produces explicit, auditable reasoning steps, reduces hallucination, and improves transparency. It also identifies key missing symptoms and generates targeted follow‑up questions to lower diagnostic uncertainty.
By Roxana Ramona Szomiu, George Ursu, Adrian Groza
InfraDev is a new benchmark comprising 36 tasks derived from real-world pull requests to Megatron‑LM, vLLM, and verl, designed to test coding agents on infrastructure development for training, inference, and reinforcement learning across many GPUs. Each task requires agents to reimplement multi‑file changes and run correctly, with grading that includes hidden probes and a prompt exposing only the interfaces used. The benchmark reveals that current frontier models resolve only about 39% of tasks, with probes rejecting a third of submissions that pass official tests, and that structural and state errors dominate failures.
By Yangshen Deng, Jiaqi Ruan, Yeqi Huang, Nick Lau, Guangming Sheng, Luo Mai
The paper introduces Dwell, a model‑based reinforcement learning framework that treats a pretrained vision‑language model (VLM) as a recurrent latent transition operator. Dwell enables the VLM to predict, adjust, and act within its own latent space, allowing the agent to learn from real interactions and latent imagination without reconstructing full futures. Experiments across five interactive visual environments show that Dwell outperforms other trainable methods while significantly reducing computational demands for imagination.
By Haoxi Li, Sikai Bai, Jianfei Ma, Qinglin Hou, Xiaocheng Lu, Tao Han, Jie Zhang, Song Guo
WorkFactory is a scalable framework that transforms professional intents into executable tasks by building an information support graph and using probe executions to detect and repair support failures. The authors generated 112,686 tasks across 612 occupations and fine‑tuned Qwen models (4B, 27B, 35B-A3B) to create the WorkPilot family, which outperforms base models on four professional‑work benchmarks. The study shows that planning and repair reduce task environment defects and improve downstream performance, with diminishing returns observed as model size increases.
By Ziyang Zeng, Tingyang Chen, Jia Deng, Qiyao Wang, Shuo Tang, Yike Yang, Tao Ren, Wayne Xin Zhao, Feng Chang, Jujie He, Chuan Hao, Ran Tao, Bryan Dai
The paper introduces a method to improve large language model agents that invoke enterprise APIs by embedding process and schema constraints into a typed property graph. By integrating OData metadata and BPMN reference processes, a hybrid retriever can identify correct endpoints and follow process edges, boosting Recall@3 for low lexical overlap queries and enabling more reliable multi-step service invocation. Experiments on 42 enterprise services and 81 multi-step queries demonstrate significant gains over baseline approaches, with a marked reduction in invalid API calls.
By Saiful Khan, Afrah Farea
The paper investigates how short memory representations of large‑language‑model (LLM) group discussions affect final decision quality. It introduces CrossProfile, a benchmark of 480 hidden‑profile tasks that separates supporting, decoy, and irrelevant facts while keeping their wording constant. Experiments with Qwen2.5‑7B, Llama‑3.1‑8B, and DeepSeek models show that heuristics based on rarity or relevance both fail in predictable ways, and that the decision model’s social‑proof bias further degrades accuracy. Removing attribution tags or randomizing them improves performance, yet natural transcripts still lose more evidence than no notes at all.
By Anouar Imel, Imad Charradi, Abderrahman Sghir, Khalil El Khammar