The paper presents a closed‑form model that compares the coordination tax of multi‑agent LLM pipelines—comprising decomposition overhead and per‑handoff protocol cost—to the heterogeneity dividend gained by routing work to cheaper models. Using the Open Floor Protocol (OFP) on 135 HumanEval and ClassEval tasks, the authors find that the protocol adds only 7.2 ms per handoff, yet a planner‑coder‑reviewer pipeline is 3.22× more expensive and slower than a single strong model call, and does not solve more tasks. The study concludes that the benefit lies in selecting the appropriate model per task rather than in pipeline decomposition, as an oracle choice between two single calls solves 94.3 % of runs at half the cost of the strong call.
By Andreas Zettl, Diego Gosmar
The paper presents a reinforcement learning framework for the 2024 roguelike deckbuilder Balatro, addressing its three main challenges: a large action space, long horizons with sparse rewards, and high stochasticity. It introduces a Rust-based simulator, a streamlined observation-action interface with dynamic masking, and a multi-agent PPO architecture that splits play into a tactical blind policy and a strategic shop policy. Training results show a sharp improvement, with the two-policy system achieving over 99% first-blind clear rates and a significant increase in episode returns, outperforming single-agent baselines.
By Ajay Vikram Periasami, Anjaneya Gupta, Ethan Cheng, Ivan Rosales
The paper introduces Proof‑Carrying Messages (PCM), a protocol that attaches verifiable evidence—Merkle log records and signed attestations—to each claim exchanged between AI agents. PCM enables receivers to evaluate claims offline, categorizing them as proven, unproven, or failed, and ensures that claims based solely on an agent’s own report can never be proven. In benchmark tests on 926 real messages, PCM accepted only 1.4% of false claims, whereas other protocols accepted many more.
By Bravish Ghosh (Cygnux Labs)
The paper investigates how self‑evolving personal agents become more vulnerable to indirect prompt injection (IPI). Using a two‑phase methodology, the authors show that when agents learn from user feedback—either through skill evolution or reinforcement learning—the attack success rate of IPI increases by 14.0–48.8 percentage points compared to a non‑evolving baseline. The study highlights a significant safety risk for personal agents that adapt based on user interactions.
By Xinyan Wang, Xiaogeng Liu, Jun Shao, Chaowei Xiao
The paper introduces Formulaic, a workflow that separates the tasks of an agentic language-model agent, a mathematical solver, and verification to co-design AI hardware and software. The agent iteratively refines problem formulations from the computation graph down to the chip, using templates that guarantee correctness. Experiments on NVIDIA GB10 and a simulated Ventus GPU show significant speedups for Qwen models and faster verification through proof reuse.
By Yi Ding, Li Shang
The paper introduces ExLabel, an LLM‑based, self‑evolving assistant that assesses task clarity and test coverage in SWE‑bench‑style software engineering datasets. ExLabel builds an experience bank from cross‑instance knowledge, enabling context‑aware judgments that outperform static rubrics. A new benchmark, SWE‑TQ, spanning eight programming languages, demonstrates ExLabel’s superior F1 scores and its ability to filter data that improves a coding model’s success rate on SWE‑bench Verified.
By Zhonghao Jiang, Xuyang Zheng, Yanfu Yan, Zhongxin Liu, Lingfeng Bao, Shuiguang Deng
The paper introduces AgentNorm, an agent‑wise test‑time normalization technique for camera‑based cooperative perception. Unlike pooled batch‑normalization, AgentNorm estimates normalization statistics separately for each collaborator and anchors the ego vehicle to its own statistics, preventing degradation when a collaborator’s cameras are corrupted by fog, motion blur, or exposure errors. Experiments on OPV2V and DAIR‑V2X‑C datasets show significant AP improvements under severe fog and motion blur, with minimal added latency.
By Yu Wang, Shaohua Wang, Yicheng Li, Kecheng Sun
The paper introduces Task-Graph Contextual Integrity (TG‑CI), a formalism that extends traditional Contextual Integrity to multi‑agent systems by modeling workflows as Multi‑Agent Task Graphs (MATGs). TG‑CI evaluates privacy not just at individual disclosures but across the entire downstream propagation, offering two enforcement policies—Full‑Scope and Auto‑Region—that balance privacy protection with task performance. A new benchmark, MAPLE‑Bench, demonstrates that TG‑CI can cut unauthorized disclosures by up to 94% while preserving over 94% of task accuracy compared to existing local‑history CI methods.
By Hengyu Wu, Blaise Delattre, Yang Cao
The paper evaluates how large language models, when integrated into a harness that provides information retrieval and software execution, perform as agents in acoustic engineering tasks. In one experiment, the agent accurately detects and locates faults in vehicle interior audio recordings, showing that a weak model limits performance regardless of harness design. In a second case study, the agent reliably designs metamaterial unit cells to meet specified insertion loss targets across a frequency range, illustrating the agent’s practical utility in acoustic design.
By Arne Hildenbrand, Philipp Diel, Zhuo Cao, Hanno Scharr, Stefan Pischinger, Steffen Marburg
InferD is a low‑interaction honeypot framework designed to emulate AI inference infrastructure without executing any real model inference or tool calls. It achieves this by answering all adversary requests from static response pools across four deception planes—transport, service protocol, vulnerability surface, and agentic session—while adding a provenance plane to link credentials to evidence of downstream model‑mediated processing. The framework addresses the unique deception challenges posed by AI services, which differ from conventional network services in observable behavior and authorization flows.
By Hicham Hammouchi, Gabriele Lenzini
The paper introduces MASCAL, a method that enables each agent in a multi-agent collaboration system to output a calibrated probability of correctness for its answer. By jointly optimizing task performance and confidence calibration through reinforcement learning, upstream agents provide downstream agents with a quantitative trust signal, which is further leveraged via confidence-aware weighting during training. Experiments across three collaboration architectures show that MASCAL improves both accuracy and calibration, and that upstream confidence influences downstream decision-making, enabling selective answering and targeted rechecking.
By Chenxiao Fan, Chongming Gao, Gangyi Zhang, Yaxin Gong, Leyang Shen, Jiakai Wang, Dong Wang, Yang Liu, Haoyan Liu, Xinting Hu, Fuli Feng, Xiangnan He
The paper introduces Branch-aware Graph Planning (BGP), a method for offline goal-conditioned reinforcement learning that tackles the problem of path mirages in stochastic, long-horizon environments. BGP models stochastic branching by treating high-variability points as graph nodes and learns goal-conditioned edges with segmented learning to avoid unreliable branches and reduce unnecessary stochasticity. Experiments on diverse OGBench tasks demonstrate that BGP substantially outperforms existing state-of-the-art offline goal-conditioned RL approaches.
By Junhyeon Bae, Jaebak Hwang, Gwanwoo Choi, Minung Kim, Seungyul Han
BURGER is a benchmark for evaluating vision‑language‑action models on robotic drawing tasks. It provides a scalable, prompt‑driven pipeline that generates continuous drawing trajectories in both simulation and the real world, with 28,200 simulated demonstrations across 282 tasks. The benchmark enables fine‑grained analysis of language grounding, spatial composition, and action generation, revealing distinct failure patterns among five representative VLA models and demonstrating that some models can generalize to unseen compositions and scale with task diversity.
By Xiangyu Xu, Bangshuai Peng, Ziyi Liu, Mengjie Zhao, Chenhao Lu, Zhengyang Fan, Huazhe Xu, Kaizhe Hu
The paper introduces Prompts to Proxies (P2P), a two‑stage inference‑time system that uses structured attribute‑based prompting and entropy‑guided sampling to build a diverse proxy pool, followed by L1‑regularized regression to create a compact weighted ensemble that matches aggregate survey responses. P2P achieves low mean squared error on American Trends Panel data and competitive performance on World Values Survey cross‑locale experiments, outperforming baseline prompting and demographic‑conditioning methods while using minimal training data. The authors argue that this preference reconstruction approach offers a lightweight, externally verifiable alternative for aligning large language model proxies with target human populations.
By Bingchen Wang, Zi-Yu Khoo, Jingtan Wang
PRISM is a framework for proactive agents that decides when to intervene by using a cost‑sensitive decision gate and a dual‑process reasoning architecture. The agent only engages a resource‑intensive slow reasoning mode near the decision boundary, guided by a calibrated probability of user acceptance and asymmetric costs for missed help versus false alarms. Training employs gate‑aligned distillation, allowing the response policy to be decoupled from the gate for tunable, auditable control, and on the ProactiveBench benchmark PRISM reduces false alarms by 22.78% while improving F1 by 20.14% over strong baselines.
By Yuxuan Fu, Xiaoyu Tan, Teqi Hao, Chen Zhan, Xihe Qiu
The paper introduces a scale‑conditioned evaluation protocol for agent memory that measures how usable evidence degrades as irrelevant sessions accumulate. It logs memory trajectories and reports four diagnostics—budget‑compliant reliability, tail memory‑call burden, failure‑regime decomposition, and the usable‑scale boundary. Experiments on LongMemEval and LoCoMo show that reliability loss varies across memory interfaces and agents, highlighting that scalable‑memory claims must be conditioned on agent, interface, scale range, and interaction budget.
By Jiaqi Shao, Yiyi Lu, Yunzhen Zhang, Bing Luo
Hedge-Bench is a benchmark comprising 92 real-world environments derived from hedge fund analysts' reasoning traces, designed to test AI agents’ ability to forecast future events from incomplete information. Agents are evaluated on joint success—correct forecasts with over 60% research coverage—across 11 frontier models, achieving rates between 2.54% and 31.52%. Training Muse‑Glimmer with GRPO on 165 environments improves coverage on held‑out tasks and demonstrates generalization to out‑of‑distribution evaluations.
By Mahesh Ramesh, Eric Cho, Alice Lu, Shawn Huang, Baiqing Lyu
RePolicy is a safeguard that learns to both select the appropriate safety policy from a library and assess agent behavior against it. It processes an agent trajectory and a set of candidate policies, chooses the best-matching policy, applies its requirements, and predicts safety. The authors created PolicyTraj-20K, a dataset of over 20,000 annotated trajectories, and trained RePolicy with supervised and reinforcement learning, achieving top performance on six benchmarks and 19 baselines.
By Houcheng Jiang, Boxuan Zhang, Qiyong Zhong, Junfeng Fang, Xiang Wang, Xiangnan He
The paper investigates how to choose the size of a language model for AI agents to balance reliability, energy, cost, and time. By treating model selection as a multi‑criteria decision problem, the authors derive task‑length thresholds—crossover points—where a smaller model becomes more efficient than a larger one, based on per‑step reliability and resource use. Experiments with 15 open‑weight models on synthetic tasks confirm that these predicted crossovers match observed data for most cases, showing that smaller models are greener only up to a predictable task length.
By Sajid Siraj, Maria Barbati
The paper revisits group recommender systems, noting that despite decades of research, few real-world implementations exist. It critiques traditional assumptions about inter-group communication and decision-making, arguing they may not match user expectations. The authors propose leveraging generative AI, such as ChatGPT, to support conversational group decision support, potentially making interactions more natural and increasing adoption.
By Dietmar Jannach, Amra Deli\'c, Francesco Ricci, Markus Zanker