The paper introduces a Bellman optimality equation for optimizing plasticity in Markov decision processes, extending recent work that reframes the stability‑plasticity dilemma as an empowerment‑plasticity tradeoff. It builds on Abel et al. (2025), which defined plasticity as the generalized directed information from observations to actions and empowerment as the reverse. This work is the first to address plasticity optimization under that new definition, providing a theoretical foundation similar to existing empowerment‑based approaches.
By Jeremy Lucas, Doina Precup
CT‑SAFR is a multi‑layered verification framework designed to enhance the safety and faithfulness of Chain‑of‑Thought reasoning in autonomous robots. The framework achieves a 94.2% hallucination detection rate with sub‑500 ms latency, and a warehouse robot case study shows an 87% reduction in unsafe reasoning outputs. The study also offers recommendations for responsible deployment of reasoning‑capable autonomous robots.
By Cagri Temel
arXiv:2511. 09902v3 Announce Type: replace Abstract: We show that using autonomous-flow-based generation, one can universally approximate orientation-preserving diffeomorphisms defined on the cube by Neural ODEs with rate $\mathcal{O}(P^{-1/d})$ with $P$ parameters.
By Hossein Rouhvarzi, Anastasis Kratsios
The paper reports a pioneering study on emergent risks in generative multi‑agent systems, focusing on scenarios such as competition over shared resources, sequential handoff collaboration, and collective decision aggregation. It finds that group behaviors like collusion‑like coordination and conformity arise frequently across varied interaction conditions, mirroring known human societal pathologies even without explicit instructions. These risks cannot be mitigated by existing agent‑level safeguards alone, highlighting a social intelligence risk inherent to intelligent multi‑agent collectives.
By Yue Huang, Yu Jiang, Wenjie Wang, Haomin Zhuang, Xiaonan Luo, Yuchen Ma, Zhangchen Xu, Zichen Chen, Nuno Moniz, Zinan Lin, Pin-Yu Chen, Nitesh V Chawla, Nouha Dziri, Huan Sun, Xiangliang Zhang
The paper introduces RD-Forget, a training‑free framework that separates what a persistent language agent stores from what it uses at answer time. It keeps a source archive of all observations while a query‑conditioned memory view filters evidence relevant to the current question, using a frozen language‑model curator to group facts into semantic slots and preserve multi‑hop relations. The approach employs rate‑distortion principles to stay within a memory budget and demonstrates improvements across conversational memory, knowledge updating, fact consolidation, long‑context reasoning, and personalization tasks.
By Yuhang Li, Yuchen Li
The paper introduces the Discovery Certification Protocol (DCP), a framework that transforms claims from AI research agents into executable tests for recovery and feedback. DCP includes multiple gates that validate improvements, provide controlled information, and measure the impact of truthful feedback, while its core requires strict controls and finite‑sample bounds. Controlled audits in SQLite optimization and virtual catalyst control demonstrated zero recoveries across 96 episodes, with rigorous verification by a deterministic, LLM‑free verifier.
By Jingjie Ning, Shanshan Zhong, Xiaochuan Li, Ji Zeng
The paper introduces FedMVLA, a modality‑decoupled federated learning framework designed for privacy‑preserving embodied intelligence in 6G networks. It addresses the unique challenges of vision‑language‑action models by employing modality‑aware federated aggregation, privacy allocation, and communication compression, along with a precision‑critical action transport slice. A case study on federated robotic manipulation demonstrates significant gains in task success, scalability, and uplink payload reduction compared to standard FedAvg.
By Zhuodong Liu, Xiangyu Li, Chunhong Yuan, Hongyang Du, Bodong Shang, Qingqing Wu, Tony Q. S. Quek, Mohsen Guizani
The paper introduces a phase‑decoupled, model‑calibrated power controller for disaggregated large‑language‑model (LLM) serving, addressing the mismatch between GPU power settings and the distinct hardware regimes of prefill and decode stages. By calibrating separate power caps for each lane based on measured throughput‑latency cliffs, the authors achieve a 20.4% increase in tokens per joule with only a 3.5% rise in mean end‑to‑end latency on an 8‑node B200 cluster, outperforming NVIDIA’s Max‑Q profile. The approach also demonstrates consistent meeting of ITL‑p99 service‑level objectives across multiple MoE models and yields a 32.3% electricity savings over a sustained three‑day run.
By Jae Gon Kim, Donghoon Yoo, Hanyul Ryu, Sungho Ha, Juyeon Lee, Soojung Ryu
The paper investigates reinforcement learning with multi‑step transition look‑ahead, where an agent can foresee the states resulting from any sequence of λ actions before choosing its next move. It proves that exact planning remains NP‑hard for every fixed rational discount factor γ in (0,1), and introduces a randomized polynomial‑time approximation scheme that works for any fixed look‑ahead depth. Extending this to unknown transitions and stochastic rewards, the authors develop an algorithm with cumulative regret matching classical tabular discounted RL up to logarithmic factors, showing that efficient near‑optimal planning and learning are achievable despite the NP‑hardness of exact planning.
By Corentin Pla, Hugo Richard, Marc Abeille, Vianney Perchet
The paper introduces an end‑to‑end framework for testing image‑triggered command injection on computer‑use agents (CUAs). It demonstrates that a local visual patch can cause verifiable environmental effects through the entire pipeline—from screenshot input, through vision‑language‑model generation, action parsing, to execution—by training and deploying patches on GitHub Pages and a CSDN clone. Across five open‑source or publicly available GUI‑agent or VLM backends, the study reports 84.5% T‑ASR, 47.0% TAPR, and 20.3% E2E‑ASR success rates, with trajectory analysis revealing that some attacks first execute malicious terminal commands before continuing the original task.
By Zhihao Liu, Hongyu Sun, Zhiyuan Fu, Xiaonan Duan, Jice Wang, Shangru Zhao, Weizhi Meng, Wuxin Yang, Yangfan Zhou, Yuqing Zhang
arXiv:2609. 09404v1 Announce Type: cross Abstract: Agentic AI frameworks let a language model plan, keep memory, and call tools that reach real files, mail, and services.
By Viet K. Nguyen, Mohammad I. Husain
KernelGenBench is a unified benchmark that evaluates large language models and agentic systems for generating efficient Triton kernels across diverse operator sources and hardware platforms. It covers 210 operators from PyTorch ATen, vLLM, and cuBLAS, and tests a 110‑operator subset on six different chips, consuming over 15 billion tokens in evaluation. The study finds that no single method dominates across all sources and platforms, with significant variations in correctness and performance depending on the operator source and hardware, and that agentic approaches require millions of tokens per successful operator.
By Peiyu Zang, Jian Tao, Jialing Zhang, Yichen Yuan, Wentao Zhang, Guang Liu, Yonghua Lin
The paper presents a taxonomy of architecture options for foundation-model-based agents, covering functional capabilities, non‑functional qualities, and operational aspects of design‑time and run‑time phases. It also introduces a decision model to guide critical design and runtime choices, aiming to streamline and improve the development of such agents. By unifying these classifications, the authors seek to reduce fragmentation in the field and provide a structured framework for architects and developers.
By Jingwen Zhou, Qinghua Lu, Jieshan Chen, Liming Zhu, Xiwei Xu, Zhenchang Xing, Stefan Harrer
The paper presents a dual‑channel graph neural architecture for selecting the best exact solver for the Maximum Clique Problem (MCP). It combines a Graph Attention Network that captures local neighborhood patterns with a Multilayer Perceptron that models global statistical descriptors, trained on a benchmark of four state‑of‑the‑art solvers evaluated across diverse graph instances. The resulting model achieves 90.43 % test accuracy, outperforming classical baselines and the single‑best solver.
By Xiang Li, Shanshan Wang, Chenglong Xiao
The article introduces AgentActionBench, a benchmark designed to evaluate agent-based experiment reproduction across machine learning and AI4Science papers. It employs an MCP-based Action Recorder to capture agents’ behavior during reproduction and assesses the resulting traces against paper-specific rubrics. The benchmark includes 150 papers, with a human-annotated subset and model-assisted augmentation expanding it to over 10,000 rubric items, revealing that current systems face execution bottlenecks but that model-generated rubrics correlate strongly with human judgments.
By Hanhua Hong, Yizhi Li, Luu Gia Huy, Jian Yang, Ming Zhou, Chenghua Lin
The study audits 576 LLM-based social simulations from 350 papers using the PIMMUR framework, which evaluates agent profile, interaction, memory, minimal control, unawareness, and realism. Results show that PIMMUR principles are met more often than minimal control, unawareness, and realism, with frontier LLMs correctly identifying the underlying experiment in 65.2% of cases and half of prompts pre‑determining outcomes. Reproducing five experiments revealed that many reported collective phenomena disappear or reverse when PIMMUR principles are enforced, suggesting that apparent emergent behaviors may be methodological artifacts rather than genuine social dynamics.
By Jiaxu Zhou, Jen-tse Huang, Xuhui Zhou, Man Ho Lam, Xintao Wang, Hao Zhu, Wenxuan Wang, Maarten Sap
The paper introduces SHAPER, a self‑evolving framework that enables embodied agents to adapt to new environments without retraining the underlying foundation model. SHAPER keeps model parameters frozen and instead evolves reusable skills and a context‑code harness through rollouts in the target environment, allowing the same model to act as both planner and optimizer. Experiments on VLABench and ESI‑Bench demonstrate that this skill‑and‑harness optimization outperforms pure execution, supervised fine‑tuning, and test‑time scaling baselines, showing that self‑evolving adaptation is viable when model training is costly or impractical.
By Peidong Wang, Zhiming Ma, Ying Chang, Xufang Luo, Yiqun Zhang, Zihan Wang, Xiaocui Yang, Shi Feng, Yuqing Yang, Dongsheng Li
CamPilot is a multi‑agent cinematic assistant that combines cinematographic planning with camera‑work control to generate more coherent and aesthetically pleasing movies from text prompts. It learns camera‑work planning from 14,000 professional films using a GRPO‑based learning paradigm, capturing motion patterns, composition principles, and cross‑shot relationships. The system is evaluated with a new benchmark, CamEval, and outperforms existing text‑to‑movie methods in cinematographic control and quality.
By Yang Wu, Stefano Petrangeli, Ishita Dasgupta, Yu Shen
V‑Retrver is an evidence‑driven retrieval framework that treats universal multimodal retrieval as an agentic reasoning process grounded in visual inspection. It allows multimodal large language models to selectively acquire visual evidence through external tools, alternating between hypothesis generation and targeted visual verification. The approach is trained with a curriculum that blends supervised activation, rejection‑based refinement, and reinforcement learning, achieving an average 23.0% improvement in retrieval accuracy across multiple benchmarks.
By Dongyang Chen, Chaoyang Wang, Dezhao Su, Xi Xiao, Zeyu Zhang, Jing Xiong, Qing Li, Yuzhang Shang, Shichao Kan
HeteroPROMPT is a real‑time, privacy‑preserving framework for heterogeneous collaborative perception in autonomous systems. It aligns features from diverse sensors and models into a unified ego‑centric space using modular prompts and lightweight tuning, while keeping encoders and fusion stacks frozen. The system employs a metadata‑free autoencoder for modality classification and routing, achieving higher average precision on OPV2V‑H and V2XSet datasets with far fewer trainable parameters.
By Armin Maleki, Hayder Radha