Spine is a new interface and runtime that aligns a language model’s context with the nesting and completion of subtasks in long‑horizon coding problems. By letting agents declare explicit subtask boundaries through work scopes, Spine records interactions in a persistent SpineTree and compiles only the necessary unfinished subtask history into the model’s prompt, replacing completed subtasks with their results while reusing prefix‑cache. Integrated into Codex and Pi, SpineCodex and the reusable Pi package achieve significant performance gains—improving SWE‑Milestone and ProgramBench scores and reducing costs—without the overhead of recursive subagent delegation.
By Jiahong Xiang, Kunqiu Chen, Lingming Zhang, Yuqun Zhang
AnalogScout is a new framework that uses a large language model (LLM) agent to perform PVT-aware analog device sizing more efficiently. By reusing elite designs, applying analytic gm/Id local sizing, and refining globally under full-corner verification, it cuts power consumption by 43.3% compared to the CMA-ES evolution strategy while using the same 16,000 simulations on three SKY130 OTAs. The method also outperforms Bayesian optimization and other LLM-driven flows on the tested OTAs.
By Yichen Long, Keren Zhu
The paper introduces the first strategyproof deep reinforcement learning mechanism for online resource allocation, combining a monotonic neural network, a generalized critical payment rule, and training via RL or imitation. This design ensures truthful reporting and individual rationality while capturing nonlinear resource interactions beyond linear models. Experiments on a four‑resource stream show the neural rule attains 97.8% of an offline welfare upper bound, outperforming linear learners and first‑come, first‑served baselines, and imitation training achieves comparable performance 89 times faster than RL.
By Fariba Dehghan, Sebastian Stein, Vahid Yazdanpanah, Stephanie Gauthier
The paper investigates how different sensing channels help diagnose failures in a simulated bimanual tool‑use task. It finds that a tool’s slip is best detected by the tool hand, while the identity of the late hand can be inferred from both arms but not reliably from a single hand’s sensor. A repair policy that escalates when the cause is unclear can fix as many failures as one that knows the exact cause within two attempts.
By Baptiste Bonin, Caro Strickland, Audrey Durand
The paper introduces Adaptive Residual Correction (ARC), a lightweight post‑training module that adds a reinforcement‑learning–driven residual correction to a frozen imitation policy. ARC uses sparse human corrections to decide when the residual should activate, where in the residual action space to explore, and what the final residual action should be, thereby treating local failures as a localized learning problem rather than a global one. In experiments on three precision‑manipulation tasks, ARC raises average task success from 33.3 % to 75.0 % and confines the residual corrections to 2–3 dimensional subspaces that capture over 96 % of the correction magnitude.
"whyItMatters":"The results demonstrate that sparse human guidance can effectively localize and constrain autonomous post‑training corrections, enabling significant performance gains without requiring full policy updates or dense human supervision."
By Anukriti Singh, Maxime Caniot, Pratap Tokekar, Firas Abi-Farraj
Gym-Anything-RT is a new benchmark that creates 100 real‑time visual minigames to test vision‑language AI agents’ ability to perceive and act while the environment evolves. The framework automatically generates these environments from a taxonomy of six real‑time interaction capabilities and allows controlled experiments by pausing or slowing the game. Evaluation shows a stark gap: GPT‑6 Astra passes only 9 % of the real‑time games, but 75 % when the environment is paused, revealing significant limitations in current agents’ real‑time performance.
By Pranjal Aggarwal, Sean Welleck
The paper introduces the notion of faithful latent world models, defined as an equivariance between an environment’s transition map and the learned latent dynamics. It proposes latent rollout consistency (LRC) as a metric for evaluating action‑conditioned latent predictions. Experiments show that current state‑of‑the‑art latent world models fail to faithfully model the environment’s dynamics, despite achieving task performance.
By Rodrigo Gonzalez Laiz, Steffen Schneider
The paper introduces HiMA-Ecom, a hierarchical multi‑agent e‑commerce benchmark with 22.8K instances, where 17.7K include memory entries. It explores training, evaluation, and management of accumulated context in long‑horizon agentic systems, and proposes HiMA‑R1, a policy optimization method that controls trajectory context budgets and uses reward‑driven memory to accelerate convergence. Experiments show that agents built on 3B and 7B open‑source models outperform DeepSeek‑V3 by an average of 6.0 points, while reward‑driven memory cuts convergence steps by about 55%.
By Junxing Hu, Ai Han, Haolan Zhan, Pu Wei, Zhiqian Zhang, Yuhang Guo, Jiawei Lu, Zhen Chen, Haoran Li, Zicheng Zhang, Ling Li
The paper introduces Metacognitive Behavioral Tuning (MBT) and its agentic variant MBT‑A, which edit long reasoning traces from a teacher model to insert seven metacognitive behaviors—such as planning, thorough verification, and avoiding shortcuts—at appropriate points. By applying these edits line‑by‑line and recording each as a git commit, MBT‑A avoids summarization and distributes insertions throughout the trace. Experiments on competition math show that a Qwen3.5‑4B‑Base student trained on MBT‑A traces achieves higher scores (85.4 vs. 83.1) than standard supervised fine‑tuning, and the inserted behaviors transfer to other student models, increasing the frequency of rare behaviors.
By Ik-hwan Kim, Hyeongrok Han, Mingi Jung, Sangwon Yu, Jinseok Hong, Sang Hun Kim, Yoonyoung Choi, Sungroh Yoon
The paper introduces Bot‑Mod, a moderation framework that detects malicious intent in multi‑agent systems by interrogating agents with diagnostic probes and updating a Bayesian posterior over candidate intents. Bot‑Mod starts with a content‑only prior and refines its inference using moderator‑assigned likelihoods until a stopping criterion is reached. Evaluations on the Molt‑Intent dataset, an external social engineering dataset, and three agent LLMs show that Bot‑Mod outperforms content‑only moderators, a fine‑tuned classifier, and judge‑based interrogators, achieving higher F1 scores both in‑distribution and out‑of‑distribution.
By Ali Al-Lawati, Nafis Tripto, Abolfazl Ansari, Fali Wang, Jason Lucas, Suhang Wang, Dongwon Lee
ScalableRAG introduces a persistent typed‑set workspace that lets an agent construct named document and value sets at query time and combine them using set algebra, typed extraction, aggregation, and diagnostics. The zero‑ingestion variant requires no preprocessing or vector index yet matches leading baselines on standard QA benchmarks and outperforms them on the Hotels‑50 aggregative task. A lightweight‑index version achieves over 80 judge score on 5 M documents with low latency and minimal ingestion time, surpassing other systems in both accuracy and efficiency.
By Hilaf Hasson, Aditya Chakravarty, Jayant Thomas, Krishna Gogineni
The paper investigates the macroscopic dynamics of large language model (LLM)-driven agents by applying a least‑action principle to estimate the generative directionality of LLMs embedded within agents. Through experimental measurement of transition probabilities between LLM‑generated states, the authors find evidence of a tendency toward detailed balance, suggesting that LLM generation may be governed by underlying potential functions rather than explicit learning rules. This approach offers a model‑agnostic, effective description of LLM generative dynamics that could serve as a foundational theory for complex AI systems.
By Zhuo-Yang Song, Qing-Hong Cao, Ming-xing Luo, Hua Xing Zhu
The paper presents a novel dimensionality‑reduction framework for reach‑avoid games involving unmanned aerial vehicles. By decomposing the three‑dimensional problem into horizontal and vertical sub‑games, each solved with Hamilton‑Jacobi reachability analysis, the authors reconstruct the full solution using a tracking control algorithm that guarantees attacker capture and subsequent tracking. Numerical simulations and Gazebo simulations confirm that the decomposition preserves optimality and achieves successful quadrotor capture in three‑dimensional space.
By Minh Bui, Simon Monckton, Mo Chen
SQuID is a benchmark dataset comprising 2,000 quantitative spatial reasoning questions over 587 satellite images, covering 24 question types across three difficulty tiers. Answers are derived from curated segmentation masks or human consensus, with numeric responses scored against tolerance ranges based on human annotations. The benchmark also provides 75,687 training question-answer pairs and reference masks, enabling modular evaluation and highlighting segmentation as a key error source for current vision‑language models.
By Peter A. Massih, Eric Cosatto, Martin Renqiang Min
AgentSentry is an inference‑time defense for tool‑augmented large language model agents that mitigates indirect prompt injection (IPI) attacks. It uses temporal causal diagnostics by performing counterfactual re‑executions at tool‑return boundaries to detect mediator‑sensitive proposals, purifies implicated content, and resumes execution from a sanitized state. Evaluations on AgentDojo show zero successful attacks (ASR = 0%) and an average Utility Under Attack of 76.10%, outperforming baselines while maintaining low false‑positive rates.
By Tian Zhang, Yiwei Xu, Juan Wang, Keyan Guo, Xiaoyang Xu, Bowen Xiao, Quanlong Guan, Jinlin Fan, Jiawei Liu, Zhiquan Liu, Hongxin Hu
SWE-Skills-Bench is a benchmark that evaluates the impact of individual agent skills on real-world software engineering tasks. It uses 98 public skills paired with specific GitHub repositories, generating 3,920 paired evaluations across four agent-model configurations. The study finds that most skills are ineffective, with only a few providing modest gains while some are harmful, raising the overall pass ratio from 76.6% to 76.9%.
By Tingxu Han, Yi Zhang, Wei Song, Chunrong Fang, Zhenyu Chen, Youcheng Sun, Lijie Hu
The paper presents an empirical investigation of scaling laws for reinforcement learning, focusing on an actor‑critic algorithm in the Starpilot and Craftax environments. It demonstrates that these scaling laws can accurately predict performance, optimal resource allocation, and hyperparameter settings for training jobs up to two orders of magnitude larger. The study also shows that Elo scores provide a more predictable performance metric than mean episode return, and uses the discovered laws to achieve a new state‑of‑the‑art result in Craftax while revealing a non‑monotonic scaling pattern linked to Ray interference.
By Zeyu Zheng, Clare Lyle, Khimya Khetarpal, Will Dabney
The paper proposes a framework for the collective collaboration of recursive autonomous agents (RAAs) in smart factories. It introduces a common task and capability interface that enables recursive decomposition of production objectives from the top down and aggregation of conditional capabilities from the bottom up. The framework includes organizational models for leaf and composite units, conditions for finite task expansion, and operational mechanisms linking task decomposition, resource commitments, result aggregation, and coordination across changing production conditions.
By Ziqing Zhou, Linqiang Hu, Yuning Chen, Hongda Zhang, Zhao Huang, Ding Zhou, Haining Xu
The paper introduces OPAL, a reactive optimizer that employs a pretrained language‑model agent to monitor training dynamics and modify optimizer code on the fly. OPAL can alter update rules, schedules, learning‑rate schedules, optimizer structure, numerical precision, and momentum for individual parameter groups without prior meta‑training or cross‑run memory. Experiments on language modeling and image generation show that OPAL matches or surpasses fixed recipes, adapts to data corruption, and reduces over‑fitting, while the authors also discuss its computational cost and implementation details.
By Xingjian Bai, Jia-Hao Liang, Jayden Teoh, Vincent Sitzmann
The paper introduces ALMOND, a cognitive framework that bridges the gap between human intent and low‑level control logic for 6G low Earth orbit satellite networks. It employs a bidirectional semantic engine to convert natural‑language requests into multi‑objective reward functions and extracts operational constraints from failed reinforcement‑learning trajectories, while a large‑language‑model‑guided graph neuroevolution designs efficient policy architectures. In a 350‑satellite digital twin, ALMOND improves polar handover success from 55% to over 95%, speeds convergence threefold, and cuts reward‑engineering time from 48 to 4.5 hours, all within a compute‑frugal on‑board policy of under 50,000 parameters.
By Zhengqian Xu, Yasheng Zhang, Chenhua Sun, Shan Lu, Chengcheng Li