RoboArchitect automates the design of agent harnesses that orchestrate frozen robot policies, vision tools, and direct arm commands. By observing rollouts, diagnosing failures, and iteratively revising the harness, it improves success rates from 9.7% to 43.3% on unseen scenes and from 12.7% to 40.2% on unseen tasks without updating model weights. The evolved harness also develops persistent memory and intuitive control strategies, outperforming finetuned policies especially when the benchmark diverges from the policy’s training data.
By Shengguang Wu, Suning Huang, Cherie Ho, Ria Doshi, Hao Zhu, Serena Yeung-Levy
WildOS is a unified system that enables long‑range, open‑vocabulary object search in unstructured outdoor environments by combining safe geometric exploration with semantic visual reasoning. It constructs a sparse navigation graph for spatial memory and uses the ExploRFM vision module to score frontier nodes, predicting traversability, visual frontiers, and object similarity in real time. A particle‑filter localization method estimates distant goal positions, allowing the robot to plan toward far‑away targets, and field experiments show WildOS outperforms purely geometric or vision‑based baselines in efficiency and autonomy.
By Hardik Shah, Manthan Patel, Erica Tevere, Deegan Atha, Marcel Kaufmann, Shehryar Khattak, Marco Hutter, Jonas Frey, Patrick Spieler
VEvo-Harness is a causally-guarded, continual self‑evolving system designed to improve agentic image generation. It introduces a visual traceback mechanism, visual continual governance, a lifecycle‑managed harness state, and a heterogeneous sandbox policy to ensure only causally valid skill updates are incorporated. The system outperforms existing agentic methods and rivals proprietary generators while preserving continual evolution and generalization.
By Yuan Wang, Yuhang Yu, Mengting Chen, Tianrun Xu, Zihao Pan, Xuetao Feng, Wei Li
Apriel Voice 27B Preview is an open speech‑language model that unifies speech recognition, contextual reasoning, conversation maintenance, and tool use within a single reasoning backbone. Built on the post‑trained Gemma 4 26B‑A4B checkpoint, it processes speech directly and generates responses and tool calls without a separate dialogue model, using a two‑stage training procedure that adapts the projector and backbone through supervised fine‑tuning, text replay, and cross‑entropy/KL supervision. The model achieves strong performance on benchmarks such as Big Bench Audio (98.90), Speech‑BFCL (81.13), and τ‑Voice (65.51% task success), with a mean latency of 1.91 s, and is deployed on a custom serving platform for interactive voice interactions.
By Aman Tiwari, Akshay Kalkunte, Akintunde Oladipo, Jash Shah, Tara Bogavelli, Hoang Nguyen, Nishanth Madhusudhan, Vishal Juneja, Sridhar Nemala, Srinivas Sunkara, Sathwik Tejaswi Madhusudhan, Shruthan Radhakrishna, Masoud Hashemi, Kelechi Ogueji, Khalil Slimi, Saloni Mittal, Pulkit Pattnaik, Shashank Maiya, Dhruv Jhamb, Esakkivel Esakkiraja, Katrina Stankiewicz, Gabrielle Gauthier Melan\c{c}on, Fanny Riols, Amit Saha, Lakshmi Sirisha Chodisetty, Nitin Aggarwal, Anil Madamala, Sagar Davasam, Segan Subramanian, Vikas Yadav, Sai Rajeswar Mudumba
The paper introduces Noura, a system that uses large language models to generate personalized diet plans that satisfy public clinical guidelines and individual health profiles. By treating rejected plans as supervision signals, Noura trains compact models to produce higher-quality, rule-compliant plans—improving complete-plan pass rates from 24% to 72.5% and cutting rule violations by up to 97.5% on a large held‑out cohort.
By Junwei Wu, Runze Yan, Yuzhang Xie, Hanqi Luo, Ziwei Dong, Jiaying Lu, Xiao Hu, Carl Yang
rl_rob is an open‑source Python library designed for teaching reinforcement learning, covering a full curriculum from tabular temporal‑difference learning to deep value‑based and policy‑gradient methods. Every algorithm is implemented from first principles, making the update rules transparent and concise. The library provides a unified interface for all agents, enabling direct comparison across stages and allowing students to implement, modify, and explain update rules rather than merely producing working code.
By Abdulrahman Altahhan
ObjectGuidedMap introduces a multitask architecture that jointly predicts bird's eye view (BEV) maps and 3D objects, enabling explicit information exchange between the two tasks. By adding an object detection branch and cross‑task feature fusion through deformable attention and cross‑attention, the model allows map instances to leverage object context while remaining robust to occlusions. Experiments on nuScenes and Argoverse 2 show consistent improvements in BEV map prediction accuracy, especially a 21% mAP gain in heavily occluded scenes.
By Christian Bohn, Harsh Yadav, Klaus Friedrichs, Hasan Tercan, Tobias Meisen
The paper introduces the phenomenon of cyclical entropy eruption in agent reinforcement learning, where entropy rises sharply and then subsides in recurring cycles, unlike the steady low entropy seen in single‑turn reasoning RL. It breaks this dynamic into three phases, offering theoretical and empirical insights into the mechanisms driving the oscillation. The authors propose SEAL, a lightweight auxiliary loss that separates correct and incorrect trajectories in representation space, which stabilizes training and improves downstream agent performance across various benchmarks and algorithms.
By Wendi Li, Shawn Im, Sharon Li
The paper introduces RAGFlow, a retrieval‑augmented generation framework that transforms an undisciplined query loop into a structured pipeline of five stages. It emphasizes knowledge compilation—enriching documents with navigation hierarchies, claim‑evidence pairs, entity graphs, timelines, mind maps, ontologies, wikis, and memory—rather than improving retrieval accuracy. The compiled context reduces query tokens by 91% when evidence suffices, cuts latency from 140 s to 30 s, and provides governance and integration benefits not captured by standard accuracy benchmarks.
By Zhichang Yu, Baining Cao, Sbaaoui Idriss, Qi Wang, Gangqiang Xu, Ling Qin, Xiaoling Jin, An Liu, Shaohao Cai, Xiaoyang Bao, Shengle Chen, Changchang Ding, Pengfei Zhu, Jingxiang Xu, Yifeng Mao, Lu Liu, Hai Jin, Yingfeng Zhang
The paper introduces FABench, a scalable backtesting environment designed to evaluate forecasting agents. FABench automatically discovers events, generates questions, and resolves outcomes for any past forecast date, allowing continuous renewal beyond model training cutoffs. In initial tests, retrieval improves Brier scores across eight models, with top agents matching or surpassing market consensus on 54 Polymarket-linked questions, yet analysis shows persistent evidence‑weighting failures that highlight room for improvement.
By Yuxuan Wang, Yishu Wang, Jiaqi Deng
The study examines how large language models (LLMs) predict the emotional impact of outcomes in gambling and stock-trading scenarios and whether these predictions influence their decision-making. When outcomes are evaluated alone, LLMs show diminishing sensitivity to magnitude—similar to human value functions—but they amplify gains relative to losses, a phenomenon termed silicon gain amplification. When outcomes are compared directly, the ratings become linear, and in choice tasks the gain-loss asymmetry largely disappears, with the models’ affective ratings not providing a better explanation of choices than monetary utility alone.
By Joseph Marks (NatWest Group, London, UK), Marius C. Vollberg (University of Zurich, Zurich, Switzerland, Swiss Center for Affective Sciences, Geneva, Switzerland)
CIRCE is a method that estimates the difficulty of a task for a large language model (LLM) by extracting the problem’s constraint structure into a Constraint Intermediate Representation, which is then compiled into a constraint graph. Unlike prior approaches that rely on the problem text or model outputs, CIRCE uses a lightweight LLM to build this representation and feeds it to a graph neural network to predict whether the LLM will solve the problem correctly. Experiments on seven reasoning benchmarks and four LLMs show that even a randomly initialized GNN improves prediction accuracy, lowering negative log‑likelihood from 0.465 to 0.423 and raising AUROC from 0.691 to 0.727 compared to the best baseline.
By Chenhao Hong, Jiakai Xu, Kaushiki Singh, Yusen Zhang, Eugene Wu
Operon‑S1 is a fast decision controller that separates a language‑model agent’s judgment from its execution, posing 13 closed questions at key decision points and using a lightweight decision model to answer them. In seven life‑science benchmarks, the local decision model handled judgments in 0.6 s on median with no measurable impact on agent performance, while the main language model’s involvement accounted for 8–35 % of task cost. The system records every decision and its outcome, allowing retraining to improve accuracy, and is already applied to wet‑lab protocol troubleshooting and germline panel design.
By Genoria AI Team, Yuejie Hou, Yueyuxiao Yang, Weiting Tang, Meng Yang
arXiv:2610.41564v1 Announce Type: new
Abstract: Language agents increasingly tackle long-horizon tasks in interactive environments, where actions often depend on latent preconditions. The core diffic...
By Mingyang Liu, Qingcan Kang, Shixiong Kai, Tao Zhong, Mingxuan Yuan, Xin Li, Linqi Song
The paper audits 25 cognitive-architecture modules in an LLM‑based agent system, using a four‑question static audit protocol with line‑number evidence and call‑chain tracing. It finds that in 18 of 25 modules the named computation never produces any output, identifies three new dead‑code patterns, and confirms or falsifies 61 pre‑registered predictions across two systems. Repairs based on the audit raise planner task success from 0.028 to 1.000 and reduce state‑controller error by 52%.
By Peter Chen, Ni Zhang, Ying Xu, Siyuan Zhang
arXiv:2610.41647v1 Announce Type: new
Abstract: Multi-target on-orbit servicing missions require planning tools that can adapt to evolving operational conditions while accounting for the outcome vari...
By Adam Abdin (LGI)
AutoHebb introduces autonomous Hebbian networks that combine external feedback with internally generated neuromodulatory signals to gate a biologically grounded 3‑factor Hebbian learning rule. These networks, optimized via meta‑learning for fixed weights, demonstrate strong inductive biases and autonomous learning control, enabling zero‑shot inference and continual learning without external optimization. They outperform contemporary models on these tasks and can be reverse‑engineered into explicit algorithms, highlighting the importance of meta‑learning, Hebbian plasticity, and autonomous learning control.
By Daniel Levine, David van Dijk, Kenneth Kay
TSFLab is a self‑improving, agent‑agnostic workbench designed to streamline time‑series forecasting research. It equips coding agents with shared resources and tools, enabling low‑cost experimentation, reproducibility, and continuous maintenance of research artifacts. Experiments show that TSFLab boosts task completion rates by 20‑163% and reduces token usage and development time significantly compared to baseline approaches.
By ChengAo Shen, Yanjun Zhao, Haoxin Liu, Bonnie Chen, Miles Ma, Dongjin Song, Hanghang Tong, Qingsong Wen, Jingrui He, Jingchao Ni
arXiv:2610.40753v1 Announce Type: cross
Abstract: Coding agents increasingly rely on automated approvers to decide whether a proposed shell command may run. Existing approvers, from rule-based checke...
By Weiheng Wu, Junkai Chen, Zhensu Sun, Jieke Shi, Yunbo Lyu, David Lo
arXiv:2610.41132v1 Announce Type: cross
Abstract: The EU Platform Work Directive, to be transposed by December 2026, requires ride-hailing platforms to state in plain language the main parameters of...
By Xinyu Wang, Andres Fielbaum