The paper evaluates GPT‑6‑Astra’s performance across 12 robotic benchmarks, showing it achieves unprecedented results when using in‑context learning from training‑set demonstrations. It highlights the benefits and costs of a dual‑system architecture (high‑level reasoning and low‑level control) and identifies practical gaps in real‑world deployment, such as expensive demonstration collection and physical constraints. The authors propose a new benchmark prototype that derives tasks from real‑world needs, omits in‑domain demos, and measures agents against a continuously running world clock.
By SenseNova, Joint Research Team, :, Zhongang Cai, Ziang Cao, Tailai Chen, Han Chong Chua, Linjun Dai, Hanming Deng, Hokin Deng, Haiwen Diao, Chenjing Ding, Weichen Fan, Xiao Fu, Zerui Gong, Chenyang Gu, Zexin He, Zhi Hou, Bo Li, Qinyue Li, Quanzhou Li, Xiaowan Li, Zhichao Liao, Dahua Lin, Zilong Lin, Bo Liu, Chunxiao Liu, Pengfei Liu, Tianqi Liu, Ziwei Liu, Chen Change Loy, Lewei Lu, Jiapeng Luo, Yuexiao Ma, Liang Pan, Xingang Pan, Oscar Qian, Ziyi Shang, Yuyao Shao, Dawei Su, Dacheng Tao, Qingyi Tao, Quan Wang, Ruisi Wang, Xiaogang Wang, Ziyi Wang, Beichen Wen, Donghao Wu, Jiaxu Wu, Silei Wu, Xiaoming Wu, Zhonghua Wu, Hongcan Xiao, Weiye Xu, Atsushi Yamashita, Lei Yang, Zhe Yang, Baiqiao Yin, Wanqi Yin, Hanfei Zhao, Tianbo Zhong, Long Zhuo, Jing Zuo
The paper introduces "Anytime Contracts" for AI services, addressing moral hazard where providers may downgrade models and users can terminate usage at any time. It extends contract theory to handle unknown exit rules by framing the problem as a Principal-Agent game and shows that optimal contracts align with anytime-valid hypothesis tests using likelihood‑ratio e‑values. The authors provide an efficient algorithm for computing such contracts and demonstrate its application on synthetic data and real large language model deployments.
By Boris Velasevic, Eden Saig, Eric Mazumdar
HardEval is an execution‑verified benchmark that tests language‑model‑assisted security hardening on a live Red Hat Enterprise Linux 8 system. It uses STIG controls that fail on an unhardened baseline, and models generate Bash scripts that are run on a virtual machine and graded by OpenSCAP/OVAL compliance checks. The benchmark reveals significant differences in model performance across security domains and highlights failures where code appears plausible but does not meet the exact compliance condition.
By Manohar Vellala, Soumya Banerjee, Sandip Roy, Ross Gore, Sachin Shetty
The paper analyzes primal‑dual algorithms for constrained Markov decision processes under a two‑timescale framework, focusing on how non‑stationarity enters via additive and parametric channels. In the additive case, oscillating constraint levels cause the dual variable to saturate at a floor determined by the moving optimum, with amplification proportional to the timescale ratio. In the parametric case, oscillating constraint sensitivity leads to a parametric resonance at twice the natural frequency, driven by mismatches between primal and dual estimates; this resonance is mitigated by aligning the updates or using a PID term.
By Maxime Baelde
The paper investigates whether the policy gradient update used in many reinforcement learning algorithms—specifically the episode-summed, on‑policy update that omits the outer discount factor—can be interpreted as the gradient of any objective function. The authors prove that this update does not necessarily correspond to the gradient of any function and provide a counterexample where the update drives the policy toward a globally pessimal deterministic limit while the parameters diverge. They also highlight that this theoretical ambiguity has led to widespread misstatements in the literature, even in highly cited works from top conferences.
By Chris Nota, Philip S. Thomas
The paper evaluates the energy consumption of sensor nodes in IoT applications, examining the entire pipeline from data acquisition to optional on-device CNN classification and LPWAN transmission using an ESP32‑S3 microcontroller. It compares on‑device inference followed by result transmission with raw data transmission across three environmental monitoring scenarios, varying model architecture and LPWAN technologies (LTE‑M, NB‑IoT, LoRaWAN). Results show that on‑device inference can reduce energy consumption by 1.7× to 6.9× compared to raw data transmission, with a tolerable loss in accuracy, and identify developer‑controllable variables that influence the energy trade‑off.
By Benjamin Karic, Nina Herrmann, Jan Stenkamp, Paula Scharf, Fabian Gieseke, Angela Schwering
The paper presents a pipeline for training, compiling, and deploying Bayesian neural networks (BNNs) that use a single probabilistic forward pass (PFP) instead of multiple stochastic samples. By assuming Gaussian-distributed weights and activations, PFP enables fully analytic uncertainty propagation, allowing a single deterministic forward pass to replace sampling. Using the TVM compiler and a library of Gaussian-propagating operators, the authors achieve up to 4200× speedups on small mini-batches while maintaining accuracy, uncertainty estimation, and out-of-domain detection comparable to traditional stochastic variational inference (SVI).
By Bernhard Klein, Falk Selker, Hendrik Borras, Sophie Steger, Franz Pernkopf, Holger Fr\"oning
The paper introduces Cast‑R1, an 8‑billion‑parameter agentic forecasting model that learns to orchestrate analytical and prediction tools through trajectory distillation and multi‑turn reinforcement learning with structured memory. The training pipeline first generates multi‑turn interaction trajectories using a teacher LLM, then fine‑tunes the agent with supervised learning and group relative policy optimization, guided by rewards for accuracy and temporal structure. Experiments across several datasets show that Cast‑R1 effectively equips smaller language models with tool‑augmented, multi‑turn forecasting capabilities within structured workflows.
By Xiaoyu Tao, Mingyue Cheng, Yaguo Liu, Xingpeng Gao, Qingsong Wen, Shijin Wang, Enhong chen
LoongSage is a lightweight, fully asynchronous reinforcement learning framework designed for large language models that perform agentic tasks. It integrates with Megatron for distributed training and SGLang for high‑throughput inference, centralizing orchestration in a single controller and using Ray actors for GPU components. The framework introduces AgentFlow for unified execution of white‑box and black‑box agents, sliding‑window strategies to limit trajectory staleness, TransferMesh for efficient weight synchronization, and a memory‑efficient Teacher Manager that supports multi‑teacher on‑policy distillation. Experiments across four training runs show steady reward improvement, controlled train‑inference gaps, and sustained hardware utilization by eliminating synchronous idle periods.
By LoongSage Team, Bowei Song, Caiwei Zhao, Chuwen Liu, Dechuang Gu, Dong Fang, Feiyu Xie, Huaiyu Li, Hui Zhang, Jian Tian, Jian Wang, Jie Zhao, Liguang Li, Maoyun Wang, Meiyan Yang, Minghan Zhu, Pai Liu, Pandeng Yao, Renbao Tian, Ru Yang, Shuai Jia, Shuai Tan, Shuailong Li, Shuliang Li, Teng Jin, Tianze Yang, Wenbo Cui, Wenjing Bai, Wenkang Wu, Xiaowei Jiang, Xiaoyi Dong, Yanpeng Wang, Yanwen Wang, Yaohui Ma, Yawen Wang, Yihang Zhang, Yingjie Yan, Zhanbiao Lian, Zhenyu Lu, Zhida Hu, Zhihui Hao, Zhihui Hu, Zhiwen Xiao, Zhiyi Gao, Zhongwei Qi, Zhuangzhuang Wu, Ziqian Liu, Ziyue Zheng
RubricEM is a unified reinforcement learning framework designed for deep research agents that plan, search, evaluate evidence, and synthesize long‑form reports. It introduces rubric‑guided policy decomposition, combining supervised fine‑tuning to scaffold agents with rubrics, stage‑structured GRPO for fine‑grained credit assignment, and a reflection meta‑policy for experience reuse. The resulting RubricEM‑8B model outperforms larger open models on four long‑form research benchmarks and approaches the performance of proprietary deep‑research systems.
By Gaotang Li, Bhavana Dalvi Mishra, Zifeng Wang, Jun Yan, Yanfei Chen, Chun-Liang Li, Long T. Le, Rujun Han, George Lee, Hanghang Tong, Chen-Yu Lee, Tomas Pfister
The paper investigates prompt injection attacks on medical multi‑agent systems that use large language models (LLMs) for clinical decision support. Across two systems, two model families, and three benchmarks, a single injected passage can cause incorrect decisions in up to 61% of cases, and existing defenses only partially mitigate the threat. The authors propose a training‑free defense that verifies each clinical judgment against the patient’s findings, which reduces attack success on MedQA to 4‑7% while preserving accuracy on non‑attacked cases.
By Fardin Ahsan Sakib, Ziwei Zhu, Ozlem Uzuner
The paper introduces ProPerMem, a benchmark for memory-driven proactive assistance that pairs egocentric video with long, distractor-rich personal memories and human-verified trigger windows. Baseline models perform poorly, achieving at most 15% accuracy, while an oracle-driven GPT-5.3-Chat model reaches 96% when given evidence inside the trigger window. Building on these insights, the authors develop ProPerAgent, a plug‑and‑play system that retrieves relevant memory from a graph store and uses lightweight reinforcement learning to improve proactive intervention accuracy to 42.4% without training.
By Shenglai Zeng, Xiao Yang, Tiffany Ma, Xinyuan Zhang, Nanshu Wang, Deren Lei, Kai Sun, Xilun Chen, Tammy Stark, Xianhui Zhu, Yuetai Li, Rulin Shao, Tianxin Wei, Rakesh Wanga, Anuj Kumar, Jiliang Tang, Wen-tau Yih, Xin Luna Dong
The paper introduces Multimodal Filesystem Memory (MFM), a memory architecture for long‑horizon embodied agents that stores episodic, spatial, semantic, procedural, and working memories. MFM uses searchable textual records to link these representations to multimodal episodes, enabling agents to retrieve past experiences via contextual cues and revisit original observations for missing details. Experiments on long‑horizon Minecraft and computer‑use tasks show that MFM improves performance over three base models, achieving higher milestone completion and reward scores compared to baselines.
By Chloe Su, Jiahan Li, Zhenting Qi, Chaoran Cheng, Haoyu Chen, Sejoon Jun, Weirui Ye, Yilun Du
The paper presents a Bengali hallucination detection system that uses response‑conditioned query rewriting, hybrid retrieval, and task‑specific language‑model judging. It employs six retrieval routes and a router that selects between a 27B textual judge and a 14B agent capable of reasoning or executing Python. The authors evaluate the system on the Olikbochon challenge, reporting a private macro‑F1 of 0.956 and detailing runtime, execution traces, and corpus overlap findings.
By Shamim Hasan, Samiya Ahmed Chowdhury
The paper introduces a controlled evaluation set focused on the Go concepts of atari and sente, aiming to test whether language models can recognize these terms and link them to move evaluation. The authors fine‑tune a Go language model, LoGos, with terminology and commentary supervision and then assess its performance on local board changes and strategic move preferences. Results show that LoGos struggles to correctly identify atari relationships and to assess their strategic implications, yet its terminology predictions still influence move choices, sometimes contradicting the reference evaluation.
By Taiyi Chen, Iskar Deng, Amandalynne Paullada
AMEND is an inference‑time framework for text detoxification that separates detoxification into role‑conditioned proposal generation, isolated multidimensional criticism, and constraint‑first selection. It evaluates diverse candidates with role‑blind critics, enforces strict semantic and fluency feasibility, and ranks them by utility. The lightweight AMEND Lite variant limits the search to a single targeted repair round when residual toxic spans remain, achieving higher joint scores and reducing inference requests by 28.4% compared to unconstrained multi‑round loops.
By Tianyu Wang
RedDarwin is an agentic framework that evolves jailbreak skills by treating reusable attack mechanisms as modular skills rather than fixed prompts. It starts with 47 base skills and autonomously revises, composes, and selects them using feedback from target models, producing 65 evolved skills that improve 44 of the original attacks. The evolved skills raise the mean attack success rate on AdvBench from 18.8% to 47.3% and achieve high success on unseen frontier LLMs, including 100% ASR on GPT‑5.6‑Sol with only ten queries.
By Zhouyuan Ma, Yutao Wu, Hanxun Huang, Xiang Zheng, Yi Liu, Xingjun Ma
The paper introduces Cambrian-BEV, a parameter‑free geometric tokenization method that projects visual patch features from multiple video frames into a shared bird’s‑eye‑view (BEV) grid, aggregating observations of the same physical location across viewpoints and time. This approach yields a persistent spatial representation while dramatically compressing the visual token set—from 100 K DINOv3 patches across 128 frames to just 729 BEV tokens—reducing LLM prefill time by 2.9× and peak memory by 3.6×. Experiments on VSI‑Bench and ReVSI show that Cambrian‑BEV outperforms frame‑token baselines, achieving significant gains in room‑size and relative‑direction tasks.
"whyItMatters":"By providing a stable, compressed spatial context, Cambrian‑BEV enables more efficient and accurate multimodal large language models for embodied agents, improving performance on spatial reasoning benchmarks while reducing computational resources."
By Faizan Siddiqui, Nirav Savaliya, Sahith Reddy Chada, Enna Sachdeva
RAP (Reconstruct A Poster) is a new dataset and benchmark for generating conference posters from research papers, comprising over 25,000 verified paper‑poster pairs from 2022‑2026. Each pair is annotated with a three‑level content taxonomy and converted into styled HTML/CSS markup through an automated derendering pipeline, enabling objective render‑and‑compare evaluation. The corpus supports both inference and training, with a retrieval‑based PosterRAG pipeline and a supervised‑fine‑tuned model, and is publicly available via an online demo.
By Haonan Han, Yukang Lin, Minliang Lin, Haoxian Tan, Jun-Yan He, Rui Yang, Zhi-Qi Cheng, Jie Hu, Guanglu Wan, Xiu Li
Agentic Image Generation in the Wild pairs an image generator with a web‑search agent to retrieve references beyond the generator’s training data. The authors introduce the Savanna benchmark, comprising 500 prompts derived from recent headlines that are difficult to resolve due to limited visual evidence and confusing search results. They propose the Locate‑Corroborate‑Isolate paradigm and train Ranger‑8B, which outperforms existing frameworks on both search accuracy and generation quality.
By Yifan Li, Bowen Qin, Kun Zhou, Haochen Zhao, Zikang Liu, Zheng Li, Minsheng Hao, Yuhao Cheng, Haosheng Zou, Jeff Chen, Wayne Xin Zhao