AgentPerturb introduces a systematic framework to evaluate large language model agents under controlled interaction perturbations, addressing the lack of unified evaluation methods for noisy real‑world tasks. The study quantifies robustness across frontier models, identifies task‑specific vulnerabilities, and shows that agents often fail to detect altered inputs, propagating errors and incurring higher computational costs. The authors provide a benchmark‑construction pipeline and plan to release code and data to support future robustness‑oriented training.
By Yuanjian Xu, Jinyi Han, Jianing Hao, Yuxuan Sun, Fanyang Lu, Richeng Xuan, Guang Zhang, Zhichao Hu
RecSoup introduces a lightweight two‑stage method for Controllable Multi‑Objective Recommendation (CMOR) that merges multiple objective‑specific models without retraining. The approach first prunes parameters to reduce cross‑objective interference, then adaptively merges them using a learnable preference‑to‑coefficient mapping, enabling dynamic trade‑off control at inference time. Experiments show RecSoup achieves superior Pareto front quality and 165–228× faster inference than parameter‑generation baselines.
By Zidong Jia, Guofu Xie, Xiao Zhang
TRACER is a training‑free multimodal harness that builds minute‑level action records from continuous video and merges them with clinical data into hourly reports, enabling an LLM agent to query and inspect both modalities. In evaluation, TRACER achieved up to 84.8% accuracy on action detection and 61.0% accuracy on clinician‑informed QA, outperforming the strongest baseline. Across 280 patients and ~60,000 camera‑hours, TRACER detected significantly less purposeful activity in patients with a positive delirium assessment compared to those without (p < 0.0001).
By Zane Durante, Paul Tang, Keshigeyan Chandrasegaran, Amit Kaushal, Shrinidhi Kowshika Lakshmikanth, Feolu Kolawole, Yu Jun Zheng, Yifan Geng, Shobhit Agarwal, Kevin Schulman, Juan Carlos Niebles, Ehsan Adeli, Arnold Milstein, Li Fei-Fei
ARIA is a framework that learns to repair ongoing multi‑agent system executions by turning verified repairs into training data. It builds a large dataset of 3,953 repair pairs from over 24,000 trajectories across multiple frameworks and benchmarks, then trains a compact model to detect when and how to intervene during runtime. The resulting model shows strong repair performance, generalizing to unseen frameworks and host models.
By Bolin Shen, Yushun Dong
MAI-Gym is an open, all-in-one platform that unifies GUI agent evaluation, data collection, and online reinforcement learning training across mobile, computer, and web environments. It supports ten established benchmarks, scales to 20,000 parallel environment instances, and enables automated synthesis of diverse GUI applications and tasks with execution-validated verifiers. Using this platform, the authors trained a 9B-parameter agent, MAI-Gym-9B, achieving state‑of‑the‑art performance on MobileWorld and OSWorld-Verified, and demonstrated that training on a mocked train‑ticketing app improves real‑world task success from 28.8% to 48.1% on a physical phone.
By Panrong Tong, Quyu Kong, Pengxiang Li, Tianyu Xia, Hanzhang Zhou, Lei Wang, Chenglin Cai, Long Li, Long Chen, Jianan Zhang, Zhengwen Zeng, Xu Zhang, Gongjie Zhang, Yue Wang, Steven Hoi
Anvil is a unified framework for embodied spatial intelligence that integrates data generation, training, and benchmarking. It introduces three components: Anvil-Gen, a heterogeneous data flywheel that generates and verifies cross‑stage consistency; Anvil-Ben, which provides scene‑aligned supervision and assessments to separate stage competence from error propagation; and Anvil‑MOPD, a 9B student model that distills spatial knowledge from specialist teachers using semantic‑first fault routing and stage normalization. The system builds 5.8 M source‑grounded records, trains on 5.22 M, and evaluates on 580 K, outperforming state‑of‑the‑art models on seven public benchmarks.
By Weilin Ruan, Bohan Zhou, Jinghao Li, Yangkai Wei, Wenqian Li, Peiran Dong, Yingjie Xu, Leyi Wu, Haojian Huang, Binghui Xie, Han Yang, Kaiwen Zhou, Yinchuan Li, Ying-Cong Chen, James Cheng
The paper introduces an open midtraining recipe that primes base models for agentic post‑training, filling a gap between pretraining and full fine‑tuning. It trains on 256K context with hundreds of billions of tokens drawn from natural software‑engineering data and a small share of agentic traces. Applying this midtraining to Nemotron 3.5 Lightning and Super models yields measurable gains on SWE‑bench and DeepSWE benchmarks after identical full SFT, and improves robustness across agent harnesses.
By Soumye Singhal, Joseph Jennings, Ka Shun Shum, Alexander Bukharin, Khushi Bhardwaj, Zirui Liu, Venkat Srinivasan, Jonathan Li, Yash Jain, Kezhi Kong, Markus Kliegl, Olivier Delalleau, Sanjeev Satheesh, Mostofa Patwary
The paper proposes a framework for aligning AI agents with deployment requirements through a process called calibration, which avoids the need for source imitation or universal superiority. It introduces three families of requirements—basic capability, technical precision, and user context—alongside a versioned mapping to manage changes. The authors outline four instruments (benchmarks, examples, skills, and role-conditioned cost/utility) that provide tunable reward channels within a fixed acceptance contract, and describe methods for adaptation, qualification, and evaluation, noting that confirmatory validation remains outstanding.
By Yaxiao Liu (PwC China AI Center), Pengbo Liu (PwC China AI Center), Yiwen Liu (PwC China AI Center), Yihua Guan (PwC China AI Center), Jiaxing Song (Tsinghua University), Xingjian Li (PwC China AI Center)
TimeCap is an open, deterministic pipeline that generates text captions for wearable health time‑series data without using a language model in the loop. It consists of three reusable extractor tiers and a cross‑channel synthesis layer, allowing new schemas to be added by defining channel mappings. Applying TimeCap to the MyHeart Counts dataset produced MHCap, the largest open caption layer for multimodal wearable data, containing 7,005,763 captions over 908,944 person‑days across 19 Apple HealthKit channels.
By Max Rosenblattl, Milana Gurbanova, Karl Aurel Deck, Narayan Schuetz, Thomas Kaar, Manuel Milling, Nicolas Zumarraga, Ning Wang, Kelly W. Zhang, Anders Johnson, Robert Jakob, Bj\"orn W. Schuller, Elgar Fleisch, Daniel Seung Kim, Allan Lawrie, Patrick Langer, Maxwell A. Xu, Euan Ashley, Paul Schmiedmayer
The paper discusses how offline-trained neural networks, when used in closed-loop control systems, can change the state distribution they were trained on, leading to performance degradation. It introduces a taxonomy of failure modes—Closed-Loop Identification Error, deployment-induced Action Error, and continuous Tracking Error—and uses discrete-time sensitivity analysis to explain how these errors arise. The authors propose time-scale separated online adaptation policies to mitigate distribution shift, validated on a nonlinear Duffing oscillator, showing that bounded continuous learning can prevent trajectory divergence while static deployments cannot.
By Vittorio Lippi
PEDBench is a new benchmark for phonetic error detection (PED) that includes natural recordings of children’s speech errors, scripted error productions by actors, and synthetic error productions generated via phone‑conditioned text‑to‑speech. The benchmark provides high‑quality phonetic annotations verified by trained annotators. Evaluation of 13 state‑of‑the‑art speech models shows that they struggle to reliably detect phonetic errors, even in controlled conditions, and often misclassify correct phonemes as errors. Fine‑tuned speech‑language models perform better on word‑level PED, but a sim‑to‑real domain gap persists.
By Zhonghao Shi, Zackary Rackauckas, Jasmine Yip, Savannah Halpern, Simon Campos, Lauren Brace, Emma Yang, Tiantian Feng, Xuan Shi, Shrikanth Narayanan, Teresa Satterfield, Ying Xu
The paper introduces ECI2D, a method that distills self-generated in-context demonstrations by converting language rationales into executable programs and applying causal interventions to ensure transfer invariance across logically similar samples. It addresses the problem of observational equivalence, where both faithful and spurious rationales can produce correct outputs, by filtering demonstrations that remain effective under causal manipulation. Experiments on math, logic, and commonsense reasoning tasks show that ECI2D outperforms observation-based in-context learning, achieving over 7‑point gains when test distributions shift while using far fewer demonstrations.
By Minming Ye, Min Liu, Haoang Chi, Ciheng Wu, Naiyang Guan
The paper introduces Un-OPD, an unbiased on‑policy distillation framework for block diffusion language models that addresses two key optimization biases: context misalignment from mismatched block boundaries and an intrinsic bias that causes overconfidence collapse. Un-OPD employs a boundary‑aware step filtering strategy to remove misaligned decoding steps and a support‑rebalanced confidence calibration to moderate optimization intensity at high‑support positions. The method also includes a rollout reuse mechanism to cut generation overhead, and experiments on math reasoning and code generation tasks demonstrate stable training, improved performance, and roughly 50% faster wall‑clock time.
By Zaiquan Yang, Fei Wei, Yong Wang, Yudong Han, Yiyu Li, Zhuofan Zong, Gerhard Petrus Hancke, Xiangxiang Chu, Rynson WH Lau
DK‑VLM introduces Dual‑Kernel Attention, where a visual token’s attention is computed from both its input position and its geometric position derived from 3D points and camera poses. A semantic kernel preserves pretrained attention while a geometric kernel modulates attention via point displacements and relative view transformations, with a learned gate fusing the two. The model also employs Reference‑Frame Composition supervision to update its reference frame across successive judgments, achieving top scores on six spatial benchmarks and strong performance on VSI‑Bench and ReVSI‑32.
By Han Su, Meng Tian, Tianyu Huang, Mingyu Gao, Lennox Wei, Dongxu Li, Jianhua Han, Hang Xu, Xiaohe Wu, Wangmeng Zuo
The paper introduces Future‑L1, an interleaved latent visual reasoning framework that alternates between language tokens and continuous latent visual spans during autoregressive decoding for video event prediction. By training on a curated dataset (Future‑L1‑50K) and optimizing latent trajectories with a latent‑aware RL objective (LA‑DAPO), Future‑L1 achieves state‑of‑the‑art results on FutureBench and TwiFF‑Bench, outperforming prior models by significant margins. The work demonstrates that preserving intermediate visual semantics in latent space, rather than verbalizing every reasoning step, leads to more accurate and visually grounded future predictions.
By Tianxiang Jiang, Linquan Wu, Sheng Xia, Songze Li, Ziang Yan, Haoyu Yang, Yu Qiao, Yi Wang
The paper introduces the concept of the effective Speed of Coherent Cognition (S_C^{eff}) as a benchmark for how fast an AI system can adapt to changing reality while maintaining reliable behavior. It defines a challenge where systems must infer reliable actions for increasingly distant future states across a growing observation–reality gap, and tests this using reinforcement‑learning agents trained at a baseline speed of reality (S_R). The study compares Transformer and gated‑MLP baselines with a novel Co^4 architecture inspired by cellular mechanisms of conscious processing, finding that Co^4 achieves a substantially higher S_C^{eff} and thus supports the proposed scaling principle for reliable AI.
By Ahsan Adeel, Talha Bin Riaz
FedSkillGit is a federated framework that enables large language model agents to evolve reusable natural‑language skills without sharing raw data. Instead of uploading entire skill documents, clients submit bounded patches and local experience summaries, which are filtered for privacy, deduplicated, and merged into a global skill candidate. The server distributes the candidate for client‑side validation, publishing it only when performance improves, and experiments show it outperforms baselines by up to 22% while approaching centralized results.
By Boxun Lan, Jiamei Li, Weijian Li, Chengyu Yang, Zhuang Qi, Chengcheng Wan, Xiaohong Chen, Ming Hu
DropKV introduces a new KV‑cache eviction framework that decouples residual‑output perturbation into independent per‑token scores, avoiding the NP‑hard combinatorial optimization of prior methods. The approach offers a provable constant‑factor approximation under a cone condition verified on three long‑context LLMs, and a high‑performance Triton‑based implementation that yields a 19.8× speedup in scoring kernels and up to 9.9% end‑to‑end prefill speedup. Benchmarks on RULER and LongBench show DropKV consistently outperforms state‑of‑the‑art baselines at the same cache budgets, achieving the lowest inference latency among eviction strategies while remaining faster than dynamic cache baselines.
By Aozhong Zhang, Selcuk Gurses, Yanxia Deng, Naigang Wang, Chi-Chun Liu, Davis Wertheimer, Derrick Liu, Xin Li, Zi Yang, Felix X. -F. Ye, Penghang Yin
FDE-Gym is a new benchmark that evaluates AI agents on 95 end‑to‑end delivery tasks drawn from real customer scenarios across 18 sectors. Unlike traditional benchmarks that use well‑specified tasks, FDE-Gym places agents in simulated customer environments where they must clarify ambiguous requirements, interact with stakeholders, and uncover unstated constraints to improve a customer’s success metric. In tests, the best model configuration achieved a mean score of 0.56 and solved only 14 of the 95 tasks, highlighting a significant gap between well‑specified problem solving and real‑world AI delivery.
By Dongxu Yang, Hongda Yan, Renxin Wang, Baoli Zhang, Haojia Zuo, Ying Zhao, Youmin Chen, Xiaojian Liao, Congming Gao, Tianyang Jiang, Yixiao Chen, Qing Wang, Zhiguo Yu, Xiaqing Li, Yuhao Zhang, Jiawen Sun, Renzhe Xu, De Mi, Tiancheng Wu, Xianya Fu, Xuanlu Zhao, Bo Feng, Meixuan Ren, Guanming Lin, Ziheng Zhou, Xinhua Wang
The paper introduces Prosodic Adjustment with Conversational Context (PACC), a deep‑learning method that converts read speech into natural conversational speech. PACC analyzes and modifies prosodic features—intonation, stress, and rhythm—using HiFi‑GAN for synthesis, aiming to preserve naturalness and intelligibility while keeping computational load low for real‑time use. Experiments show significant improvements in speech conversion, higher model accuracy, and better Mean Opinion Score (MOS) results, establishing new benchmarks for the task.
By Parshav Singla, Agnik Banerjee, Aaditya Arora, Shruti Aggarwal, Anil Kumar Verma, Vikram C M, Raj Prakash Gohil, Gopal Kumar Agarwal