Daily AI Papers — October 25, 2026

1. EvalResearchBench: Can AI Agents Build Their Own Feedback for Recursive Self-Improvement?

Authors: Yaolun Zhang, Tianyi Xu, Yujie Zhao, Jishen Zhao, Qingyun Wu, Huazheng Wang arXiv: arxiv.org/abs/2610.04184 Summary: EvalResearchBench tests whether AI agents can design compact, executable evaluators for coding, co-work, and reasoning under fixed time and API budgets. Across 9 researcher agents, 13 candidate models, and 14 target benchmarks, the best evaluators agree with target orderings on about 75% of model pairs, below a 91% target-agreement ceiling, while human-designed public-task samples remain strong baselines.

2. F2G-Pose: Geometry-Aware Foundation Feature Lifting for Direct RGB-D Category-Level Object Pose Estimation

Authors: Jinyu Zhang, Haitao Lin, Jiashu Hou, Xiangyang Xue, Yanwei Fu arXiv: arxiv.org/abs/2610.31623 Summary: F2G-Pose lifts visual foundation features into partial point clouds and directly predicts category-level object pose, metric size, and completed camera-frame shape from one RGB-D observation. With its task-specific network trained on synthetic SOPE data and a frozen pretrained RADIO backbone, it reports state-of-the-art SOPE and ROPE accuracy, 28 FPS pose-only inference, and transfer to real objects and manipulation settings.

3. UniAVC: Unified Audio-Video Creation via Progressive Training and Asymmetric Scheduling

Authors: Kien T. Pham, Yuhao Liu, Chong Hou Choi, Zhenqi He, Yanghao Wang, Chenyang Lei, Rui Liu, Qifeng Chen, Long Chen arXiv: arxiv.org/abs/2610.24791 Summary: UniAVC first trains audio and video branches on abundant unimodal data, then uses paired data to learn cross-modal correspondence for joint generation and editing. Independent modality noise schedules and clean-modality conditioning support asymmetric generation, intent-aware editing, and selective self-refinement from one set of weights without task-specific adapters.

4. What Makes a Good Unmasking Policy? A Controlled Study of Learned Unmasking in Diffusion Language Models

Authors: Jinya Sakurai, Xun Xu, Jaehong Yoon arXiv: arxiv.org/abs/2610.31576 Summary: This controlled study separates the effects of teacher targets, training states, policy inputs, and overhead in learned token-unmasking policies for diffusion language models. Its 856K-parameter policy uses trajectory-contraction targets, confidence, and hidden states to achieve the strongest reported GSM8K accuracy-versus-evaluation frontier and throughput on LLaDA-8B-Instruct, with competitive results on other tasks and substantial transfer to Dream-7B-Instruct.

5. Foundation4D: Towards a unified perception and reconstruction model for autonomous vehicles

Authors: NVIDIA: Sérgio Agostinho, Maximilian Apodaca, Emmanuel Attia, Guillem Braso, George Dong, Ilya Fedorov, Michele Fenzi, Gang Huang, Jiahui Huang, Shengyu Huang, Bingxin Ke, Laura Leal-Taixé, Kuan-Ting Lee, Mingxuan Liu, Tian Liu, Tim Meinhardt, Aljoša Ošep, Naveen Kumar Rai, Kefei Ren, Cristiano Saltori, Keqiang Wu, Ning Xu, Fei Xue, Jin-Bey Yu, Lei Zhang, Qunjie Zhou arXiv: arxiv.org/abs/2610.30340 Summary: Foundation4D unifies geometry estimation, 3D detection, semantic occupancy, and Gaussian reconstruction from multi-camera driving video with a shared spatiotemporal backbone. The authors release a 1,253-clip benchmark, report leading results among compared methods after benchmark-specific fine-tuning, and show that joint training improves detection and geometry while supporting lidarless driving-scene reconstruction.

6. Who Takes the Lead? Model Ability and Task-Conditioned Collaboration with Humans

Authors: Weien Li, Haochen Liu, Zeyu Li, Rui Song, Zixuan Dong, Ye Yuan arXiv: arxiv.org/abs/2610.27148 Summary: This study uses a shared-control maze to distinguish appropriate division of labor from poor collaboration by varying whether a human or an LLM agent has the shorter route to a common goal. Across four models, stronger solo solvers involve humans selectively, including roughly 80% of initial decisions when humans have a route advantage, whereas the weakest model lacks that selectivity.

7. MambaAct: Efficient Mamba Action Experts for Flow-Matching Vision-Language-Action Models

Authors: Khuyen Pham, Tran Nguyen Le, Edoardo Fazzari, Cesare Stefanini arXiv: arxiv.org/abs/2610.22951 Summary: MambaAct replaces the transformer action expert in a flow-matching vision-language-action policy with a compact Mamba expert and compressed visual-language conditioning while retaining the pretrained prefix encoder. It reports a 10-fold smaller action expert, denoising-latency reductions of 64.8% on LIBERO and 58.8% in real-world deployment, and a 1.08-percentage-point mean LIBERO success gain across three fine-tuning seeds.

8. TILDE: Teacher-in-the-Loop Distillation from Embodied Foundation Models

Authors: Yuxiang Lu, Yantong Liu, Jinghua Hou, Xueli Sun, Ruihua Han, Zhe Liu, Hengshuang Zhao arXiv: arxiv.org/abs/2610.25376 Summary: TILDE transfers embodied foundation-model capabilities to student policies by combining on-policy teacher supervision with teacher-executed recovery trajectories. Its AutoDAgger pipeline uses a robotic reward model to trigger interventions without human monitoring, collecting recovery experience more than twice as fast as a human-in-the-loop method and improving performance by up to 4.1 times over behavior cloning across three real-world tasks.

9. HyperDrone: Simulation-to-Reality Transfer at Millions of Frames per Second

Authors: Jonas Eschmann, Giuseppe Loianno arXiv: arxiv.org/abs/2610.28950 Summary: HyperDrone combines GPU multirotor dynamics with the HyperLux ray-tracing engine to render rich visual observations at millions of frames per second on consumer GPUs, reporting more than 20-fold faster batch simulation than compared state-of-the-art quadrotor simulators. The authors demonstrate simulation-to-reality transfer by training a vision-based yaw-control policy entirely in simulation and running it onboard a camera-equipped Crazyflie quadrotor.

10. Diffusion Language Models Change Speed Across Tasks and Languages

Authors: Mohsen Fayyaz, Ali Modarressi, Daniel Israel, Lucas Bandarkar, Nanyun Peng arXiv: arxiv.org/abs/2610.25835 Summary: An analysis of two diffusion language models across 10 tasks finds that forward passes per token depend on output novelty, word rarity, task type, and language tokenization, with story writing costing over three times as many passes as grammar correction in DiffusionGemma. Decoding effort also tracks correctness, and weighting math-answer candidates by decoding speed improves over majority voting by up to 4.2 points without extra inference cost.

11. LLM-Attack Suite: A Unified Red-Team Framework for Privacy Evaluation of Split and Distributed LLM Systems

Authors: Georgios Politis, Evangelos Pappas arXiv: arxiv.org/abs/2610.22722 Summary: LLM-Attack Suite standardizes privacy evaluations of split language models by recording attacker access, datasets, scoring rules, undefended controls, and untested attacks while saving predictions for independent scoring. In a controlled GPT-2 experiment, exposing returned training gradients raises token recovery from 94.20% to 97.38% and exact-document recovery from 13.71% to 37.77%, illustrating how attacker access changes privacy conclusions.

12. SplatLab: Autonomous Scientific Discovery for 3DGS Research

Authors: Seemandhar Jain, Manmohan Chandraker arXiv: arxiv.org/abs/2610.30468 Summary: SplatLab autonomously reproduces 3D Gaussian Splatting baselines, proposes and evaluates extensions, runs ablations, and writes manuscripts while binding every reported number to a logged training run. Its PSNR-targeted experiments report gains of 0.18 to 3.78 dB over reproduced targets, using evolutionary tournaments and a standing falsifier to select candidates and test claims.

13. IF-MAP: Influence-Function-Inspired Agent Prioritization in Multi-Agent Systems

Authors: Jie Mao, Pingchen Lu, Zikun Qu, Jiahang Sun, Xiangyi Wang, Zhiwei Shang, Zhiyong Wang, Guojun Xiong, Zhongxiang Dai arXiv: arxiv.org/abs/2610.24383 Summary: IF-MAP uses controlled semantic prompt perturbations and local response models to prioritize agents whose improvement is most likely to increase end-to-end multi-agent performance, without model parameters or gradients. Across six workflows, optimizing selected prompts yields a 3.81-percentage-point macro-average gain versus 1.85 for uniform selection, with substantial token savings relative to improving every agent.

14. TARSphere: A Robot-Oriented Full-Stack Framework for Long-Horizon Embodied Agents in Evolving Physical and Social Environments

Authors: Yicheng Xu, Zhikai Xu, Ziyuan Chen, Yuhang Dong, Tao Zhang, Zhucun Xue, Yong Liu, Jiangning Zhang arXiv: arxiv.org/abs/2610.30043 Summary: TARSphere combines physics simulation, embodied non-player characters, constrained dialogue, hierarchical memory, and executable skills to evaluate long-horizon robot agents in changing physical and social environments. Its benchmark contains 360 episodes across 30 indoor scenes, and evaluation of eight multimodal models exposes persistent difficulties with task updates, human interaction, collisions, and falls.

15. Skeleton-Grounded Latent Action Model for Robot World Models

Authors: Qi Lv, Jiayi Lin, Yuxin Jiang, Jianzong Wang, Xiang Deng, Mike Shou Zheng arXiv: arxiv.org/abs/2610.31474 Summary: SkeLAM separates skeleton-grounded robot kinematics from residual visual dynamics and learns a lightweight adapter from robot controls to support controllable world-model rollouts. On DROID it reports 19.5% lower end-effector displacement error and 11.1% lower LPIPS than DreamDojo, alongside generated human-to-robot motion transfer and agreement between predicted and real robot policy success rates.

16. Bridging Exploration and Persona: Information-Directed Dueling for Test-Time Agent Personalization

Authors: Jie Mao, Pingchen Lu, Xiangyi Wang, Hengquan Huang, Shuang Qiu, Min Zhang, Zhongxiang Dai arXiv: arxiv.org/abs/2610.24385 Summary: ID-TAP personalizes LLM agents at test time through binary response comparisons, combining a query-conditioned neural utility model with a lightweight Bradley-Terry explorer and similarity-weighted warm starts. Across seven user-conditioned generation benchmarks with simulated preferences, it achieves the strongest aggregate performance among compared baselines under a fixed ten-round comparison budget.

17. SEAM: Spectral Embedding Alignment and Merging for Low-Resource TTS

Authors: Walaa Othman, Baher Mohammad, Stamatios Lefkimmiatis, Ammar Ali arXiv: arxiv.org/abs/2610.26252 Summary: SEAM initializes embeddings for unseen text-to-speech tokens without paired audio by preserving subtoken phonetic order with a spectral filter and aligning embeddings through closed-form ridge regression. Across three languages and two TTS architectures, it approaches or matches baseline performance with up to four times less training data than random initialization and mean pooling.

18. Realism VS Accuracy: Event Sequence Forecasting from a Generative Modeling Perspective

Authors: Dmitry Osin, Egor Surkov, Petr Mokrov, Igor Udovichenko, Ivan Karpukhin, Evgeny Burnaev, Egor Shvetsov, Alexander Korotin arXiv: arxiv.org/abs/2610.12586 Summary: This work frames multi-horizon event-sequence forecasting as conditional generative modeling and introduces a benchmark that measures both prediction accuracy and distributional realism. Its evaluation finds a trade-off between the two, while the proposed Wasserstein BaryBooster converts samples from a realism-strong diffusion model into a state-of-the-art predictor.

19. MMedAgent-Plus: Scaling Multimodal Medical Agents across Modalities, Tasks and Workflows

Authors: Ruoxin Wang, Yuanting Pan, Binxu Li, Ruiyang Ji, Yaqian Chen, Jiayuan Ding, Qifan Wang, Lifu Huang, Mingjie Li, Yu Zhang, Ehsan Adeli, Kilian Pohl, Zhe Xu, Zihao Lin, Haoyu Dong, Yixin Wang arXiv: arxiv.org/abs/2610.27233 Summary: MMedAgent-Plus combines an instruction-tuned vision-language planner with 18 specialist models across 12 medical tasks and 8 modalities to coordinate tool calls and synthesize grounded responses. In 350 end-to-end evaluation cases across seven tasks, it invokes the required specialist correctly in 90.6% of cases and reaches 0.98 final-answer accuracy, matching the dedicated leading specialist.

20. A Fix for Softmax and Cross-Entropy

Authors: Jonathan T. Barron, Dan Gnanapragasam, Michael Tschannen arXiv: arxiv.org/abs/2610.31642 Summary: This paper identifies numerical bugs in standard softmax and cross-entropy implementations that can zero the target gradient or produce nonzero gradients when student and teacher logits match, and supplies corrected implementations. The fixes eliminate the reported Slingshot phenomenon, accelerate grokking with weight decay, prevent accuracy collapse without it, and stop distillation students from drifting away from their teachers.