SP$^2$ec is a training‑free bandit algorithm that adapts the draft configuration during generation for Vision‑Language Models, exploiting a single‑peak throughput pattern observed in existing self‑speculative methods. It minimizes a novel wall‑time regret metric, theoretically guaranteeing near‑optimal draft selection with negligible exploration cost. Experiments on Qwen2.5‑VL‑7B‑Instruct and LLaVA‑Next‑Mistral‑7B show consistent improvements of up to 11% over prior self‑speculative methods across visual and textual tasks.
By Yuqi Huang, Xingyao Li, Yunlong Hou, Fengzhuo Zhang, Jiachun Pan, Vincent Y. F. Tan
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
OrbiSim is a new robotic simulation paradigm that treats world models as a fully differentiable physics engine for embodied intelligence. It integrates structured scene assets, neural dynamics, and reinforcement learning into a single, physically grounded, end‑to‑end differentiable pipeline. The approach enables tasks such as differentiable contact modeling, gradient‑based policy optimization with sparse rewards, and intuitive physical inference, and it outperforms state‑of‑the‑art world models in predictive fidelity and control performance.
By Jiajian Li, Jingyuan Huang, Junru Gong, Qi Wang, Xiaokang Yang, Yunbo Wang
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
Cross-Centaur is a new language‑model framework that predicts individual decision‑making across diverse psychological tasks. It is trained on the Cross‑Psych‑101 dataset, which contains 591 participants and over 260,000 decisions from 12 different tasks. The model outperforms existing LLMs and the Centaur baseline on held‑out participants, and its internal representations appear to capture participant‑specific behavioral tendencies that transfer across tasks.
By Hikaru Asano, Hidezo Suganuma, Ryutaro Mori, Yukino Baba
PragReST is a self‑supervised framework that enhances large language models’ pragmatic reasoning by generating counterfactual reasoning traces and fine‑tuning models with supervised learning and reinforcement learning, all without human‑labeled data. The method constructs pragmatic QA data, trains models to internalize counterfactual reasoning, and demonstrates significant accuracy gains on four pragmatic benchmarks compared to backbone models and other baselines. Error analysis shows that counterfactual reasoning reduces mistakes from failing to contrast observed utterances with plausible alternatives, and removing it markedly degrades performance while preserving out‑of‑domain general‑knowledge and mathematical reasoning.
By Jihyung Park, Minchao Huang, Leqi Liu, Elias Stengel-Eskin
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
The paper introduces TIDER, a time‑aware retrieval framework that uses a reinforcement‑learning trained temporal interval locator to identify relevant time windows for evidence retrieval in long‑horizon agent memory. By combining temporally grounded candidates with semantic retrieval and reranking them by proximity to the inferred windows, TIDER mitigates the displacement of target memories caused by temporally mismatched yet semantically similar entries. The lightweight 1.7B locator achieves state‑of‑the‑art accuracy on LifeBench and LoCoMo benchmarks, demonstrating the importance of temporal cues for effective evidence retrieval.
By Dabin Sheng, Zhe Wu, Jinming Zhao, Junliang Xing, Yuanchun Shi
The paper presents the first adversarial linear bandit algorithm based on Follow-the-Perturbed-Leader (FTPL) that achieves √T regret over general action sets. It attains a ≥O(d^{5/4}√T) regret using only linear optimization oracles, and a near‑optimal ≥O(d√T) regret when an optimal‑design exploration distribution is available. The authors introduce adaptive perturbation distributions—including Gaussian, random‑weighting, and elliptical perturbations—and develop a systematic analysis of the FTPL stability term via Fisher information, extending the FTPL framework beyond previous approaches.
By Chen-Yu Wei
The paper introduces PrefMind, an online dual‑agent reinforcement learning framework that jointly trains a response Generator and a preference Critic to explicitly learn and apply human preference judgments. By having the agents discuss and correct each other’s evaluations of candidate responses, PrefMind embeds preference understanding directly into the generation process rather than relying solely on a scalar reward model. Experiments on models of varying sizes demonstrate gains in both preference judgment accuracy and downstream response quality.
By Ziyun Zou, Jiwoo Hong, Yifan Lu, Shao Tang, Weiyi Wu, Aman Gupta, Jin Tian, Imran Razzak
The paper introduces Self-HPO, a recursive framework that alternates between optimizing the harness and the policy of large language model agents for coding tasks. By letting the current policy propose new harness candidates based on its failures, the method aligns harness evolution with policy improvement, leading to consistent performance gains across three benchmarks. Experiments show that Self-HPO raises the average resolve rate of Qwen3.5-35B-A3B from 57.8% to 64.8% and improves reliability and capability metrics, with benefits that generalize to out‑of‑distribution settings.
By Zhepei Wei, Wenlin Yao, Baolin Peng, Danqing Wang, Qianhui Wu, Hao Cheng, Yu Meng, Jianfeng Gao
COMET is a task‑agnostic continual learning framework for vision‑language models that addresses modal‑consistency drift, representation‑space interference, and per‑task capacity shortfall. It uses a frozen teacher for feature reference, Tri‑Affinity Distillation to preserve image‑image, text‑text, and image‑text geometry, and Student‑Aware Cross‑Expert Distillation to align new experts with prior ones, while a contextual bandit and reconstruction router manage capacity and inference. Experiments on the X‑TAIL benchmark show that each component improves continual learning performance, demonstrating that separating feature reference, expert shaping, routing, and capacity control enables effective multimodal adaptation.
By Seungyoon Woo, Keonhee Park, Gunhee Kim
The paper investigates how small changes in prompt characteristics, such as demographic attributes, can alter the semantics and tone of large language model outputs. It shows that standard alignment methods like reinforcement learning can worsen counterfactual bias by increasing response quality differences across demographics. To mitigate this, the authors propose a policy‑gradient approach with a group‑level reward that operates on matched counterfactual prompt pairs, and they validate its effectiveness on fairness benchmarks while maintaining overall response quality.
By Basar Demir, Emily Shen, Ksheeraj Sai Vepuri, Sourabh Kulkarni, Harihar Subramanyam, Ankit Jain, Nan Jiang, Chirag Nagpal
QD-Rec is a quality–diversity approach that expands recommendation systems by generating diverse, high‑quality suggestions paired with natural‑language hypotheses. It uses a contextual bandit to select items and a foundation model to craft reasons, storing new recommendations in an archive when they are novel or superior to existing ones. Experiments on mergers & acquisitions, customer targeting, and content recommendation show that QD‑Rec achieves higher mean quality and larger archives than baseline methods.
By Takashi Ishida, Taro Makino, Issa Sugiura
HydroShear is a hydroelastic tactile simulator that models stick‑slip transitions, path‑dependent force and shear buildup, and full SE(3) object‑sensor interactions. It uses Signed Distance Functions to track indenter displacements on a sensor membrane, producing physics‑based force fields from arbitrary watertight geometries while remaining engine‑agnostic. Experiments on GelSight Minis show HydroShear reproduces real tactile shear more faithfully than existing methods, enabling zero‑shot sim‑to‑real reinforcement learning with a 93% success rate across four dexterous tasks.
By An Dang, Jayjun Lee, Mustafa Mukadam, X. Alice Wu, Bernadette Bucher, Manikantan Nambi, Nima Fazeli
Group-Graph Policy Optimization (G2PO) is a novel group-based reinforcement learning algorithm designed for long-horizon agentic tasks. It transforms linear interaction trajectories into a global state‑transition graph, aggregates identical observations across trajectories to reduce sampling variance, and redefines actions as transitions between state nodes. By employing an edge‑centric advantage estimation and globally standardizing TD errors, G2PO identifies and prioritizes critical transitions, achieving up to 22.2% higher success rates than state‑of‑the‑art baselines on benchmarks such as WebShop, ALFWorld, and AppWorld.
By Yunan Wang, Minghui Song, Zihan Zhang, Shaohan Huang, Haizhen Huang, Furu Wei, Weiwei Deng, Feng Sun, Qi Zhang
arXiv:2610.22438v1 Announce Type: new
Abstract: Studies of what fine-tuning does to a language model often summarise the weight update $\Delta W$ by its rank. We compare the rank measures used for th...
By Muhammad Usman Altaf
arXiv:2610.22695v1 Announce Type: new
Abstract: We study personalized LLM routing, where a router selects among heterogeneous models with different costs and capabilities. The router must minimize cu...
By Ameya Anjarlekar, Rasoul Etesami, R Srikant
arXiv:2610.23402v1 Announce Type: new
Abstract: In the classical principal-agent setting, the principal delegates some task to the agent, who receives a payment conditioned on the outcome of the task...
By Elif Y{\i}lmaz, Jakub T{\l}uczek, Victor Villin, Christos Dimitrakakis
arXiv:2610.21711v1 Announce Type: cross
Abstract: Repeated Bertrand competition is a stringent benchmark for algorithmic pricing: in homogeneous-goods markets, standard learning procedures typically...
By Martin Bichler, Philip Hierhager, Kassian K\"ock, Fabian Pieroth