The paper introduces the concept of proof‑carrying cognition, aiming to close the verification gap in language‑model reasoning by using reality‑settled rewards. It presents a theoretical framework linking verifier‑gold correlation to compute‑capability trade‑offs, demonstrates that unsound verifiers degrade under best‑of‑N selection while sound verifiers improve, and proposes a new benchmark metric, Soundness‑under‑Pressure, for evaluating reality‑settled reasoning systems.
By Eshwar Reddy M, Sourav Karmakar
The paper introduces the Belief-State Engine (BSE), an inference module that supplies a large language model (LLM) with a Bayesian posterior over hidden states in a partially observable Markov decision process (POMDP). By keeping the raw action‑observation log hidden from the LLM, the BSE ensures the agent behaves as a sound Markov policy on the belief MDP, thereby inheriting classical POMDP optimality guarantees. Experiments on the Tiger POMDP and a red‑team attack‑graph task show that BSE‑augmented agents outperform six baselines in task return, belief calibration, and decision consistency.
By Arnab Chattopadhayay, Debdipta Halder
HiRAD is a hierarchical reinforcement learning framework designed for continuous-space routing of large-scale AGV fleets, offering real-time guarantees. It introduces a step-level spatiotemporal representation, separates heading selection from velocity control to shrink the action space, and employs an asynchronous event-driven decision pipeline that reduces inference complexity from O(n²) to O(n) and cuts per-step latency by up to 71%. Experiments on random graphs and two warehouse maps show that HiRAD decreases makespan by 45% to 63% and shortens overall runtime.
By Yunjie Huang, Ruizhong Wu, Mengxuan Zhang, Frodo Kin Sun Chan, Yan Nei Law, Lei Li
The paper presents the KuaiRP series of role‑playing models, detailing a multi‑stage training pipeline that balances deep domain knowledge injection with the preservation of general agent capabilities. The approach includes a standardized character template, a supervised fine‑tuning (SFT) data pipeline, a rule‑based reward function for reinforcement learning, and a novel two‑stage on‑policy distillation (OPD) with cumulative‑divergence decay (CDD) to recover general skills. Experimental results show that the models achieve state‑of‑the‑art role‑playing fidelity in target domains while maintaining low deployment costs.
By Yipeng Wang, Ziwei Zhang, Jiahui Zhang, Qi Gan, Kai Sheng
RetroThinker is a multi-stage post‑training framework that enhances SpeechLLMs by enabling them to self‑verify and forward‑correct Chain‑of‑Thought reasoning steps during inference. It combines supervised fine‑tuning on curated retrospective thinking data with length‑based direct preference optimization to improve reasoning while the user speaks. On the GSM8K benchmark, RetroThinker achieves an 11% absolute accuracy gain over non‑retrospective baselines while maintaining comparable latency.
By Yi-Jen Shih, Puyuan Peng, Abdelrahman Mohamed, David Harwath
Negative Self-Distillation (NSD) is a new framework for improving large language models by encouraging them to diverge from their own flawed reasoning rather than imitate privileged solutions. Unlike On-Policy Self-Distillation, which can suppress uncertainty and exploratory behavior, NSD generates a question‑specific negative condition (e.g., a careless reasoner) and uses a dynamic gating mechanism to target only reasoning‑critical tokens for penalization. This approach preserves foundational language capabilities while consistently outperforming OPSD and other label‑free self‑bootstrapping reinforcement learning baselines.
By Rongcan Pei, Zhepei Wei, Shuyao Xu, Xinyu Zhu, Wei-Lin Chen, Yu Meng
The paper introduces a multi‑signal pipeline for detecting hallucinations in large language models, combining fine‑tuned DeBERTa‑v3 classification, Monte Carlo Dropout uncertainty, and temperature‑scaled calibration. On the HaluEval benchmark it achieves high performance (F1 = 0.915, AUROC = 0.977) across QA, summarization, and dialogue, and shows that 25 % of training data yields 77 % of full‑data performance. The authors also demonstrate that applying Direct Preference Optimization to a Qwen2.5‑0.5B generator cuts hallucination rates from 85.5 % to 37.7 %, and that domain‑specific fine‑tuning (PubMedBERT on SciFact) outperforms general‑domain models for biomedical text.
By Varun Teja Chundru, Debasmita Biswas
The paper surveys how large language models (LLMs) are being applied in legal tasks such as judgement prediction, document analysis, and drafting. It reviews the benefits of automation while highlighting legal challenges like privacy, bias, and explainability. The authors also discuss data resources for legal domain specialization and outline future research directions.
By Zhongxiang Sun
The paper introduces GenQAS, a tensor network‑guided reinforcement learning framework that uses a learned local transition model to generate synthetic circuit transitions for prioritized generative replay. By mixing these synthetic transitions with real experience during Double Deep Q‑Network updates, GenQAS addresses sample starvation in quantum architecture search. Across benchmarks ranging from 6 to 15 qubits, the method improves success probabilities, identifies compact circuits, and reduces steps to chemical accuracy by up to 92.7%.
By Akash Kundu, Amit Kumar Jaiswal, Sebastian Feld, Prayag Tiwari
arXiv:2609. 10981v1 Announce Type: new Abstract: Bakhtiari, Lattimore and Szepesv\'ari (COLT 2025) proved that Thompson sampling (TS) has Bayesian regret $\tilde O(d^{5/2}\sqrt n)$ for bandit convex optimisation with convex \emph{monotone} ridge losses $f(x)=\ell(\ip{x}{\theta})$, and asked whether monotonicity of the link is necessary.
By Xuan Li
LOCUS is a post‑training technique that selects a task‑aware low‑rank adaptation subspace to reduce token generation cost while preserving a model’s preference alignment. By updating only a tiny fraction of parameters (0.24–0.28 %) on 3 B‑parameter backbones, LOCUS cuts continuation length by up to 39.84 % on Pythia‑2.8B and 14.87–17.58 % on Qwen2.5‑3B, without materially affecting internal preference diagnostics. The method demonstrates that careful subspace selection can make large language models more token‑efficient without sacrificing utility.
By Dongfang Zhao
The paper introduces T1, a 122‑billion‑parameter Mixture‑of‑Experts model trained with reinforcement learning to perform long‑horizon terminal tasks such as coding and scientific discovery. T1 operates a real shell in a cloud sandbox, making over 300 tool‑call turns per task and receiving rewards from task‑specific verifiers. The authors detail a training recipe that includes aggressive warm‑starting, TITO construction with drift repair, and rollout‑routing replay, achieving significant performance gains on Terminal‑Bench 2.1 and surpassing GPT‑5.4 and GLM‑5.1 on the Long‑Horizon Terminal Bench.
By Junyao Yang, Yucheng Shi, Zhongzhi Li, Ruhan Wang, Zongxia Li, Haitao Mi, Leowei Liang
MindTopo is a benchmark that tests foundation models on topological reasoning, covering five cognitive properties—continuity, separation, order, enclosure, and knots—across two cognitive levels: reasoning and planning. It contains 11,030 instances from 13 procedurally generated task types, and evaluates 14 multimodal large language models, including agent configurations with image and video generation. Results show that models perform better on reasoning than planning, and even the best model lags far behind human performance, with fine‑tuning and reinforcement learning improving reasoning more than planning.
By Yunfei Ge, Anbang Liu, Qineng Wang, Johnalbert Garnica, Jianwen Lyu, Zihan Wang, Reuben Tan, Jianfeng Gao, Ruohan Zhang, Yining Hong, Jiajun Wu, Manling Li
The paper presents SEAR, a system for Task 2 of the second Multilingual Conversational Speech Language Model Challenge. It adapts Qwen3-Omni-30B-A3B-Instruct with a segment‑evidence‑aware pipeline that converts timestamped ASR into event spans, expands and crops them, and then generates both semantic and acoustic multiple‑choice questions using Qwen3.6-27B and Gemini 3.1 Flash‑Lite. After rigorous checks, 359,825 verified MCQs across 21 languages and accents are produced, and the system achieves 90.92% accuracy on the official evaluation set.
By Huy Hoang Le, Long-Bao Nguyen, Minh Tri Dao
arXiv:2609. 10866v1 Announce Type: new Abstract: Reinforcement learning (RL) agents deployed in real-world environments are often vulnerable to adversarial perturbations in state observations, creating risks in safety-critical applications.
By Tong Li, Saunak Kumar Panda, Yisha Xiang
The paper critiques the common reinforcement‑learning approach of sampling tool subsets when the full set of tools is enumerable, showing that sampling leads to degraded policy estimates and increased reward sparsity in genomic reasoning tasks. It proposes Full‑Group Policy Optimization (FGPO), which evaluates every tool subset and precomputes rewards in a table, thereby eliminating the need for frozen‑reasoner calls during training. Experiments across five frozen reasoners and three genomic benchmarks demonstrate that FGPO consistently outperforms GRPO, improving average scores by 6.75 points and reducing the number of invoked tools per question.
By Haoyue Liu, Xiaoyu Ma, Ye Chen, Zhichao Wang, Xiaoying Tang
The paper presents a method where self‑evolving scientific agents design explicit, neural‑network‑free white‑box controllers for fluid dynamics tasks. By iteratively interpreting simulation data, the agents accumulate control knowledge and refine controller code, ultimately achieving robust control of an underactuated two‑joint swimmer in unsteady flows. The resulting controllers generalize across varying target positions, wake geometries, cylinder counts, and inflow speeds, and 2D control priors successfully transfer to accelerate 3D adaptation.
By Boai Sun, Wenjin Guo, Zongmin Yu, Liu Yang
The paper tackles two main issues in multi‑subject video generation—uncontrollable fidelity strength and semantic drift—by exploiting intrinsic attention patterns in Diffusion Transformers. It introduces an Intrinsic Spatial Grounding Map (ISGM) that accurately locates reference subjects and a Dual‑phase Intrinsic Attention Leveraging (DIAL) framework that uses ISGM during both training and inference. DIAL guides attention in low‑noise stages for precise fidelity control and builds preference pairs in high‑noise stages for reinforcement learning, resulting in superior identity consistency and controllable fidelity on the OpenS2V‑Eval benchmark.
By Niange Yu, Ye Tian, Biaolong Chen, Miao Lu, Aixi Zhang, Hao Jiang, Yunhai Tong, Pipei Huang
The paper introduces a statistical framework for Inverse Entropy-regularized Reinforcement Learning that resolves the non-uniqueness of reward functions by combining entropy regularization with a least-squares reconstruction of the reward from the soft Bellman residual. It models expert demonstrations as a Markov chain, estimates the expert policy via penalized maximum likelihood, and provides high-probability bounds on the excess Kullback–Leibler divergence between the estimated and true policies. These results yield non-asymptotic minimax optimal convergence rates for the least-squares reward function, highlighting the trade-offs among smoothing, model complexity, and sample size.
By Denis Belomestny, Alexey Naumov, Artemy Rubtsov, Sergey Samsonov
The paper presents a gamified 20Q-style recommender for cybersecurity education that uses reinforcement learning and explainable AI to guide learners through interactive questioning. By acting as a knowledgeable questioner, the system narrows down user-described security scenarios, identifies the underlying threat, and transparently explains its reasoning. The authors detail the system architecture, algorithmic foundations, and provide case studies covering attack vectors such as the Cyber Kill Chain, phishing, ransomware, and web application vulnerabilities.
By Mary Nusrat, Sarfuddin Bhuiyan, Gahangir Hossain