Reinforcement_Learning
-
UCVE: Graph Structured Retrieval for Compliance Validation
2026-10-27 14:35:00Emily Chen, Michelle Menachery, Himanshu Kumar, Rahul Nenavath, Amirfarrokh IranitalabAbstract
Structured compliance validation requires determining whether every obligation in a predefined requirement space is satisfied, violated, or unaddressed by a given document. This can be described as a coverage problem that existing retrieval and classification systems, by treating requirements as independent, are not designed to solve. We introduce \textbf{\ac{UCVE}}, a domain-agnostic framework that represents compliance requirement spaces as a \ac{CRG} and performs retrieval via spectral reranking over its reinforcement learning-weighted Laplacian. Edges between requirement nodes are learned via reinforcement learning with a traversal reward grounded in violation detection, reshaping the graph's community structure such that low-pass spectral filtering distinguishes relevant from irrelevant obligation clusters. Evaluating across three compliance domains, \ac{UCVE} averages 4.3\% matched-$K$ recall lift over flat cosine retrieval and consistently outperforms strong retrieval and graph-based baselines.
Abstract (translated)
URL
https://arxiv.org/abs/2610.37323
PDF
-
RPSD: Retrieval-Privileged On-Policy Self-Distillation with Optimal Transport
2026-10-27 14:08:38Dongyue Li, Zebang Cheng, Zhongyi Deng, Hongboxu, Yi Bin, Fei MaAbstract
Sequential recommendation predicts a user's next interaction from historical behavior. Recent LLM-based methods recast this task as generative recommendation and increasingly adopt reinforcement learning such as GRPO for post-training. However, these methods suffer from two limitations, namely credit-blind optimization, where a sequence-level reward is shared across all generated tokens, and collaborative-blind supervision, where each training signal is derived from a single user and ignores behavioral patterns shared across similar users. To address both issues, we propose Retrieval-Privileged on-policy Self-Distillation (RPSD). For credit-blind optimization, we introduce on-policy self-distillation (OPSD), in which a teacher with privileged knowledge re-evaluates the student's on-policy response, and the token-level confidence difference between the student and teacher reshapes the shared sequence-level advantage into token-specific credit. However, the privileged knowledge in vanilla OPSD is still derived from a single user, leaving collaborative-blind supervision unresolved. We therefore ground the teacher in collaborative behavioral evidence from similar users. Specifically, we use temporally aware optimal transport (OT) to compare interaction trajectories across users by jointly considering item similarity and temporal structure, and further apply diversity-aware selection to retain complementary trajectories as privileged evidence. In this way, on-policy self-distillation provides fine-grained token-level credit, while OT enriches the teacher with structured cross-user evidence. To the best of our knowledge, this is the first work to bring on-policy self-distillation to sequential recommendation and to use cross-user behavior as its privileged signal. Extensive experiments across multiple real-world datasets demonstrate that RPSD achieves state-of-the-art performance.
Abstract (translated)
URL
https://arxiv.org/abs/2610.37281
PDF
-
Uncovering Shortcut Amplification in Video RLVR through the Lens of Perceptual Uncertainty Collapse
2026-10-27 13:26:13Jiafeng Liang, Xianhao Ou, Tao Ren, Rongchuan Mu, Runxuan Liu, Dandan Tu, Ming Liu, See-Kiong Ng, Bing QinAbstract
Reinforcement learning with verifiable rewards (RLVR) has emerged as an effective post-training paradigm for enhancing the video reasoning capability of Large Multimodal Models (LMMs). Despite its empirical effectiveness, we show that outcome-level optimization can introduce an undesirable shortcut amplification effect, whereby models increasingly favor text-prior answers even when they are not supported by the visual evidence. To characterize this behavior, we investigate the training dynamics of video RLVR through the lens of perceptual uncertainty, analyzing how predictive uncertainty evolves under video-conditioned and text-only contexts. Our analysis identifies a perceptual uncertainty collapse phenomenon, in which RL training continuously sharpens the model's answer distribution while the uncertainty reduction attributable to visual evidence progressively diminishes. A further gradient-direction analysis reveals that such confidence enhancement becomes increasingly less aligned with visually grounded uncertainty reduction and is instead dominated by text-prior compression. Motivated by these findings, we propose Bifocal Self-distillation Policy Optimization (BSPO), a self-distillation framework that regulates policy updates with two complementary distributional signals. Specifically, BSPO derives empirical reasoning principles from intra-group correct and incorrect rollouts to identify reasoning-sensitive tokens, and introduces shortcut-aware visual perturbation to penalize predictions that remain invariant when task-relevant visual evidence is weakened. By adaptively reweighting the original RL advantage, BSPO preserves the task-solving benefits of RLVR while promoting updates that are both reasoning-sensitive and visually grounded. Extensive experiments validate the effectiveness of BSPO in improving video reasoning performance and strengthening visual faithfulness.
Abstract (translated)
URL
https://arxiv.org/abs/2610.37211
PDF
-
Remember Where You Walk: Learning Spatial Memory for End-to-End Legged Robot Exploration
2026-10-27 13:22:51Leon Magnus, Nico Bohlinger, Jan Peters, Oleg ArenzAbstract
Humanoids and other legged robots operating in complex environments must remember information that leaves their field of view and only becomes relevant much later. Reinforcement learning policies address this using memory networks such as LSTMs that are trained end-to-end with the policy to memorize such task-relevant aspects. Yet, such architectures often fail to preserve sufficient information over long horizons. In this work, we propose a network architecture based on Hadamard memory that specifically targets robot navigation. Our Local Spatial Matrix Memory (LSMM) enables RL agents to store generic task-relevant information at specific, position-based memory cells, enabling flexible and spatially grounded memorization. We evaluate LSMM on the task of autonomous maze exploration, where it matches the maze coverage of strong recurrent and finite-history baselines while being the only architecture that remembers the maze layout, reducing the reconstruction error threefold.
Abstract (translated)
URL
https://arxiv.org/abs/2610.37203
PDF
-
Action-Value Calibration via Continuation Sharing for Long-Horizon Agentic Reinforcement Learning
2026-10-27 13:16:08Yingjian Chen, Lang Zhou, Zixin Zhao, Pingfan Li, Weishi Zheng, Zhilin ZhaoAbstract
Reinforcement learning for long-horizon agents relies on feedback from sequences of interactions, making credit assignment to individual actions difficult. Existing methods such as GiGPO compare actions at shared states to compute relative advantages, but estimate each action's value from a single sampled continuation. The resulting local credit remains sensitive to decisions made later in the trajectory. To address this problem, we propose ConCal (Continuation Calibration), a recursive continuation-value estimator that incorporates continuation evidence already present in a rollout batch. ConCal matches visits from the same task by their incoming state-action-state transitions, combines shared continuation evidence with each visit's own transition estimate, and propagates calibrated values backward to refine earlier action-value estimates. This reduces dependence on individual sampled continuations and mitigates local credit distortions without additional rollouts. Held-out return diagnostics show lower error in within-group value differences and higher agreement with reference action orderings than estimates based on individual continuations. With Qwen2.5-1.5B-Instruct, ConCal improves success rates over GiGPO by 3.39 and 5.86 percentage points on ALFWorld and WebShop, respectively. Our code is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.37195
PDF
-
VGIQA: Fine-Grained Image Quality Assessment with Visual Grounding Rewards
2026-10-27 13:09:36Weichang Wang, Liyu Chen, Shuyue Guan, Jie Hu, Fei Gao, Wei Li, Nannan WangAbstract
Image quality assessment (IQA) aims to evaluate the perceptual visual quality of images. Recently, the rapid development of Multi-modal Large Language Models (MLLMs) has significantly advanced the IQA field. By integrating chain-of-thought (CoT) reasoning and reinforcement learning (RL), these models can generate more accurate and interpretable results while relying on less training data. However, existing MLLM-based methods are predominantly constrained to holistic, image-level evaluations and lack the capability to produce fine-grained, localized assessments, which substantially limits their practical applicability. To address this limitation, we propose VGIQA, a novel IQA model endowed with both local and global quality assessment capabilities. VGIQA is trained via a two-stage paradigm comprising a cold-start phase followed by reinforcement learning. Driven by carefully designed local quality assessment reward functions, our model exhibits exceptional image quality understanding on VGSR-2K, a new dataset we propose specifically for local quality evaluation, significantly outperforming existing open-source MLLMs such as Qwen2.5-VL-72B-Instruct and Qwen3-VL-32B-Instruct. Furthermore, the local quality evaluation task can effectively enhance the model's performance on the degradation perception task.
Abstract (translated)
URL
https://arxiv.org/abs/2610.37186
PDF
-
Learning to Look Beyond Screens: Transferable Visual Observation from GUI Grounding
2026-10-27 11:56:00Ruilin Yao, Xinyu Qiu, Tianyu Zou, Yi Rong, Shengwu Xiong, Shuheng Shen, Changhua MengAbstract
Visual perception is evolving from passive recognition toward active, task-oriented observation. We investigate GUI grounding as a scalable source of spatial supervision for learning observation behaviors that transfer beyond interfaces. We introduce LBS (Look Beyond Screens), a vision-language model trained through supervised initialization on teacher-generated GUI grounding trajectories, followed by reinforcement learning for multi-turn crop-and-inspect decisions. Under fixed inference settings, LBS models based on Qwen3.5-35B-A3B and Qwen3.8-27B consistently improve over their corresponding baselines. Gains beyond supervised fine-tuning transfer to PerceptionBench for both backbones and to HR-Bench and ChartQAPro for Qwen3.5-35B-A3B, supporting observation strategies that generalize beyond GUI-specific imitation. Further analysis suggests complementary roles for the two stages: SFT establishes visual inspection primitives, while RL improves their downstream effectiveness. These findings support GUI grounding as a practical supervision paradigm for transferable active visual perception and broader multimodal reasoning. Our code and model are publicly available at this https URL and this https URL, respectively.
Abstract (translated)
URL
https://arxiv.org/abs/2610.37093
PDF
-
Guided-RLVR: Guided Exploration in Reinforcement Learning with Verifiable Rewards
2026-10-27 11:13:13Bo Yang, Zishuai Zhang, Weihua Li, Qinnan Zhang, Sijia Wen, Hainan ZhangAbstract
Reinforcement learning with verifiable rewards (RLVR) has shown strong empirical gains in improving the reasoning ability of large language models (LLMs) on tasks such as mathematics and programming. However, recent studies have shown that RLVR rollouts often collapse onto a narrow set of reasoning trajectories, limiting the explored reasoning space. Motivated by findings that thinking tokens at critical steps serve as critical structural elements in LLM reasoning pathways, we propose Guided-RLVR, a novel approach that augments RLVR with a lightweight controller. The controller strategically injects semantically meaningful steering tokens into the sampling distribution, encouraging the model to approach problems from multiple reasoning perspectives and thereby expanding the exploration space. Experiments on mathematical reasoning and code generation benchmarks show that Guided-RLVR consistently outperforms existing RLVR methods while producing more diverse reasoning trajectories.
Abstract (translated)
URL
https://arxiv.org/abs/2610.37041
PDF
-
Omni-IGN: Onboard Image-Goal Navigation for Drones with Omnidirectional Vision
2026-10-27 10:24:19Zichen Yan, Rui Huang, Shenao Wang, Yucong Wang, Nuthasith Gerdpratoom, Lin ZhaoAbstract
Image-goal navigation (ImageNav) requires an agent to reach a location specified by a goal image in an unknown environment. For drones, this task demands exploration, goal- directed navigation, and collision avoidance under limited onboard computing resources. Existing aerial ImageNav methods are built on single-view observations, requiring additional rotation and exploration to search for goal-relevant cues outside the field of view. In this paper, we present Omni-IGN, a fully onboard ImageNav system that integrates quad-fisheye perception, an attention-based reinforcement learning (RL) policy, and a collision-aware safety layer. The policy adaptively fuses goal-relevant features across multiple views to provide directional guidance for continuous navigation. The quad-fisheye RGB images support both ImageNav and obstacle avoidance without a dedicated depth sensor. On Habitat benchmarks, Omni-IGN achieves the best performance among continuous-action baselines. Moreover, flight tests in IsaacSim validate the effectiveness of integrating learning-based depth estimation with the action safety layer for collision avoidance. Finally, real-world experiments demonstrate zero-shot sim-to-real transfer without any fine-tuning or external localization. The code and model checkpoints will be released upon acceptance.
Abstract (translated)
URL
https://arxiv.org/abs/2610.36968
PDF
-
PailitaoMMSearch2.0 Technical Report: Scaling Generative Retrieval to Unified End-to-End Multimodal Search
2026-10-27 10:17:33Xiaohan Ye, Zihan Gong, Jingqian Zhao, Lianyu Du, Yunmeng Shu, Xu Chen, Zesheng Zhou, Renchun You, Bocheng Chen, Jinsong Lan, Haihong Tang, Bo ZhengAbstract
Generative retrieval models have been increasingly adopted in search and recommendation systems. However, industrial generative retrieval models are commonly specialized to a fixed online scenario: they learn from scenario-specific interactions, support a narrow set of query forms and retrieval objectives. We study how to remove these couplings and build a unified end-to-end multimodal search system for open-ended image--text queries, heterogeneous shopping intents, and online scenarios with different resource constraints. We present \textbf{PailitaoMMSearch}, which directly generates product Semantic IDs from a raw query image and free-form instruction and scales along three dimensions. \textbf{Data scaling} uses a reusable toolset that can be composed into intent- and lifecycle-specific workflows for cold-start synthesis, quality control, real-traffic refinement, and continuous data production. \textbf{Intent scaling} progressively equips one model with e-commerce product knowledge, multiple retrieval relations, and adjustable reasoning budgets through continual pre-training (CPT), mixed-budget supervised fine-tuning (SFT), and unified reinforcement learning (RL). The RL stage combines intent-balanced optimization and Sample--Beam rollouts aligned with multi-candidate serving. \textbf{Scenario scaling} exposes quality--cost operating points across model capacity and reasoning budget. The system is deployed in multimodal search applications on Taobao Pailitao. A controlled online A/B test reports statistically significant improvements, including item page views $\textbf{+17.90\%}$, transactions $\textbf{+27.04\%}$, and gross merchandise value $\textbf{+8.94\%}$.
Abstract (translated)
URL
https://arxiv.org/abs/2610.36958
PDF
-
Vision-Language Agents Can Learn in Their Own Latent Worlds
2026-10-27 09:51:08Haoxi Li, Sikai Bai, Jianfei Ma, Qinglin Hou, Xiaocheng Lu, Tao Han, Jie Zhang, Song GuoAbstract
Learning from experience is essential for building the next generation of agents that continually expand their capabilities. Pretrained vision-language models (VLMs) already internalize rich world knowledge, yet this knowledge is rarely leveraged to drive policy improvement through online interaction. Realizing this potential efficiently remains challenging, as policy updates shift the agent's interaction distribution, requiring the world model to adapt accordingly, while continually updating large pretrained models is computationally expensive. To address this challenge, we introduce Dwell, a model-based reinforcement learning (RL) framework that reframes a VLM as a recurrent latent transition operator, endowing it with the ability to predict, adjust, and act in latent space. Leveraging the VLM's own representations and its dual role as both world model and policy, the agent continually learns from real interactions to anticipate and evaluate action outcomes, and from latent imagination to improve its behavior. Across five interactive visual environments, Dwell achieves the strongest overall performance among trainable methods while substantially reducing the compute required for imagination, showing that effective world models for VLM agents need not reconstruct the future; they need to predict futures that are useful for action. Our code is publicly available at this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2610.36925
PDF
-
Noise Steering: Exploiting Wake Steering and Acoustic Directivity for Noise-Constrained Wind Farm Control via Reinforcement Learning
2026-10-27 09:23:45Mart\'in de Frutos, Baptiste Corban, Claire Bizon Monroc, Ana Bu\v{s}i\'c, Donatien Dubuc, Jiamin Zhud, Esteban FerrerAbstract
Existing noise mitigation strategies for wind farms usually consist in reducing turbines power output, without accounting for the wake interactions between them and the noise emission directivity. This leads to over-conservative policies and unexploited wind power. Yet, turbine yaw can simultaneously affect wake interactions and the direction of noise emission, offering an additional degree of freedom for farm-level noise control. We therefore propose Noise Steering, a model-free reinforcement learning (RL) strategy that jointly controls turbine yaw angles and power setpoints to maximise farm power while satisfying a local noise constraint in a defined region. To enable training at a reduced computational cost, a low-fidelity acoustic model is coupled with the steady-state FLORIS simulator. The RL agent effectively learns how to redirect wakes and noise directivity away from the protected region while redistributing power across the farm. On a six-turbine test farm, the policy satisfies the 40 dB noise constraint throughout the full range of turbulent winds, with an average power deficit of only 2% relative to greedy unimpeded control, and up to 7.7% at high wind speeds. In comparison, the industry-standard noise-reduced operation (NRO) baseline here generates an average power loss of 17%. When deployed in the higher-fidelity this http URL simulator, the policy largely retains its performance under steady and turbulent inflow, despite differences in simulation fidelity. These results demonstrate the potential of combining yaw control and power derating through RL to improve the power-noise trade-off in wind farm operation.
Abstract (translated)
URL
https://arxiv.org/abs/2610.36888
PDF
-
When All Rollouts Fail: Policy-Compatible Joint Decoding for Hard-Problem Reasoning
2026-10-27 08:47:35Zehao Wang, Hanzhu Chen, Runquan Gui, Zhe Wang, Lin YangAbstract
Reinforcement Learning with Verifiable Rewards (RLVR) derives its learning signal from binary rewards via on-policy exploration and becomes a powerful paradigm for improving the reasoning of large language models. Yet it faces a learning cliff on problems beyond model's current ability: when every rollout fails, the samples reveal nothing about a correct solution, and group-relative estimators (e.g., GRPO, RLOO) collapse the update to zero advantage. Expert trajectories offer a remedy, but they are incompatible with the current policy and often lie in the student's low-probability regions, so the surviving updates demand large parameter shifts and decay into surface-level imitation rather than transferable reasoning. We argue that a useful learning signal on hard problems must be both verifiably correct and sufficiently compatible with the current policy. We propose Policy-Compatible Reinforcement Learning (PICO), which performs distribution-constrained joint decoding: expert data guides decoding at the token level while preserving the student's relative token probabilities, and a dynamic top-k mechanism confines the sampling distribution under a finite divergence budget, yielding trajectories that are verified and policy-compatible by construction. These trajectories are optimized jointly with on-policy rollouts under a mixed-behavior objective that mitigates their distribution mismatch via importance sampling. Experiments across reasoning diverse datasets and agentic benchmarks demonstrate that PICO consistently outperforms standard RLVR and prior off-policy baselines.
Abstract (translated)
URL
https://arxiv.org/abs/2610.36849
PDF
-
MVRCap: Towards Comprehensive Video Captioning via Hierarchical Multi-Dimensional Verifiable Rewards
2026-10-27 08:37:07Weihuang Lin, Xianing Chen, Zhibin Wang, Congzhi Zhang, Yiwei Ma, Jiayi Ji, Miao Chen, Jun Song, YuCheng, Bo Zheng, Xiaoshuai Sun, Rongrong JiAbstract
Despite significant progress, current models struggle to generate comprehensive video captions, constrained primarily by coarse-grained supervision, which often fails to capture spatial-temporal dynamics and the underlying logical intent. To address this issue, we introduce MVRCap, a novel framework that reformulates video captioning as a reinforcement learning task steered by structured verification. Central to our approach is a Fine-Grained Verifiable Reward mechanism, serving as dense supervision to enforce rigorous visual-semantic alignment. Specifically, we construct this reward system anchored by three complementary pillars: a Hierarchical Concept Coverage Reward utilizing our newly constructed MVRCap-QA dataset to ensure multi-dimensional completeness and semantic depth, a Factuality Verification Reward to enhance factual consistency, and a Conciseness Constraint to maximize semantic density. Furthermore, we curate the high-quality MVRCap-SFT dataset, providing structured annotations across six key dimensions, to serve as a robust foundational initialization for the RL policy in this training pipeline. Extensive experiments demonstrate that MVRCap achieves state-of-the-art performance on authoritative benchmarks including VDC, CAPability, and VC4VG.
Abstract (translated)
URL
https://arxiv.org/abs/2610.36838
PDF
-
SetGRPO: Calibrating Grader-Induced Policy Optimization
2026-10-27 08:11:30Jiale Huang, Zixu Li, Zhiwei Chen, Zhiheng Fu, Jinhe Bi, Yupeng HuAbstract
Rubric-based reinforcement learning uses fine-grained grader feedback to train large language models on open-ended tasks, but grader randomness can alter GRPO updates in ways that score discrepancies alone cannot characterize. Because GRPO jointly normalizes rewards within each response group and combines the resulting training coefficients with response-level gradients, a scoring discrepancy can shift the coefficients of the entire group and induce policy-dependent update changes. Selective execution further affects the direction-reversal risk of retained updates. To address this, we propose Set-Valued Group Relative Policy Optimization (SetGRPO). It comprises three modules: (a) Joint Reward Set Calibration constructs a joint uncertainty set for the training coefficients; (b) Geometry-Guided Safe Recovery maps candidate coefficients into the current policy's local update geometry and determines the largest safe recovery ratio for each response group; and (c) Selective Risk-Gated Optimization selects a global threshold through independent risk calibration to control the conditional direction-reversal risk of accepted updates while preserving the original GRPO objective. Under exchangeability, SetGRPO provides finite-sample marginal coverage for the joint coefficient set and simultaneous risk control across candidate thresholds. On ten open-ended benchmarks spanning three domains, SetGRPO achieves the highest macro average among all compared training methods on both Qwen3.5 and Llama3.2 model families.
Abstract (translated)
URL
https://arxiv.org/abs/2610.36807
PDF
-
Learning Side-by-Side Navigation for Personal Mobility Vehicles from Human Drivers
2026-10-27 06:45:45Luis Yoichi Morales, Florent FerreriAbstract
This paper presents an approach to achieve side-by-side autonomous motion with a pair of personal mobility vehicles (PMVs). Given a global path, two PMVs plan cooperatively so as to hold a social formation that lets the passengers hold a conversation, while keeping the ride safe and smooth. Each vehicle runs its own sampling-based model-predictive planner; the planners exchange nothing but the odometry each vehicle could observe of the other, and the joint decision is expressed as a pair of utility matrices over pairs of candidate trajectories, resolved by a common rule. The behavioural parameters are fitted by inverse reinforcement learning from 409 minutes of driving by ten pairs of people in human-driven PMVs in an indoor environment, and the lanes the pair can use are derived offline from the corridor itself rather than fixed by hand. A safety layer the cost cannot override executes a move only if the vehicle can still brake clear, measured from its whole body, and a vehicle that is stuck reverses along its own path and continues. In closed-loop simulation over 24 runs in the conditions the drivers were recorded in and 24 in four generated obstacle layouts, no vehicle entered a wall or a cone. The pair holds conversational formation for 60.5 and 68.9 percent of the time among obstacles (anticlockwise and clockwise) and 81.2 and 91.5 percent in a clear aisle, against the drivers' 34.0 and 36.0 percent, and 52.9 and 56.0 percent, and it matches their speed. It keeps one lateral formation where the drivers alternated between a close diagonal and a wide abreast one, and it varies far less. Experimental results are based on a differential drive kinematic simulator running on ROS 2.
Abstract (translated)
URL
https://arxiv.org/abs/2610.36726
PDF
-
InfraDev: Can Coding Agents Develop AI Infrastructure, and How Can We Tell?
2026-10-27 03:47:47Yangshen Deng, Jiaqi Ruan, Yeqi Huang, Nick Lau, Guangming Sheng, Luo MaiAbstract
Coding agents are beginning to develop the production-grade AI infrastructure that trains and serves large models, yet existing benchmarks cover only single kernels or single-GPU tasks, and grade with tests and prompts that misjudge both correct and incorrect patches. We introduce InfraDev, a benchmark of 36 tasks built from real-world pull requests to Megatron-LM, vLLM, and verl, covering infrastructure development for training, inference, and reinforcement learning. Each task asks an agent to reimplement a real change that spans many files and must run correctly across many GPUs. To grade such changes, we jointly synthesize hidden probes, which enforce the requirements the original tests leave unchecked, and a prompt that states every interface the probes call without revealing the solution. We also design a shared execution environment that reuses dependencies and artifacts across tasks, cutting storage by 52-57% and build time by 13-14% on Megatron-LM and vLLM. Across six frontier models, the best resolves only 38.9% of the tasks; the probes reject 32% of the submissions that pass every official test, and with the original pull request descriptions as prompts, no model resolves a single task. Failure analysis shows that structural and state errors dominate and that agents' own tests rarely check memory, time, or traffic. Turning these findings into an agent skill raises the resolved rate of a frontier model on the Megatron-LM tasks by 10.5%.
Abstract (translated)
URL
https://arxiv.org/abs/2610.36589
PDF
-
AudioRSI: Recursively Self-Improving Audio Reasoners
2026-10-27 03:33:35Jiajun Fan, Roger Ren, Jingyuan Li, Rahul Pandey, Prashanth Gurunath Shivakumar, Ankur Gandhe, Yile Gu, Ge Liu, Ivan BulykoAbstract
Reinforcement learning has made Audio Large Language Models (Audio LLMs) better reasoners, but the reward that drives it still comes from outside the model: a final-answer check, a hand-crafted process score, or an external LLM judge that stays fixed while the policy improves. We show that an Audio LLM can supply this reward itself, and can do so recursively: each improved model becomes the critic that trains its successor, so the training signal improves along with the model. We introduce AudioRSI (Recursively Self-Improving Audio Reasoners), a reinforcement learning framework in which a single model plays two roles: as an actor it generates reasoning paths, and as a critic it compares them pairwise, turning its own judgments into a process-level reward. The critic is periodically replaced by the current policy, which closes the recursive loop. Because the critic hears the same recording as the actor, it can reject reasoning that cites acoustic evidence the audio does not contain, rather than reward reasoning that merely sounds plausible. The recursion is what sustains progress: under a frozen external judge, out-of-distribution accuracy plateaus during training, while a critic refreshed from the improving policy keeps it rising. Across four out-of-distribution benchmarks, all disjoint from the AVQA training distribution, AudioRSI outperforms strong baselines including Audio Flamingo 3 and Gemini 2.5 Pro on MMAU test-mini, ranks first on MMSU, achieves a clear lead on the challenging Multi-Audio split of MMAU-Pro, and leads the 7B open-source tier on MMAR among methods without external reward models. With no reward specifying how to reason, AudioRSI also develops structured reasoning behaviors (logical reasoning, perception-aware analysis, and reduced hallucination), corroborated by an external LLM-as-Judge evaluation and a keyword-based behavior analysis.
Abstract (translated)
URL
https://arxiv.org/abs/2610.36578
PDF
-
Confident First, Reflective Second: Reinforcement Learning of Confidence-Preserving Reflection for Video Understanding
2026-10-27 01:23:41Zihao Lin, Jingyuan Liu, Yucheng Zhang, Yuguang Yao, Lifu HuangAbstract
Video models answer from sampled frames and may miss the decisive moment. A common remedy is to commit to an answer, re-examine the video, and confirm or revise it, a loop now mainly learned with reinforcement learning (RL). We find that RL erodes the initial answer, a phenomenon we term confidence erosion, where confidence means the label-verified reliability of the initial answer instead of self-reported certainty: because errors can be repaired later, a correct initial answer earns less reward, and the initial answer stalls even as the final answer's gain over it rises. We propose the principle confident first, reflective second: the initial answer must stay reliable on its own, above a floor set by a reference policy, and reflection may only add to it. Our method, CoFire, implements this principle with one mechanism for each erosion channel: reward incentives, parameter drift, and gradient interference. CoFire reaches 67.0% mean accuracy on Video-MME, MLVU, and LongVideoBench, 1.7 points above the strongest compared 7B method, and improves initial-answer and final accuracy over outcome-reward RL by 4.1 and 3.7 points on average: a stronger initial answer carries through to a stronger final one.
Abstract (translated)
URL
https://arxiv.org/abs/2610.36467
PDF
-
CONCUR: Concordant Auxiliary Rewards for Multi-Task, Multi-Reward Reinforcement Learning
2026-10-27 00:36:59Jixuan Leng, Si Si, Fei Wang, Cho-Jui Hsieh, Jiong Zhang, Hsiang-Fu Yu, Inderjit S. DhillonAbstract
When a Large Language Model (LLM) is trained with Reinforcement Learning (RL) on multiple tasks, a task's primary reward can gate auxiliary rewards such as a length reward: on a failing response, an auxiliary cannot exceed the primary. However, group-relative advantage estimators treat a gated auxiliary like any other component, so it can reverse the primary's order, and at the primary's ceiling only auxiliaries order responses. Keeping gated auxiliaries hurts some tasks, while dropping them hurts others. We propose CONCUR, an advantage estimator that replaces the standardized scores of each gated auxiliary with the closest scores that satisfy two conditions: responses with higher primary rewards cannot receive lower scores, and responses at the primary's ceiling receive equal scores. CONCUR-T applies both conditions throughout training, and CONCUR-A lifts the second for an auxiliary that agrees with the primary's order more often than it reverses it among passing responses. On three models from two families and across math, code, instruction following, and writing, both variants score higher on average than all baselines and than when all the gated auxiliaries are dropped.
Abstract (translated)
URL
https://arxiv.org/abs/2610.36437
PDF
