Reinforcement_Learning
-
Learning to Reason with Selective Internal Feedback
2026-10-15 07:08:46Xun Lu, Shumeng Yang, Xinrui Wang, Songjun Tu, Linjing Li, Daniel Dajun ZengAbstract
Reinforcement learning from internal feedback can improve language models without ground-truth answers or external verifiers. However, confidence and correctness can diverge during training, and continued optimization of average self-certainty can increase confidence while degrading reasoning accuracy. We introduce SIF, a reinforcement learning method that combines comparative reward construction with selective policy updates. SIF compares ranked token-level self-certainty profiles across multiple windows and peers, assigning positive rewards only when both levels of voting provide sufficient support. Sample and group level filtering determines which responses contribute to policy optimization. Experiments on three backbones across 12 mathematical and scientific reasoning benchmarks show that SIF's final checkpoints achieve higher average accuracy than the best-validation checkpoints of Intuitor and Vigor, with a gain of 4.71 percentage points over Intuitor on Qwen2.5-Math-1.5B. Further analyses show better retention of accuracy during continued optimization, while test-time training experiments demonstrate sustained improvements with as few as four rollouts per prompt.
Abstract (translated)
URL
https://arxiv.org/abs/2610.21392
PDF
-
Perception or Reasoning? Diagnosing and Tackling Reinforcement Learning Bottlenecks in VLMs
2026-10-15 07:00:46Zeyuan Yang, Zhuohao Chen, Jiaben Chen, Yao Zhang, Maohao Shen, Chuang GanAbstract
Reinforcement learning has brought strong gains to language model reasoning, but its improvements on visual tasks remain less consistent. One difficulty is that visual task failures can arise from different sources: the same benchmark may be limited by perception for one model and by reasoning for another. We introduce Perception and Reasoning Attribution (PARA), which measures the remaining perceptual and reasoning headroom of each model--benchmark pair by comparing how many current errors are recovered with more accurate visual information versus a stronger reasoner. Across 4 vision-language models and 17 multimodal benchmarks, PARA broadly agrees with the expected characteristics of existing benchmarks while revealing substantial model-specific variation. More importantly, it shows that RL gains become smaller as perceptual headroom increases, but larger when reasoning explains more of the recoverable errors, suggesting that current RL exploits reasoning headroom more effectively than perceptual headroom. Motivated by this gap, we analyze how existing RL objectives affect visual perception and find that reward- and token-level signals influence visual attention only indirectly. We therefore propose GRAD, a group-relative attention distillation objective that uses the best rollout in each group to directly guide the visual attention of weaker rollouts. GRAD consistently improves performance across two model scales and seven benchmarks, and further complements existing RL methods.
Abstract (translated)
URL
https://arxiv.org/abs/2610.21380
PDF
-
Hindsight-Privileged RLVR: Single-Rollout Token Credit Without Group Sampling
2026-10-15 06:36:16Wei Chen, Yubing Wu, Zhou Xu, Haoyue Zhang, Delu Zeng, Jiahao Wang, Junmei Yang, Pan tiewen, Jingfeng Liu, Zhiheng Zhou, Bo Li, Min ChenAbstract
Reinforcement learning with verifiable rewards (RLVR) trains language models from checked task outcomes, but assigning a terminal reward to individual reasoning tokens or actions remains difficult. Group-relative methods such as Group Relative Policy Optimization (GRPO) sample several attempts per prompt, but lose their reward-based learning signal when all attempts receive the same outcome. We propose Hindsight-Privileged RLVR (HP-RLVR), a single-rollout method that turns a completed trajectory's own future into token-level update weights. The generating policy snapshot scores each eligible token under its causal history and under a hindsight context containing the realized future and verifier feedback. A normalized Jensen--Shannon divergence between the two views weights the reward relative to a historical baseline, yielding bounded token credit for both successful and failed attempts, with no reference solution or learned critic. In paired Qwen3-8B runs over a shared $100$-step schedule, HP-RLVR trains $1.21\times$, $1.46\times$, and $2.30\times$ faster than GRPO on mathematics, code generation, and ALFWorld, using one eighth the rollouts per prompt. Accuracy is comparable overall, with HP-RLVR ahead on mathematics at all three sizes and GRPO ahead on code and ALFWorld. In ALFWorld, $43.2\%$ (4B) and $49.0\%$ (8B) of GRPO's gradient-norm records are zero, versus none for HP-RLVR.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20497
PDF
-
Ranking and Partitioning of Attention Heads for Sparse and Continual Reinforcement Learning Fine-tuning
2026-10-15 06:21:56Ruizhe Shi, Zaiwen Yang, Qiubo Wang, Banghua Zhu, Maryam FazelAbstract
We study sparse and continual reinforcement learning fine-tuning of language models through attention-head ranking and partitioning. For sparse learning, we rank attention heads by gradient norm to identify a small subset of heads that can be trained effectively. For continual learning, we partition heads across tasks to reduce cross-task interference and preserve previously learned capabilities. Theoretically, we relate gradient norm to training effectiveness, and further characterize why sparse, separated head updates can remain stable. Empirically, across math, instruction-following, and coding tasks, tuning higher-ranked heads consistently outperforms tuning lower-ranked ones while often matching denser baselines, and head partitioning improves retention while preserving adaptation to new tasks.
Abstract (translated)
URL
https://arxiv.org/abs/2610.21336
PDF
-
Constrained Multi-Agent Reinforcement Learning in Imagination
2026-10-15 06:09:04Raz Shmueli, Shahaf S. ShperbergAbstract
Predicting cost in a multi-agent world model moves most constrained policy updates into imagination: the stronger model-free baseline needs $10$ to $20$ times CA-MAMBA's interactions on the median SMAC configuration, and CA-MAMBA earns $2.8$ to $12$ times the baselines' MAMuJoCo reward at equal budget. In controlled tests on one task per benchmark, pricing the model's cost advantages at parity more than halves a cost-blind learner's cost and meets a tight continuous-control limit the learned multiplier misses, suggesting the model's cost estimates are not the bottleneck.
Abstract (translated)
URL
https://arxiv.org/abs/2610.21319
PDF
-
What Should We Remove? Evaluating the Reliability and Costs of RLVR Data Decontamination
Abstract
Reinforcement learning with verifiable rewards (RLVR) math corpora are routinely screened for benchmark overlap, but the reliability and data-loss costs of cleaning policies remain underexamined. We audit sixteen open RLVR math corpora comprising 1,673,514 corpus-specific training items against six benchmarks containing 1,949 questions. We evaluate eight policies based on exact, lexical, and semantic matching using 594 independently and blindly dual-reviewed training-benchmark pairs. Calibration and held-out sets have disjoint benchmark questions and training-example families. A separate review of all 83 candidate matches for 80 sampled training items measures item-level false removal. At cosine threshold 0.80, 52 of 59 sampled removals (88%) were confirmed false removals; this is not a demonstrated population rate. No semantic threshold in our preregistered grid (0.80-0.98) meets the 90% conservative-precision target on calibration. The existing judge-adjudicated policy achieves 98.1% resolved precision and 98.2% candidate-frame recall on held-out corpora. It has zero confirmed false removals among six sampled removals, with one unresolved case; this does not establish zero risk. Within the evaluated candidate frame, this policy offers a more favorable precision-recall balance than the semantic-threshold alternatives tested. Recall does not measure contamination outside the candidate frame. We release reviewed labels, per-policy decisions, and coverage accounting to support evaluation of cleaning reliability and cost rather than overlap counts alone.
Abstract (translated)
URL
https://arxiv.org/abs/2610.21229
PDF
-
Rubric Judgments Lack Uncertainty: Conformal Weighting for LLM Evaluation
2026-10-15 04:06:12Licheng Pan, Haocheng Yang, Haoxuan Li, Yuzhou Chen, Yichen Sun, Yuan Lu, Xiaoxi Li, Hao Wang, Zhixuan Chu, Kui RenAbstract
Rubric-based evaluation is widely used to assess large language model (LLM) outputs by decomposing quality into multiple criteria and assigning each a binary judgment from an LLM verifier. However, judgment certainty varies substantially across criteria. Existing methods uniformly aggregate these judgments into a final score, implicitly treating every criterion as equally certain and overlooking this heterogeneity. As a result, uncertain judgments remain incorporated into reward estimation without distinction, potentially yielding misleading training signals for downstream reinforcement learning (RL). To address this limitation, we propose Uncertainty-aware Rubric Weighting (URW), a training-free framework that quantifies criterion-level judgment uncertainty via conformal calibration. Specifically, URW elicits a soft uncertainty signal from a fixed pool of paraphrased judgment templates and converts it into a per-criterion weight using split-conformal $p$-values with provable coverage guarantees. Extensive experiments demonstrate that URW consistently improves pairwise preference accuracy and delivers stronger downstream RL alignment.
Abstract (translated)
URL
https://arxiv.org/abs/2610.21214
PDF
-
Judging the Judges: Confidence and Disagreement in Rubrics as Rewards
2026-10-15 03:35:00Michael Tian-Yue Liu, Anne Harrington, Trevor Darrell, Yutong BaiAbstract
Rubrics as Rewards (RaR) extends reinforcement learning to non-verifiable domains by using an LLM judge to score responses against a structured rubric. Prior work has focused on improving the rubric or strengthening the single judge, with comparatively little attention to how different single or multiple judges might be used during RaR training. We first show that judge identity alone produces a substantial spread in scores when training Qwen3-4B-Base on RaR-Medicine and RaR-Science across five frontier and open-weight judges. Building on this, we propose RaRE (Rubrics as Rewards with Ensembles), a novel judge ensemble strategy for RaR with two mechanisms: a soft-Likert reward that uses judge logprobs to weight more confident judges more heavily, and a drop-on-disagreement mask that excludes high-disagreement rollouts from the GRPO update. We find that this design is critical: while a naive average of judge scores fails to outperform the best single judge, RaRE outperforms every single-judge RaR variant on both HealthBench and GPQA Diamond, with per-theme analysis showing that it inherits the complementary strengths of its constituent judges. Our findings establish judge ensembling as a compelling axis for improving RaR.
Abstract (translated)
URL
https://arxiv.org/abs/2610.21179
PDF
-
All-in-One Memory: Process-Rewarded Reinforcement Learning for Memory Agents
2026-10-15 03:02:24Bingbing Wang, Jingjie Lin, Weiming Qiao, Jing Li, Ruifeng XuAbstract
Large language models increasingly rely on external memory to answer queries grounded in long-horizon histories such as multi-session dialogues and lengthy documents. However, existing memory methods often optimize retrieval and memory construction separately, without a shared reward space for evaluating search, writing, and stopping decisions. To address this, we propose All-in-One Memory (AiO-Mem), which organizes these operations as transitions over a shared evidence state and directly evaluates the semantic quality of each action in a unified reward space. AiO-Mem compiles heterogeneous histories into source-traceable multi-view events, providing a shared substrate for retrieval and memory writing. A unified memory policy actor then rolls out complete interaction trajectories over this shared evidence state. A separate trajectory-conditioned process reward model is trained through evidence-grounded supervised fine-tuning and trajectory-level DPO to assign process scores to the collected trajectories. These scores are aggregated into an action-balanced trajectory reward and used by Process-Aligned DAPO (PA-DAPO) to optimize the memory policy. Across long-context memory benchmarks, AiO-Mem consistently improves answer accuracy over both static memory systems and RL-trained memory agents.
Abstract (translated)
URL
https://arxiv.org/abs/2610.21147
PDF
-
Manifold Generative Annealing for Contact Constraint-aware Impedance Control with Reinforcement Learning Prior
2026-10-15 02:51:51Zhilin He, Guoquan Qi, Changliu LiuAbstract
Contact-rich robotic tasks require coordinated motion and impedance under uncertain interaction dynamics and physical constraints. While recent methods on learned policies provide useful interaction strategies, they may fail when contact conditions differ from those encountered during training. Model-based generative methods show promising receding-horizon trajectories optimization results, but often lack structured behavioral guidance and refinement that jointly accounts for constraint geometry and actual contact responses. To this end, we propose Manifold Generative Annealing (MGA), a joint model-free and model-based generative framework for motion--impedance optimization. MGA first organizes pretrained reinforcement learning policies into structured sequence proposals, retaining task-specific interaction strategies without committing to their execution. It then evaluates these proposals alongside exploratory samples through model-based rollouts for refinement or separate execution selection. Crucially, realization-aware manifold refinement aligns updates from the refinement bank with both task constraints and locally achievable physical responses, allowing learned experience to guide online adaptation under changing contact conditions. Experiments cover surface scanning, peg insertion, and humanoid pushing, with improved safety success rate of impedance control under nominal and sensing-shift conditions.
Abstract (translated)
URL
https://arxiv.org/abs/2610.21140
PDF
-
MemComposer: Learning to Assemble Memories for LLM Agents via Reinforcement Learning
2026-10-15 01:47:06Yixuan Xie, Xueqiang Xu, Xinyi Fan, Xinxi Lyu, Tao Feng, Xing Fan, Chenlei Guo, Jie Hao, Shaowen Wang, Jiashuo Sun, Jiawei HanAbstract
Agent memory enables large language model (LLM) agents to reuse past experiences without updating their parameters. However, semantic similarity alone does not guarantee memory usefulness: downstream performance depends on how memories are selected and presented to the agent. We propose MemComposer, a memory composer that jointly examines the current query and retrieved candidates to decide which experiences to include, how many to select, and how to order them. We train the composer using Experience Policy Optimization (ExPO), a GRPO-based framework that evaluates assembled memory contexts through complete agent interactions and uses task-success feedback to optimize the assembly policy. Across three $\tau^2$-bench domains with two frozen executors, MemComposer outperforms the strongest baseline for each executor while reducing average memory-token usage by 16.9% and 4.9%, respectively. The learned policy also transfers to unseen domains, executors, and memory banks, and benefits from memories generated by stronger models without further training. These results demonstrate the value of learning query-specific memory assembly using downstream task feedback.
Abstract (translated)
URL
https://arxiv.org/abs/2610.21091
PDF
-
Recursive Reward Improvement via Adversarial Meta-Prompt Evolution
2026-10-15 00:31:18Wei Dai, Qianchu Liu, Guanghui Qin, Xing Han, Tristan Naumann, Paul Pu Liang, Sheng ZhangAbstract
Recursive Self-Improvement (RSI) has gained wide attention for its ability to work with limited data. In a proposer-solver loop, a proposer model writes training tasks and a solver model learns to solve them through reinforcement learning. Such loops have succeeded in mathematics and coding, where every answer can be checked automatically. Open-ended expert tasks, like clinical report generations, have no answer key, which makes simple question-answer generation not viable. Here the proposer must also write a rubric that lists the content an answer should include and the errors it should avoid. However, the unique challenge behind model-written rubric is the comprehensiveness of it. For naively generated rubric, solvers tend to exploit it, naming the expected topics without the substance behind them, which corrupts the learning process. To solve this problem, we introduce RRI, Recursive Reward Improvement, which uses a reward hacking adversary to supervise the reward. Before the solver trains on a generated task, an adversary writes an answer that games its rubric while being as useless as possible, the same model writes an honest answer without seeing the rubric, and up to two repair steps revise the rubric to make the honest answer win. The exploits found also rewrite how future tasks and rubrics are generated. On HealthBench Professional, RRI improves the adjusted score over the Generation Only self-improvement baseline by 10.9 points at 9B and 12.4 at 27B, and by 10.3 points without any external model. In particular, the gains are largest on difficult and writing tasks. We show that the 27B model trained on the HealthBench Professional description improves by 6.6 points on PRBench Hard and 8.0 on ProfBench without seeing a task from either. We further analyze the exploitable points of model-generated question-rubric pairs to facilitate future research on open-ended self-play.
Abstract (translated)
URL
https://arxiv.org/abs/2610.21040
PDF
-
Learning at the Right Pace: Adaptive Data Scheduling Improves LLM Reinforcement Learning
2026-10-14 23:58:09Zicheng Xu, Ruixuan Zhang, Yu-Neng Chuang, Xiuyi Lou, Hoang Anh Duy Le, Oren Gal, Alexander S. Szalay, Zhaozhuo Xu, Guanchu Wang, Vladimir BravermanAbstract
Large Language Models (LLMs) achieve remarkable reasoning capabilities through reinforcement learning (RL) post-training. However, existing RL post-training commonly relies on uniform data sampling, which ignores the semantic structure of the training data and the evolving capability of the training policy. To address these limitations, we propose Adaptive Data Scheduling (ADS), a dual-level data scheduling framework that combines adaptive sampling across semantic clusters with policy-boundary sample selection. At the cluster level, ADS organizes samples according to semantic patterns and adapts the inter-cluster sampling distribution to prioritize clusters within the policy's current reach. At the sample level, ADS tracks prompts near the policy's capability boundary within each cluster to obtain informative relative advantages. Experimental results across three LLMs and seven reasoning benchmarks demonstrate that ADS improves average accuracy by 5.5% over Group Relative Policy Optimization (GRPO). Notably, ADS consistently improves RL methods with different objective designs, highlighting its potential as a general data scheduling strategy. The source code is available at: this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2606.22305
PDF
-
Deep Reinforcement Learning for HPC Job Scheduling: A Statistical Evaluation
2026-10-14 23:02:02Justin M. Cheney, Clement N. NyirendaAbstract
Deep Reinforcement Learning (DRL) offers learned job selection for High-Performance Computing (HPC), but performance against heuristics alone does not establish value beyond action masking. Six DRL treatments are compared: Proximal Policy Optimization (PPO), Advantage Actor-Critic (A2C), and Deep Q-Network (DQN), each with and without masking. Three heuristics and a seed-paired masked uniform random control provide baselines across ten seeds on two real Slurm (Simple Linux Utility for Resource Management) workload traces, one homogeneous and one heterogeneous, under a common non-parametric framework. All masked treatments complete evaluation, whereas unmasked treatments are truncated. On the homogeneous trace, no masked DRL treatment significantly improves average waiting time or average and maximum bounded slowdown over the control. On the heterogeneous trace, MaskablePPO improves all three metrics relative to the control. Its average waiting time is equivalent to that of the three heuristics within a 10% margin, and its maximum bounded slowdown is 22% lower than that of Last-Come, First-Served (LCFS). MaskablePPO has the leading average waiting-time rank among complete treatments on both traces, but Nemenyi comparisons detect no significant difference from MaskableA2C. Among masked variants, MaskablePPO has the lowest recorded mean training time, while MaskableDQN has the lowest mean decision latency. These findings distinguish completion from baseline-specific scheduling gains and support including a masked random control when assessing the value of learned selection.
Abstract (translated)
URL
https://arxiv.org/abs/2610.21008
PDF
-
How to Train Your Little Dragon: Teaching a Small Language Model to Construct and Refine GPU Optimization Scripts
2026-10-14 21:49:24Annabelle Sujun Tang, Wei Tang, Songyao Jiang, Tianhui Li, Shengzhe Xu, Jiin Woo, Kaan Ozkara, Youngsuk Park, Zhen Jia, Hadi EsmaeilzadehAbstract
Compiler transformations interact, making GPU kernel optimization a search for combinations that preserve the computation while improving performance. We introduce HOW TO TRAIN YOUR LITTLE DRAGON, a training method that teaches a small language model both to construct complete optimization scripts and to improve existing ones. Instead of ranking candidates or generating kernel code, our model generates optimization scripts. We devise a compact, lossless representation whose deterministic decoding reconstructs the exact compiler trace, preserving its transformation parameters and dependencies. We use a compiler framework TVM MetaSchedule to build training data on 54,116 distinct workloads, and construct 108,594 examples in two training sets: 60,236 teach the model to generate an optimized script from an unoptimized kernel, while 48,358 teach it to produce a better script given the kernel and a somewhat optimized script. Supervised fine-tuning teaches the model to construct and improve optimization scripts from these search-derived examples. We then add reinforcement learning with a shaped reward that combines correctness with measured performance, awarding a latency-based bonus only to numerically correct kernels. We apply this training method to a model with 8 billion parameters. After supervised fine-tuning and reinforcement learning, the model achieves per-attempt correctness of 91.27% for construction and 92.01% for refinement, while reinforcement learning produces faster kernels than supervised training on 88 of 102 construction tasks and 145 of 170 refinement prompts. With only 8B parameters, our LITTLE DRAGON achieves higher per-attempt correctness and better kernel performance than six frontier LLMs: GPT-6 Astra, Opus 5, Kimi-K3, and GLM-5.3 among them, each with hundreds of billions to trillions of parameters, for which task-specific training is unavailable or expensive.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20980
PDF
-
SOMA: Stack-of-Objects Manipulation using a Quadruped Robot
2026-10-14 21:43:14Stergios E. Bachoumas, Panagiotis ArtemiadisAbstract
Executing heavy-duty physical manipulation in unpredictable environments remains a major challenge for quadrupedal robots lacking dedicated manipulators. This paper presents SOMA, a hierarchical reinforcement learning framework for whole-body non-prehensile manipulation of multi-object stacks using a quadruped robot. Unlike traditional methods restricted to two-dimensional planar pushing, SOMA integrates explicit robot base-height control, unlocking full three-dimensional pushing strategies. This capability enables the quadruped to manipulate complex payload stacks as a unified system while preventing relative sliding or tipping of upper layers. The architecture factorizes control into a high-level manipulation policy using a Simplified Structured State Space Sequence (S5) encoder to reason over long-horizon temporal history and predict bounded target tracking commands via Beta distributions, and a low-level locomotion policy that dynamically adapts leg stance configurations. In extensive simulation benchmarks, SOMA achieves a $90.3\%$ success rate for single-object reposing and up to $90.1\%$ for dynamic two-box stacks under stringent error margins ($0.05\text{ m}$ position error, $1^\circ$ orientation error). Finally, we validate zero-shot sim-to-real transfer on a Unitree Go1 quadruped robot across diverse payload masses and geometries, demonstrating active base-height adjustments to maintain traction and balance during heavy pushing tasks.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20968
PDF
-
MAI-Gym: An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments
2026-10-14 21:06:01Panrong 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 HoiAbstract
Online reinforcement learning (RL) with verifiable rewards has driven rapid progress in mathematics, coding, and tool use, but scaling this success to GUI agents remains challenging due to the high cost of interactive environments and the scarcity of realistic training tasks. We introduce MAI-Gym, an open all-in-one platform that unifies GUI agent evaluation, data collection, and online RL training across mobile, computer, and web environments. Through a shared rollout pipeline, MAI-Gym ensures consistent agent behavior across evaluation and training, supports ten established benchmarks, and scales to 20,000 parallel environment instances. Beyond scalable infrastructure, MAI-Gym enables automated synthesis of diverse GUI applications and challenging tasks, including resettable mocked apps and coherent cross-app worlds with execution-validated task verifiers. Using MAI-Gym, we train MAI-Gym-9B, a unified 9B agent for mobile and computer use, achieving state-of-the-art performance among open GUI agents under 10B parameters: 78.4% on MobileWorld and 76.2% on OSWorld-Verified. Importantly, RL in a mocked train-ticketing app transfers to its real-world counterpart, improving task success on a phys- ical phone from 28.8% to 48.1%. We release the framework, synthesis skills, mocked apps, 2,000 verified tasks, and the model checkpoint to support repro- ducible research and further advances in scalable GUI agent training.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20939
PDF
-
Wrong Votes, Right Gradients: Why Majority-Vote Rewards Track Verifiable RL
2026-10-14 20:25:42Dylan P. Wilson, Chengsong Huang, Jiaxin HuangAbstract
Reinforcement learning with self-generated labels, such as majority vote over a model's own samples, often recovers most of the gain of training with verifiable rewards. This is puzzling from the usual prompt-level view, in which majority vote only returns a model's modal answer and cannot add information about correctness. We argue that the relevant unit is the batch-level update. Because parameters are shared, a prompt's own gradient accounts for less than 1\% of the change in that prompt's output at practical batch sizes, so a wrong label matters only to the extent that label errors share a direction across prompts. We decompose the majority-vote update into shrinkage and misdirection relative to the gold update, and show that misdirection vanishes with batch size whenever errors are incoherent across prompts, however often the labels are wrong. Measuring per-response gradients on five models, we find that majority-vote errors are as incoherent as a random-reward null, that the majority-vote update agrees with the gold update as closely as two independent gold updates agree with each other, and that its noise-corrected cosine with the gold update is 1 within error. Randomly flipping up to 30\% of gold labels leaves the update direction intact, while the same fraction flipped by a shared rule destroys it, and rewards that favor a kind of response on every prompt, such as a learned reward model, fall measurably short. The equivalence holds at every checkpoint we examine but is local: gold-reward training still ends ahead on held-out majority accuracy. Error coherence, not error rate, determines when self-generated labels can stand in for verification.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20898
PDF
-
Race to Recovery: Designing Advanced Safety Systems for Autonomous Driving via Learning Autonomous Racing
2026-10-14 20:04:54Teodor Ilie, Kleber Cabral, Sidney GivigiAbstract
Vehicles that lose control at the handling limit through icing, hydroplaning, or an abrupt evasive maneuver fall outside the operating envelope of production safety systems such as electronic stability control, which focus on preventing instability rather than recovering from it. Expert drivers regain authority in these conditions by drifting. Existing autonomous recovery approaches are validated on idealized scenarios, define "recovered" states empirically and inconsistently, or rely on offline-planned trajectories rather than acting online. We address these gaps by first training a racing policy with reinforcement learning that learns drifting as a time-optimal maneuver, then adapting it to recovery through continued training with a minimum-time reward. We define recovery relative to a phase-plane-grounded stable equilibrium and introduce Gym-Khana for controller training and evaluation. Across 294 initial loss-of-control states, the transferred policy improves recovery by 186.8% over a Stanley path tracker and outperforms both a nonlinear model predictive controller and a recovery-only policy trained from scratch, while also recovering faster on states the baseline can already handle. An ablation shows that racing pretraining, rather than broader initial-state exposure alone, drives this gain. Deployed on a physical 1:10-scale F1TENTH vehicle using system identification, actuator-lag modelling, and domain randomization, the transferred policy outperforms Stanley on every tested loss-of-control state, including recovering 50% of the time from a hard state the baseline never recovers from.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20873
PDF
-
Gibbs-Gap Policy Optimization for Off-Policy RL with Verifiable Rewards
2026-10-14 18:59:29Xuanchang Zhang, Tong ZhangAbstract
Reinforcement learning with verifiable rewards often trains language models on responses generated by a different, possibly stale behavior policy. Correcting this mismatch with learner-to-behavior importance weights can introduce high variance for long responses. We introduce Gibbs-Gap Policy Optimization (GGPO), which minimizes a log-mean-exp contrast of KL-adjusted rewards among responses to the same prompt, without learner-to-behavior importance weights. We derive GGPO from KL-regularized policy optimization: on policy, its population loss is exactly the KL-regularized optimality gap, and under suitable coverage the learner's single-response expected reward is at least the behavior policy's best-of-$K$ reward, minus terms for the raw finite-group loss, regularization, and finite training data. We characterize reference refresh through exact Gibbs-target composition and a finite-data final-policy reward bound under local reference control and tokenwise behavior mismatch. This refresh is empirically important: in an ablation, fixed-reference GGPO scores $32.96$ on the five-benchmark problem-weighted average, versus $46.58$ with the default refresh schedule. Experiments on mathematical reasoning further show that GGPO outperforms GRPO and AGRO under training-inference mismatch and on fixed offline data, while maintaining its performance as controlled policy staleness grows.
Abstract (translated)
URL
https://arxiv.org/abs/2610.20814
PDF
