Reinforcement_Learning
-
LongZero: Enhancing Long-Context Reasoning in LLMs from Zero Data
2026-10-21 17:59:14Zhennan Wan, Jianfei ChenAbstract
Reinforcement learning (RL) is an effective way to strengthen the long-context reasoning of large language models (LLMs), but it needs training data that are long, diverse, challenging, and verifiable. Human annotation of such data is expensive, naturally occurring data are too scarce to sustain RL, and existing synthesis methods depend on external corpora or stronger teacher models, which become harder to obtain as the trained models get stronger. Recently, RL from zero data, in which a model learns from tasks it proposes itself, has achieved excellent performance, but existing methods target short-context tasks. Therefore, we propose $\textbf{LongZero}$, a method that lets an LLM improve its own long-context reasoning without any external data. Instead of writing questions and answers over existing documents, LongZero first has the model synthesize the question and answer with its supporting evidence in a compact structured form, and only then separately expands the evidence into a long multi-document context. LongZero combines these self-written multi-hop reasoning tasks with self-written multi-round retrieval tasks and trains the model with GRPO. We evaluate LongZero on two LLMs across five long-context reasoning benchmarks and compare it with RL on data synthesized from external corpora. Experimental results show that LongZero can generalize to real tasks, and outperforms both baselines, suggesting a viable and promising path to improving long-context reasoning of LLMs.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30132
PDF
-
Asymmetric On-Policy Distillation Learns In-Context Exploration
2026-10-21 17:50:31Sriyash Poddar, Yanda Bao, Jacob Krantz, Matthew Chang, Xavier Puig, Roozbeh Mottaghi, Natasha Jaques, Abhishek GuptaAbstract
Sequential decision-making under partial observability is difficult because agents must explore to infer hidden, task-relevant information before acting effectively. Learning such information-gathering behaviors tabula rasa via online reinforcement learning is sample-inefficient, whereas offline imitation averages expert actions across hidden states and fails to learn exploration. In this work, we propose ASTEROID, a simple and effective student--teacher supervised learning framework that can learn test-time exploration. Although a privileged expert never needs to explore, we demonstrate that a history-conditioned student can learn to do so when trained with on-policy contexts: the student first collects on-policy context under partial observability, and the privileged expert provides action labels conditioned on that context. By distilling expert actions directly conditioned on the student's on-policy history, the student learns to systematically resolve partial observability, i.e., search while uncertain and exploit once the hidden state is revealed. This enables efficient test-time exploration (similar to Bayesian posterior sampling) without incurring the impractical training-data costs of reinforcement learning. Across a diverse range of eleven different environments, including discrete exploration benchmarks, robotic manipulation, navigation, and Procgen environments, ASTEROID achieves ~2x higher returns under fixed exploration budgets, uses ~100x fewer interactions than RL baselines, and outperforms frontier models at exploration. Finally, our method transfers policies trained in simulation to vision-denied real-world robot manipulation, exhibiting efficient exploration and increasing success by +40% over standard distillation. Overall, it provides a simple yet surprisingly effective framework to learn efficient test-time exploration.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30105
PDF
-
AtomicRL: Atomic Reinforcement Learning for Long-Horizon Robotic Manipulation
2026-10-21 17:45:24Likui Zhang, Zhe Zhu, Zhihao Zhan, Yibo Bai, Hanlin Wang, Dongwei Li, Bin Wang, Tianle Zhang, Xiaodan Liang, Liang LinAbstract
Vision-language-action (VLA) models have shown promise in robotic manipulation, but imitation learning (IL) alone often struggles to master complex long-horizon skills. Reinforcement learning (RL) offers an effective avenue for further improving VLA policies. However, existing VLA-RL methods typically suffer from delayed and biased credit assignment due to trajectory-level sparse rewards, while imbalanced subtask mastery creates bottlenecks that limit overall performance. To address these challenges, we propose AtomicRL, an RL post-training framework that leverages atomic task structures for reliable credit assignment and adaptive exploration. AtomicRL introduces Semantic-Aware Policy Optimization (SAPO), which uses semantic-aware gating to constrain TD residuals from failed atomic subtasks from propagating across task boundaries, mitigating credit contamination while preserving intra-subtask credit assignment and value bootstrapping. In addition, Bottleneck-Adaptive Target Exploration (BATE) dynamically coordinates task sampling and local exploration incentives based on the mastery of individual atomic subtasks, directing learning toward the remaining bottlenecks. Finally, we develop a Real2Sim2Real pipeline that bridges online RL in simulation with real-world deployment, facilitating practical transfer to physical robots. On LIBERO-LONG, AtomicRL achieves absolute performance gains of 3.6% and 4.0% over PPO across two VLA architectures. On CALVIN ABC-D, it improves full-task performance by up to 9.5% over PPO. Combined with the Real2Sim2Real pipeline, AtomicRL further improves real-world task performance by an average of 21.3%. These results demonstrate that AtomicRL effectively improves the learning efficiency and manipulation performance of VLA policies on long-horizon tasks.
Abstract (translated)
URL
https://arxiv.org/abs/2610.30097
PDF
-
RangeNav: Task-Aligned LiDAR Range-Image Pretraining for Robot Navigation
2026-10-21 15:06:08Samuel Foo Enze, Manthan Patel, Fan Yang, Linus Kramer, Pascal Roth, Marco Hutter, Cesar CadenaAbstract
Autonomous mobile robots operating in dense crowds and cluttered sites require $360^\circ$ perception to react to surrounding hazards. While panoramic LiDAR offers full-surround coverage, learning representations within reinforcement learning (RL) is sample-inefficient under sparse rewards, and existing 3D pretrained point-cloud encoders are too memory-heavy for massively parallel RL simulation. We present RangeNav, a compact (183k parameters) LiDAR range-image encoder pretrained entirely on synthetic data. RangeNav uses anisotropic convolutions and circular padding to handle the extreme aspect ratio and azimuthal seam of range images. We pretrain the network using a navigation-aligned objective combining range reconstruction, semantic segmentation, and bird's-eye-view (BEV) prediction, then freeze it for policy learning. In simulation, adding semantic and BEV supervision increases success rates over a reconstruction-only baseline by an average of 32.03 and 23.73 percentage points (pp) in procedurally generated training and out-of-distribution realistic test environments, respectively. Zero-shot real-world deployments on a Unitree B2-W quadruped demonstrate robust navigation through dense crowds, low-light settings, and construction zones where limited-field-of-view depth policies struggle. Code and models will be open-sourced.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29845
PDF
-
DriveMLP: When Training Strategy Outperforms Model Complexity in Autonomous Driving
2026-10-21 12:53:21Alessandro Canevaro, Julian Schmidt, Hang Yu, Ye Jin, Sajad Marvi, Georg Martius, Julian JordanAbstract
Learning motion planning policies that remain reliable under closed-loop execution is a central challenge in Autonomous Driving (AD). Imitation Learning (IL) scales well to large AD datasets but suffers from distribution shift, causal confusion, and compounding errors under closed-loop execution. In contrast, Reinforcement Learning (RL) typically requires handcrafted rule-based rewards and massive simulation. We study a simple closed-loop IL framework that uses RL purely as an optimization tool, guided by a trajectory-distance reward derived from logged expert data. We refer to this training strategy as Expert-Guided RL (EGRL). Through a controlled empirical study we find evidence that the training procedure, rather than model size and complexity, accounts for a substantial part of the performance gains. We then introduce DriveMLP, a lightweight MLP policy trained with EGRL, as a simple instantiation of these findings. It achieves a score of 86.51 on the nuPlan Val14 Reactive benchmark, a +3.2 point improvement over the previous best IL planner (Flow Planner), with >15x faster inference. Code is available at: this https URL .
Abstract (translated)
URL
https://arxiv.org/abs/2610.29676
PDF
-
Uni-Mem: Learning a Long-Horizon Agentic Memory Manager with Unified Online Policy Optimization
2026-10-21 12:48:10Zhiqian Xia, Jielong Tang, Ling-An Zeng, Shixiong Kai, Rong Yao, Hao Kong, Xiao Dong, Wei Wang, Mingxuan Yuan, Siyu Xia, Haifeng XiaAbstract
Long-term agent memory is shifting from predefined workflows toward learnable policies for memory construction and retrieval. Existing learning-based methods either optimize construction alone or combine retrieval-side online reinforcement learning (RL) with offline construction supervision, limiting co-adaptation between the two stages. In this paper, we explore end-to-end online RL for a unified memory manager (Uni-Mem) that learns both capabilities under a shared policy. However, joint optimization with existing RL algorithms such as GRPO faces confounded group comparisons and sparse feedback over long memory-management trajectories. To this end, we propose Stage-Decoupled GRPO, which alternates construction and retrieval optimization while controlling variation from the other stage within each rollout group, yielding stage-specific learning signals. We further introduce Event-Wise Credit Assignment (ECA), which localizes feedback to responsible events and propagates event-level advantages to their constituent operations. Experiments on LoCoMo and LongMemEval show that Uni-Mem outperforms the strongest RL-based memory, demonstrating the effectiveness of jointly learning memory construction and retrieval via Stage-Decoupled GRPO and ECA.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29671
PDF
-
CineEvolve: An Agentic Self-Evolution Framework for Cinematic Multi-Shot Video Generation
2026-10-21 11:30:09Jingzhe Zhang, Peixuan Zhang, Genying Li, Bing Ma, Shuchen Weng, Si Li, Boxin ShiAbstract
Cinematic multi-shot video generation requires both narrative planning and iterative refinement. Existing methods either lack narrative reasoning or rely on feed-forward workflows without policy updates. In this paper, we propose CineEvolve, an agentic self-evolution framework for cinematic multi-shot video generation that unifies a Director and a Critic within a single vision-language policy. Emulating real-world filmmaking that couples iterative direction with critical feedback, we train CineEvolve via alternating reinforcement learning, where the Director surfaces flaws to sharpen the Critic, and the updated Critic guides the Director's targeted revisions. To stabilize co-evolution, we introduce Evolutionary GRPO (E-GRPO), leveraging historical anchoring and revision safeguarding to mitigate advantage inflation. Additionally, we contribute CineTraj, a 40K cinematic trajectory dataset spanning creation, critique, and revision. Extensive experiments demonstrate that CineEvolve outperforms existing approaches. Our code will be available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29590
PDF
-
End-to-End Reinforcement Learning for Quadrupedal Jumping in Lunar Gravity
2026-10-21 11:27:05Levin M\"uller, Oliver Fischer, Marco Trentini, Yusuke Tanaka, Marco HutterAbstract
Legged robots can extend planetary exploration to rugged and discontinuous terrain that is inaccessible to conventional rovers. In lunar gravity, jumping further expands this mobility by enabling robots to cross obstacles and gaps that cannot be traversed by walking alone. We present a multistage reinforcement-learning approach that produces a single end-to-end control policy for commanded point-to-point jumping under lunar gravity. The resulting policy has a simple architecture and relies exclusively on mission-compatible onboard sensing, making its deployment on a real lunar mission plausible. In simulation, the learned policy performs accurate point-to-point jumps with landing errors smaller than the robot's body size. We analyze the torque and power demands of lunar jumping and characterize performance degradation under extreme outlier events, including control-update freezes of up to \SI{1}{s} during flight. Hardware transfer demonstrates that the policy can successfully execute a vertical jump under lunar-gravity offloading. Furthermore, the measured torque and power profiles are consistent with the simulation results and remain within the robot's hardware limits.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29586
PDF
-
Generalized CycleGRPO for Fine-Grained Multimodal Understanding and Grounding
2026-10-21 10:39:56Xin Zhang, Haochen Wang, Beibei Lin, Zhaoxiang Zhang, Robby T. Tan, Jason Li, Dacheng TaoAbstract
Fine-grained multimodal intelligence requires models to both understand a designated region and localize it from language. Existing methods learn these directions with separate objectives and task-specific annotations. We present Generalized CycleGRPO, the first unified reinforcement learning framework coupling region understanding and localization across spatial, temporal, and spatio-temporal domains. Given a region, a multimodal large language model generates candidate descriptions and then localizes each description to reconstruct the input region. Reconstruction overlap provides a verifiable reward for captioning, while individual localization rewards improve the inverse mapping. A common region-space formulation extends this cycle from image masks and boxes to video intervals and timestamped spatial evidence, without reference captions or an external language-model judge during reinforcement learning. In videos, the cycle encourages descriptions that distinguish a target event from similar recurrences. We introduce DOVE (Distinctive Occurrence-level Video-caption Evaluation), a human-verified contrastive benchmark for this temporal distinctiveness. CycleGRPO improves SAMTok's overall gIoU from 57.5% to 68.2% on GroundingSuite and average gIoU from 71.2% to 78.7% on GRES. For temporal tasks, it achieves 51.9% tIoU on Charades and 33.3% Temporal Evidence Coverage on DOVE, compared with 45.6% and 14.8% for the Qwen3-VL initialization. These results establish region-text cycle consistency as a general self-evaluation principle for multimodal understanding and localization. Models, code, and DOVE are available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29552
PDF
-
SafeDojo: Safe Reinforcement Learning for VLA via Interactive World Model
2026-10-21 10:06:03Kai Tang, Peidong Jia, Zhong Chu, Jiajun Cao, Jixian Wu, Bohua Zhang, Rui Ma, Fangyuan Zhao, Sixiang Chen, Yichen Guo, Xiaowei Chi, Chun-Kai Fan, Kevin Zhang, Jinchang Xu, Fubing Yang, Weishi Mi, Xiaozhu Ju, Jian Tang, Shanghang ZhangAbstract
Safe control is a prerequisite for real-world embodied intelligence, for which safe reinforcement learning has emerged as a promising paradigm. However, existing safe reinforcement learning methods either require costly real-world exploration or depend on hand-crafted safety functions. Neither scales to vision-language-action models deployed in open-world physical environments. We propose SafeDojo, a model-based safe reinforcement learning framework that improves vision-language-action policies through world-model imagination. SafeDojo performs online reinforcement learning on top of an interactive video world model. A ResNet success classifier scores predicted future frames, while a lightweight safety head predicts per-step contact probabilities from the current latent context and proposed action chunk, jointly assessing task execution and trajectory safety. These decoupled signals are balanced through Lagrangian-constrained GRPO. On SafeLIBERO Level I, SafeDojo improves average safe-success rate by 8.25 percentage points over the strongest baseline; on the zero-shot Level II setting, it is competitive with the strongest model-free baseline. Real-world Franka deployment further shows the best average task and safe-success rates across five tasks. These results support world-model-based safe reinforcement learning as a promising path toward safer embodied intelligence.
Abstract (translated)
URL
https://arxiv.org/abs/2606.20698
PDF
-
Harness Scaling for Generalizable Agentic Reinforcement Learning
2026-10-21 09:15:50Fei Bai, Huatong Song, Chuan Hao, Ran Tao, Jian Yang, Bryan Dai, Wayne Xin Zhao, Ji-Rong WenAbstract
Agent harnesses have become a central component of modern agentic systems, shaping how models use tools, manage context, and coordinate long-horizon interaction. Realizing the benefits of these harnesses depends on how effectively the underlying model can make use of them, motivating reinforcement learning directly through the harness. However, existing approaches typically train through a single fixed harness, which can lead to specialization and limited transfer across harness configurations. To address this, we introduce \textbf{Harness Scaling}, a new scaling dimension for agentic RL that extends training from a single harness to diverse execution configurations. Concretely, we construct a factorized harness space, select a compact training subset with balanced factorial coverage, and organize cross-harness training to combine global coverage with controlled variation for the same task. Across multiple benchmarks, model scales, and RL algorithms, Harness Scaling consistently improves generalization to unseen white-box configurations. Moreover, models trained through structured white-box variation transfer effectively to unseen external black-box harnesses, outperforming models trained through a single complex black-box harness. These results establish harness diversity as a meaningful scaling dimension for agentic RL and highlight the importance of how this diversity is organized for cross-harness generalization.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29458
PDF
-
Distance Field Driven Reinforcement Learning for Reactive and Robust Collision Avoidance
2026-10-21 09:10:03Taiga Ikeshima, Sheila Sutjipto, Lan Wu, Francisco J. Naranjo-Campos, Taiki Ishigaki, Taisuke Kobayashi, Teresa A. Vidal-Calleja, Eiichi YoshidaAbstract
Robust and reactive collision avoidance remains a central challenge for deploying Reinforcement Learning (RL) policies on real robotic manipulators, particularly when transferring behaviours learned in simulation to the physical world. This paper presents an RL framework that integrates the Euclidean Distance Field (EDF) directly into the observation space, providing the agent with continuous geometric cues that support stable and responsive avoidance behaviours. By leveraging the rich spatial information encoded in the EDF, the method enables a streamlined reward formulation that achieves effective obstacle avoidance both with and without Artificial Potential Field guidance. The proposed approach is validated on a 7-DoF redundant manipulator, demonstrating zero-shot transfer from simulation to real hardware. Despite being trained exclusively with randomised static objects, the deployed policy generalizes to previously unseen real environments and exhibits robust online avoidance of dynamic obstacles. These results highlight the potential of EDF-based representations to mitigate the sim-to-real gap and support reliable and real-time collision-aware behaviour in complex and changing environments.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29452
PDF
-
BIRD-RL: Scaling Agentic Reinforcement Learning over Stateful Data-Centric Environments
2026-10-21 08:50:39Ge Qu, Jinyang Li, Xiaohan Xu, Xiaolong Li, Bowen Qin, Nan Huo, Ziwei Tang, Chenhao Ma, Reynold ChengAbstract
While agentic reinforcement learning (RL) has emerged as a promising paradigm for training small language models (SLMs), existing frameworks are predominantly tool-centric and struggle with data-centric environments. Specifically, applying tool-centric frameworks to stateful tasks like SQL debugging introduces severe spatial inefficiencies since proper environment isolation typically requires a separate database container per rollout. Naively reusing databases without tracking evolving states across multi-turn interactions causes trajectory contamination and cross-trajectory interference, which severely destabilizes policy updates. To overcome these bottlenecks, we introduce BIRD-RL, a novel four-stage agentic RL curriculum that alternately enhances foundational agent reasoning and trajectory-adaptive on-policy exploration to circumvent the instability inherently associated with direct end-to-end learning. Crucially, to support this curriculum, BIRD-RL is driven by a Trajectory-scoped Persistent Execution Infrastructure, which enables safe database replica reuse while preserving trajectory isolation. This design improves spatial RL efficiency by 12.8x, successfully enabling scalable agentic RL for complex data-centric tasks without needs of Kubernettes. Empirical evaluations on the BIRD-CRITIC-SQLite benchmark demonstrate that our 7B and 14B models achieve Success Rates of 44.40% and 48.00%, respectively, showing competitive performance against strong proprietary model-based agents such as Claude-Opus-4.6 Agent (48.20%). Furthermore, BIRD-RL successfully trains unified SLMs capable of both advanced SQL debugging and generation, highlighting its broader potential for multi-objective optimization across heterogeneous, real-world data-centric tasks.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29426
PDF
-
AdvFidelity: Unifying Preference Alignment and Few-Step Image Generation within Reinforcement Learning
2026-10-21 08:09:55Sicong Kang, Fangao Zeng, Yuhan Li, Hao Zhou, Mengfei Xu, Wenxiang Shang, Wei Li, Bingbing Ni, Pipei HuangAbstract
Practical image generation requires both human preference alignment and few-step efficiency, yet combining reinforcement learning (RL) with distillation is often multi-stage and may erode RL gains We observe that few-step RL progressively improves structural coherence, even with specialized rewards, but leaves low-level artifacts that these rewards may fail to penalize. To address this gap, we propose AdvFidelity, which integrates artifact-sensitive fidelity supervision into RL through an online adversarial reward. This enables joint reward alignment and high-fidelity few-step generation within a single RL stage, without distribution matching distillation (DMD) or other explicit distillation objectives. We train a discriminator to rank high-quality references (positives) above content-matched reconstructions (negatives) produced by the current generator. Reconstructing moderately noised references largely preserves style and scene layout while exposing low-level degradations. This encourages the discriminator to assess artifacts rather than content differences. With negatives reconstructed online, alternating updates adapt fidelity supervision to the generator's evolving failure modes. References can come from existing target-domain data or full-step model samples. Built on LeapAlign, AdvFidelity backpropagates gradients from both rewards through the few-step sampling trajectories used at inference. Across two model backbones and three task settings, AdvFidelity matches full-step LeapAlign in alignment, task performance, and overall generation quality and remains highly competitive with other evaluated full-step RL methods, using only 4 or 8 steps without CFG.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29380
PDF
-
Training in Deployment: Human-in-the-loop Online Post-Training for Robot Manipulation
2026-10-21 06:15:00Chun Gu, Xiaofei Wei, Yiming Chen, Bozhou Zhang, Juntong Zhang, Xiatian Zhu, Li ZhangAbstract
Large pretrained robot foundation models provide strong priors for manipulation, yet adapting them to target tasks with offline teleoperated demonstrations often fails to cover the contact-rich, recovery, and precision-critical states encountered during deployment. Existing methods mitigate this issue by augmenting offline demonstrations with human-in-the-loop data or online reinforcement learning. Despite this, offline post-training adapts slowly because it cannot immediately absorb deployment experience, while online RL typically relies on value functions or auxiliary actor-critic models, increasing both training complexity and instability. We present TIDE (Training In DEployment), a human-in-the-loop online post-training system. TIDE introduces a provenance-conditioned supervised learning framework that updates the policy online from human interventions and policy-generated rollouts collected during deployment, without a learned critic, reward model, or action scorer. Across four real-world bimanual manipulation tasks, TIDE improves milestone-normalized task progress and binary full-task completion over the warm-start policy and matched additional offline demonstrations, and performs comparably to the fixed-policy collection baseline while making updated checkpoints available during deployment. Each task-specific online post-training session, from the start of deployment rollouts until the final checkpoint is available, completes in less than one hour while learning and checkpoint handling run asynchronously.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29264
PDF
-
Reasoning-Search-Augmented LLMs: An Operational Taxonomy, Benchmark Analysis and Controlled Evidence Synthesis
2026-10-21 04:19:06Biswas Poudel, Nilson Chapagain, Amit Kumar, Xianshun JiangAbstract
Large language models (LLMs) demonstrate impressive emergent reasoning capabilities, yet they remain constrained by fixed model parameters and finite context windows. Even when augmented with static retrieval, they can produce statements that are unsupported by, or even contradictory to, retrieved evidence. This leads to hallucinations, difficulty incorporating up-to-date information, and failures to correctly interpret or reconcile sources. Recent work addresses these limitations by tightly coupling step-by-step reasoning with interactive search, treating retrieval as an action within an iterative agentic loop rather than a one-shot preprocessing step. This study reviews frameworks that implement such reasoning-search loops by interleaving planning, querying, retrieval, integration, verification, and stopping. We analyze these systems through four architectural axes: retrieval coupling, controller design, knowledge source, and grounding strictness. We separately treat the learning signal as a cross-cutting dimension because prompting, supervision, preference optimization, and reinforcement learning can instantiate the same architecture while producing materially different search, stopping, and grounding behavior. We further distinguish benchmark difficulty from RSA-loop demand and synthesize controlled within-study ablations and quality-cost deltas, avoiding invalid pooling across heterogeneous model backbones, retrievers, corpora, and evaluation protocols. Our goal is to clarify the landscape of reasoning-search-augmented LLMs and outline a unified perspective on the design choices shaping future research.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29189
PDF
-
State-Dependent Objective Coordination for End-to-End Balance and Navigation
2026-10-21 03:24:37Xuhao Huang, Xitao Xu, Yuxin Zheng, Yunfan Ren, Zheng GongAbstract
Traditional robotic navigation systems separate mapping, planning, and low-level control through cascaded modules, introducing computational overhead and accumulated latency. End-to-end reinforcement learning offers a compact unified policy that maps onboard observations directly to control actions. However, on inherently unstable platforms such as two-wheeled inverted-pendulum robots, navigation progress and obstacle avoidance can conflict with attitude stabilization in a strongly state-dependent manner. We propose a lightweight state-dependent objective coordination method for integrated navigation and control. A pitch-conditioned scalarization coefficient $\beta(s)\in[0,1]$ adjusts the relative contributions of balance stabilization and navigation objectives during training, without extra control hierarchies or auxiliary attitude-stabilization controllers, such as proportional-integral-derivative (PID) or linear-quadratic regulator (LQR) controllers. To support zero-shot deployment, simulation training models execution delay, action hold, asynchronous and noisy observations, and dynamics uncertainty. Gradient analysis reveals persistent directional interference and state-dependent variation in the objectives' relative scales. Comparisons with alternative reward-weighting, constraint, training, and control-structure strategies show that PPO+$\beta(s)$, where PPO denotes proximal policy optimization, achieves 98% and 95% success rates with small attitude excursions on the fixed obstacle map and the randomly generated high-obstacle-density map, respectively. Zero-shot hardware experiments further demonstrate bounded attitude behavior and consistent obstacle avoidance using only onboard observations. Training and evaluation code will be made publicly available upon acceptance.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29142
PDF
-
Long-MDR: Long-Context Reinforcement Learning for Multimodal Deep-Research Agents
2026-10-21 01:27:11On Tai Tang, Huanyao Zhang, Jiepeng Zhou, Wentao ZhangAbstract
The next generation of multimodal research agents must reason over long-lived research histories rather than short model completions. During a single task, an agent may repeatedly search the web, inspect visual evidence, revisit earlier hypotheses, and accumulate tens of thousands of tokens of multimodal context. Despite this trend, online RL for multimodal research agents remains largely confined to shorter contexts and interaction horizons. We push online RL training to 128k context and 75+ tool-interaction turns. To our knowledge, this is the first online multimodal deep-research RL study trained at 128k context, and the first trained with a 75 tool-turn horizon. Scaling to this regime exposes several practical limitations of conventional RL training. Early in training, weak policies make poor use of large interaction budgets, causing expensive rollouts with little reward improvement. Later, policy entropy can collapse before performance has saturated, prematurely ending useful learning. We introduce Long-MDR, a three-component training recipe designed specifically for this setting: On-Policy Distillation Warmup, Progressive Horizon Expansion, and Entropy-Triggered Rescue. Together, these techniques improve both the learning efficiency and stability of long-horizon RL, enabling continued gains in a regime where direct training is slow and costly. Code will be available at this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2610.05195
PDF
-
HybridResearcher: Bridging the Structured-Unstructured Boundary for Deep Research Agents
2026-10-21 00:41:07Soyoung Yoon, Peiran Xu, Han Wang, Ruofan Wu, Xiaodong Yu, Feng Yan, Mohit Bansal, Debiao Li, Seung-Won Hwang, Yuxiong He, Zhewei YaoAbstract
While LLM agents are increasingly capable of gathering and synthesizing evidence from a single homogeneous source, such as structured tables or unstructured data, real-world analytical questions frequently span heterogeneous sources. For example, answering one enterprise question may require both a result computed over an internal database and evidence found on the web. Recent evaluations on benchmarks like HybridDeepResearch show that even frontier agents struggle when evidence must cross this boundary. While these benchmarks have exposed the limitation, methodologies to explicitly teach cross-source reasoning remain absent. To address this, we introduce HybridResearcher, a 4B agent trained to bridge evidence across the structured and unstructured boundary. To build it, we synthesize cross-source bridge trajectories by starting with validated single-source tasks-such as existing text-to-SQL or web search queries-and fabricating the missing source around a bridge entity drawn from the seed's own output. We retain a sample only when strict construction gates confirm that neither side alone suffices, ensuring the data requires true cross-source interdependence. Finally, we train on this curated dataset using supervised fine-tuning followed by reinforcement learning. Evaluated on the existing HybridDeepResearch benchmark, HybridResearcher improves over our base model, Qwen3.5-4B, by 30.0 points pass@8 and surpasses the far larger Qwen3.5-397B-A17B by 8.3 points, while also improving over its base model on established, held-out single-source benchmarks (BrowseComp and BrowseComp-Plus), demonstrating that this complex cross-source reasoning can be taught directly to small models without triggering catastrophic forgetting.
Abstract (translated)
URL
https://arxiv.org/abs/2610.24899
PDF
-
FlashNext-RL: A System Optimization Study for Long-Horizon Agentic RL of Qwen3.8-Flash-Next
2026-10-20 23:53:09Zhewei Yao, Yuxiong HeAbstract
Long-horizon agentic reinforcement learning (RL) requires both efficient rollout generation and training on long sequences. However, the memory requirements and model-specific operators of Qwen3.8-Flash-Next make this difficult with existing open-source implementations. To resolve these limitations, we introduce FlashNext-RL, a model-specific RL training optimization framework based on prime-rl, with vLLM for software engineering (SWE) rollout generation. Our main focus is BF16 training of the full 48-layer model on a single 8xB200 node, with the first two layers, the n-gram table, the sparse-attention indexer and the multi-token prediction head frozen. We optimize host offload, gradient accumulation and the kernels for mixture-of-experts, sparse attention and the gated residual stream. In the full-model development study, non-padding training throughput increases from approximately 395 tokens/s/GPU in the initial 32K-token configuration to 4,438 tokens/s/GPU in the optimized 131K-token configuration, with hardware matmul throughput (aka without considering the cost of attention components) increasing from approximately 20 to 213 TFLOPs/GPU. A controlled cumulative study on an 8-layer model further quantifies the contributions of individual optimizations. Together with three inference nodes, the system supports a live SWE RL run with a target batch of 1,024 trainable rollouts and a 131,072-token context limit. Over steps 5-15, an iteration takes 22.4 minutes on average, with bounded entropy and trainer-sampler mismatch KL. These results highlight the importance of tailoring RL training to the model's architecture, memory layout and kernel requirements, rather than relying only on general-purpose RL infrastructure.
Abstract (translated)
URL
https://arxiv.org/abs/2610.28998
PDF
