Reinforcement_Learning
-
Flat-RLVR: Sharpness-Aware Reinforcement Learning with Verifiable Rewards for Reasoning Capacity in LLMs
2026-10-21 06:35:19Yeonbum Kim, Hyun Kyu Lee, Sung Whan YoonAbstract
Reinforcement learning with verifiable rewards (RLVR), particularly through group-relative policy optimization (GRPO), has emerged as a central paradigm for the post-training of large language models (LLMs). While RLVR is highly effective at enhancing the reasoning capabilities of LLMs, it commonly induces \textit{distribution sharpening}, whereby the policy distribution becomes excessively peaked and rare-yet-correct trajectories are progressively suppressed, thus diminishing sample diversity. To this end, we promote flat-reward optima in RLVR to smooth the policy distribution. We theoretically link flat reward RLVR to reduced distribution sharpening, and empirically show improved sample diversity while maintaining single-attempt accuracy. Our method, called \texttt{Flat}-RLVR, searches for flat reward regions via sharpness-aware minimization during LLM post-training, combining this with GRPO variants to achieve better trade-offs between sampling efficiency and reasoning capacity.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29281
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
-
Value-Rectified Distillation for Flow-based Offline Reinforcement Learning
2026-10-21 01:58:00Ke Jiang, Wen Jiang, Yoshinobu Kawahara, Xiaoyang TanAbstract
We aim to address the problem of Out-of-Distribution (OOD) Mode Averaging, a critical problem that emerges when distilling Flow Matching (FM) based behavior policies in offline Reinforcement Learning (RL): When learning over multi-modal datasets, standard distillation inadvertently forces the policy to interpolate across different behavioral modes, causing the generated actions to collapse into invalid, OOD regions. To resolve this, we propose Value-Rectified Distillation (VRD), a novel framework that reformulates offline policy distillation as a value-guided generative trajectory alignment problem. By employing a dynamic coupling mechanism, VRD actively rectifies flow dynamics to explicitly decouple distinct behavioral modes, cleanly isolating high-value actions. Specifically, we instantiate VRD through two complementary algorithms: VRD-Q, which utilizes value-quantile behavior distillation to statistically filter out low-value behaviors, and VRD-R, which reorganizes the latent space to geometrically isolate high-value generative trajectories from the FM-based behavior policy. Extensive empirical evaluations demonstrate that VRD improves existing generative baselines across diverse continuous control and planning tasks on the D4RL, OGBench, and Minari benchmarks, especially over those multi-modal, non-expert datasets.
Abstract (translated)
URL
https://arxiv.org/abs/2610.29075
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
-
Black-Box Red-Teaming of Multi-Agent Systems via Reinforcement Learning
2026-10-20 22:32:22Yuqi Jia, Hangsheng Zhang, Chenxi Liu, Junjie Liang, Xun ChenAbstract
Large language model (LLM) agents are increasingly deployed in multi-agent systems (MAS) to accomplish complex tasks. Prior black-box red-teaming attacks mainly focus on single agents, but we find that these methods are far less effective in MAS, where multiple sub-agents may not directly interact with the user. Therefore, we introduce ReMAS, the first reinforcement learning-based red-teaming framework tailored to MASs, which fine-tunes attacker LLMs to generate effective adversarial prompts for system prompt extraction. The framework follows a two-step process: first, a rewrite stage refines base attack prompts to increase extraction success, and second, a template generation stage constructs attack templates that improve the likelihood of invoking specific sub-agents and thus revealing their system prompts. Extensive experiments show that our method substantially improves attack success rates compared to existing black-box attacks, with transferability across different backend LLMs and MAS structures. These results underscore the vulnerabilities of multi-agent systems and the importance of developing stronger defenses against adaptive black-box adversaries.
Abstract (translated)
URL
https://arxiv.org/abs/2610.28966
PDF
-
Mitigating Reward Hacking in Preference Reward Models with Evolving Penalties
2026-10-20 22:14:00Afonso Baldo, Ant\'onio Farinhas, Marcos Treviso, Ricardo Rei, Nuno M. GuerreiroAbstract
Reinforcement learning for language models has moved from preference reward models to verifiable rewards, which sustain much longer optimization. Preference models are prone to reward hacking: the policy raises their score without improving response quality, or even while degrading it. Yet for open-ended tasks, where quality can neither be verified nor fully written down as rubrics, they remain valuable. We introduce Dynalize, which mitigates reward hacking online, during policy training. Periodically and asynchronously, a language model analyzes training rollouts and writes penalties for undesirable behaviors that the reward model appears to reward; a judge applies them to every rollout. In clinical counseling and creative writing, training on the reward model alone raises its score continuously but ultimately leads to a decline in external benchmarks; Dynalize sustains longer training and results in improved downstream performance. We find that our method suppresses patterns incentivized by preference models that, without penalties, spread to most responses, and even to the model's reasoning, which the reward model does not explicitly score. Furthermore, we show that its gains hold in multi-turn training, and even after we retrain the reward model on the discovered behaviors.
Abstract (translated)
URL
https://arxiv.org/abs/2610.26496
PDF
-
SLoSa: Small, Low-Power, and Safe Foundation Models for Robotic Task Planning
2026-10-20 22:09:44Weizhe Xu, Jialiang Fan, Essien Taylor, Zijun Wang, Russell Joseph, Fanxin KongAbstract
Safe robotic task planning requires planners that can reliably achieve task goals while satisfying safety constraints. Classical planners search over explicit symbolic task representations and can provide formal guarantees with respect to encoded safety constraints, but require rigid input formats. Their computational cost also increases rapidly with the search space. In contrast, large language models (LLMs) offer flexible task interfaces and strong generalization capabilities, but they are large, power-hungry, and pre-trained with little robot-specific safety data and knowledge. Thus, they can be difficult to deploy locally on robotic systems with limited resources. To address these limitations, we present SLoSa, a neuro-symbolic compression and fine-tuning framework for training foundation-model-based robotic task planners with Small size, Low power demand, and Safety awareness. The framework first constructs a safety-aware dataset using classical planners, an advanced teacher LLM, and formal verifiers, followed by two training stages: supervised fine-tuning and reinforcement learning (RL) fine-tuning guided by a verification-based reward machine to enforce safety alignment. We evaluate SLoSa across four Planning Domain Definition Language (PDDL) domains, two household simulation environments, local edge platforms, and a desktop robotic arm. Across the PDDL tasks, verifier-guided post-training raises the success rate of compact 4B-14B planners to 70.5-76.0%, while INT4 deployment on a Jetson Orin requires 17.85-26.41 W of average power. Results in household simulators and on the physical robot further demonstrate the applicability of SLoSa across diverse robotic task settings.
Abstract (translated)
URL
https://arxiv.org/abs/2610.28955
PDF
-
One Policy, Many Styles: Where to Aggregate Preferences for Steerable Humanoid Control?
2026-10-20 20:59:05Ziqian Huang, Yuzhuo Chen, Shimin Zhang, Junyang He, Xinyue Annie Yang, Animesh GargAbstract
Modern robotic agents are increasingly expected to support steerable control, where a single policy can adjust its behavioral style at runtime under changing user preferences. Preference-conditioned reinforcement learning provides this interface, but conditioning alone does not ensure that a requested behavior change is strong, does not leak into unrelated behaviors, and does not compromise task performance. Our evaluation combines Preference Reward Responsiveness (PRR) with coverage and utility metrics. We use these measures to compare reward-level, advantage-level, loss-level, and PCGrad-based aggregation across two humanoid-control frameworks. No aggregation level performs best across all environments and criteria. Experiments show that aggregating objectives early produces a wider range of behavioral responses but greater unintended changes in other objectives, while aggregating them later improves control precision at some cost to responsiveness.
Abstract (translated)
URL
https://arxiv.org/abs/2610.28925
PDF
-
JaxMARL-BC: Multi-Agent Reinforcement Learning for Macroeconomic Modelling, in Seconds
2026-10-20 20:35:44Daniele Montagnani, Andrea Coletta, Antonietta Mira, Aldo GlielmoAbstract
Multi-agent reinforcement learning (MARL) offers a principled way of populating simulated economies with many heterogeneous decision makers. Its adoption, pioneered in finance, has nonetheless been slow in macroeconomics because of two main limitations: the lack of a shared frame of reference connecting MARL systems to dynamic stochastic general equilibrium (DSGE) models the profession already trusts, and the very long training times. The first limitation was recently addressed by the Multi-Agent Reinforcement Learning Business Cycle (MARL-BC), which places deep RL households in a real business cycle economy and reduces to classic DSGE models in specific limits. Here we address the second limitation. We present JaxMARL-BC, a JAX-accelerated simulator that trains MARL-BC economies in seconds (between 75x and 724x faster than the reference implementation), and scales to 200,000 independent learning households. We verify its correctness against the known DSGE limits, characterise its scaling through a simple cost law, and demonstrate its value by a proof-of-concept calibration of an extended model with heterogeneous returns on capital, in a two-hour experiment that would previously have taken over a month. We release JaxMARL-BC as open source.
Abstract (translated)
URL
https://arxiv.org/abs/2610.28907
PDF
-
RL-IF: Tracking Non-local Training Data Influence for Online Reinforcement Learning
2026-10-20 20:01:25Shixuan Liu, Cheng Tang, Yuzheng Hu, Fan Wu, Han Zhao, Jiaqi W. MaAbstract
Online reinforcement learning (RL) has achieved significant success in decision-making tasks, but modern online RL remains highly sensitive to the quality of training experience. Understanding the contribution of individual training instances is therefore central to data-centric RL, with applications ranging from experience selection to policy interpretation. Existing data attribution methods for online RL primarily focus on local influence within a single training round, overlooking the non-local data influence across multiple training rounds. In this work, we formalize data attribution in online RL through a trajectory-level leave-one-out quantity, RL-LOO, which measures how one training sample influences a downstream target after subsequent online updates have taken place. We then derive RL-IF, a first-order estimator that propagates data influence through both optimization effects and policy-induced sampling effects, and provide theoretical guarantees on its approximation accuracy under smoothness assumptions. Our analysis further disentangles these two effects and shows both theoretically and empirically that the full RL-IF estimator can often be well approximated using the optimization effect alone. Building on this observation, we develop RINSE, a practical data filtering method for online RL. Experiments on standard control benchmarks and an RLHF-style toxicity-mitigation setting show that RINSE improves training efficiency and final performance over standard PPO and local attribution baselines. These results suggest that RL-IF provides a practical and principled framework for understanding and improving online RL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.28883
PDF
-
DualHeads: Reward-tilted few-step diffusion made simple
2026-10-20 19:31:03Yuanzhi Zhu, Eleftherios Tsonis, Simo Alami Chehboune, Xi Wang, St\'ephane Lathuili\`ere, Vicky KalogeitonAbstract
Distilled diffusion models generate high-quality images in a few sampling steps, but aligning them with task requirements and human preferences is difficult. Methods developed for multi-step diffusion models rely on differentiable rewards, or on score fields and stochastic trajectories that distilled generators may lack. We ask whether a few-step generator can be adapted using only external reward evaluations and reference samples, and show that two binary classifiers suffice. Our method, DualHeads, implements them as lightweight heads on a frozen diffusion backbone. The reward head learns which samples of the same prompt score higher, the reference head separates reference images from generated ones, and the generator maximizes a weighted sum of their logits. When the heads are optimal, this objective is KL-regularized maximization of the learned reward, and the ratio of head weights sets the regularization strength. Empirically, with both rule-based and model-based rewards, a few-step DualHeads generator reaches task rewards close to those of 40-step reinforcement learning methods while preserving image quality, at no extra inference cost. We hope that this simple and elegant dual-head formulation will help advance few-step reward-tilted generation.
Abstract (translated)
URL
https://arxiv.org/abs/2610.28858
PDF
-
Self-organizing reinforcement learning: a learner-free swarm approaches the explore-exploit optimum
2026-10-20 19:20:05Galen J. WilkersonAbstract
The explore/exploit trade-off is the core problem of reinforcement learning (RL). We show that a minimal foraging swarm approaches its optimum - the ideal-free / Nash allocation of a multi-agent resource-sharing game (an equilibrium, not a return maximum) - with no learner and no reward signal. Agents explore by heavy-tailed (Levy) flights, recruit neighbours to a discovery through a threshold cascade on their proximity network, and disperse as a patch crowds; this cascade cycle balances exploration and exploitation and drives foragers toward the ideal free distribution. The swarm attains the shape of this optimum (committed foragers track richness with slope 0.94) but not its exact equilibrium - it stays measurably exploitable - reaching it dynamically (~150 steps) through the coupling of recruitment to movement: remove the coupling and matching collapses (0.94 -> 0.08). The RL sits at the level of the problem and the allocation reached, not the micro-rule: no step evaluates payoff (reward-free by construction, though the crowd signal is informationally tied to payoff), and the swarm reaches the allocation without learning it (quality plateaus and re-settles after resources move, no cumulative gain). Against a trained baseline the comparison is two-sided: in an easier, disembodied game a learner is more precise, while in the same embodied environment the reward-free swarm reaches ~65% of a validated oracle ceiling at zero training and leads until trained PPO and centralised-critic (CTDE) learners overtake it (~74-75%, 16 seeds, all above the swarm mean) after ~4,000 of 15,000 training steps - its advantage is in budget, not asymptote. In this one setting we read it not as beating RL but as a strong zero-shot, reward-free operating point: a demonstration that the self-organizing physics of a minimal swarm can carry a collective close to the optimum RL seeks, with none of RL's machinery.
Abstract (translated)
URL
https://arxiv.org/abs/2610.25113
PDF
-
HorizonComposer: Spatiotemporally Consistent Driving Video Editing with Enriched Traffic Semantics
2026-10-20 18:51:48Mauricio Soroco, Yuqiu Liu, Zaid Tasneem, Francesco Pittaluga, Abhishek Aich, Wuyang Chen, Manmohan Chandraker, Ziyu JiangAbstract
Instruction-guided video generative models offer a scalable solution for stress-testing autonomous driving systems by simulating diverse environmental conditions. However, driving scenes operate within a dense Operational Design Domain (ODD) governed by strict traffic semantics. When applied to global driving-condition edits, general-purpose instruction-guided video editors frequently corrupt these semantics, leading to structural hallucinations and ``temporal shock.'' This work distills and protects traffic semantics during video editing at three levels. First, we construct a hybrid dataset by combining temporally aligned real-world videos with synthetic edits, balancing foundational scene structure with unambiguous appearance signals. Second, during training, we adapt advantage conditioning from offline reinforcement learning to align the model using dense rewards. We extract dense, pixel-level rewards from the hybrid data to serve as a localized signal and actively steer generation toward structurally reliable pixels. Finally, our model mitigates temporal shock by introducing thinking frames, a transition mechanism that seamlessly reconciles static image conditioning with dynamic video context. Extensive experiments demonstrate that our method prevents the corruption of multi-actor dynamics, establishing state-of-the-art spatiotemporal consistency, yielding consistent user preference (39% relative improvement in content preservation and 12% in temporal consistency), and achieving a significant 5% gain in downstream road line segmentation and 7% gain in downstream multi-object detection.
Abstract (translated)
URL
https://arxiv.org/abs/2610.28829
PDF
-
LIFT: Scoped Constraints for Evaluating and Training Instruction Following
2026-10-20 18:40:22Eran Goldstein, Yuval Belfer, Yonatan Belinkov, Daniel GissinAbstract
Common instruction following (IF) benchmarks, such as IFEval, evaluate the ability of language models (LMs) to follow simple constraints that can easily be verified programmatically. These same constraints, in turn, are used as reward signals for reinforcement learning. This design may lead to reward hacking of constraints, pushing models away from the main task. Recent attempts to address this resulted in benchmarks with richer constraints, but required expert labor, remained closed-access, and cannot be easily extended. We present Lift-Bench, an IF benchmark that asks a model to draft a document given a topic and a structured outline. Constraints are scoped to certain document sections, providing finer control over constraint density. Lift-Bench includes both verifiable and judged constraints, as well as a training counterpart, Lift-Train. Importantly, Lift-Bench provides both main task and instruction-following constraints, mitigating shallow reward hacking. On experiments with multiple models, we find that frontier models struggle with the harder subsets of our benchmark and analyze model performance along multiple axes. We also find that training on Lift-Train is much less susceptible to reward hacking than training on the common IFTrain. Lift-Bench and Lift-Train are publicly available, as well as the data generating pipeline.
Abstract (translated)
URL
https://arxiv.org/abs/2610.28819
PDF
-
Brittle by Design? Multi-Scenario Reinforcement Learning for Robust Mission Coordination in Systems of Systems
2026-10-20 18:40:12\.Ibrahim O\u{g}uz \c{C}etinkaya, Sajad Khodadadian, Taylan TopcuAbstract
As systems engineering (SE) shifts from monolithic systems to Systems of Systems, Mission Engineering (ME) has emerged as a new line of thinking in SE. Mission environments are uncertain and dynamic, and outcomes depend on how mission assets interact with them, rendering static architectures brittle and calling for flexible yet analytically rigorous ME approaches. We propose an intelligent mission coordination methodology integrating high-fidelity digital mission models with Reinforcement Learning (RL) for adaptive task allocation. Using a Digital Engineering (DE) infrastructure comprising a digital mission model and agent-based simulation, we formulate mission tactics management as a Markov Decision Process and train an RL coordinator with Proximal Policy Optimization to map mission states to tactical actions from realized outcomes. An aerial firefighting mission with three scenarios demonstrates the approach. First, all trained coordinators significantly outperform random and heuristic baselines. Second, single-scenario coordinators can be brittle: the coordinator trained on the more forgiving scenario degrades to random tactics performance in more demanding scenarios, whereas the demanding-scenario coordinator retains most of its advantage when transferred. Third, within the same training budget, a coordinator trained on both scenarios concurrently matches the scenario-specific coordinators in their own scenarios and significantly outperforms every other policy. Sensitivity analyses on ignition location, weather, scenario, and training curricula show that the variability represented during training shapes post-deployment performance. This proof of concept shows DE-enabled mission simulations combined with RL offer a reusable methodology for improving ME practice, which can be extended to fleet design and selection from a mission-first perspective.
Abstract (translated)
URL
https://arxiv.org/abs/2610.28818
PDF
-
ArcFlow: Training Flow Policies for Multiple Solver Budgets
2026-10-20 18:18:42Hieu Tran, Hong YuAbstract
Flow-matching policies adjust inference cost through the number of solver steps, but reinforcement learning at a single fixed solver budget can hide degradation at cheaper deployment budgets and even erase useful one-step behavior present in a pretrained policy. We introduce Action-Ratio-Clip Flow (ArcFlow), a deployment-aware algorithm for training one flow policy across a specified range of solver budgets. ArcFlow combines action-aligned supervision, budget exposure, and worst-budget checkpoint selection. For reconstruction objectives, its bridge aggregates temporal probes into one surrogate ratio and one PPO-style clip per action, with a boundary residual exactly equal to deployed one-step reconstruction error. Cycling exposes the shared field to the deployment range, and selection optimizes its performance floor. For native velocity objectives, multi-budget endpoint supervision applies the same principle without replacing the native update. Both realizations preserve the architecture and deployment-time cost. Across DM Control, manipulation, and locomotion, ArcFlow attains the highest mean one-step return among five flow-policy baselines on 25 of 30 tasks under each method's reported training settings. Component studies identify complementary improvements from boundary supervision, the fixed-grid bridge, budget exposure, and deployment selection. These results establish deployment-budget-aware training as an effective route to strong low-step control with one policy serving multiple inference costs.
Abstract (translated)
URL
https://arxiv.org/abs/2610.28788
PDF
