RL Fine-Tuning

When reinforcement learning teaches a language model a new behavior, can we find the training rollouts that taught it? And when an attribution method says it can, how do we know the answer is real? We study both questions on online RL fine-tuning with GRPO, using a planted behavior with a known cause. We release BehaviorTrace, an open evaluation harness that combines full-gradient sketching, the planted-behavior setup, and controls for gradient magnitude, fluency, headroom, and variation across seeds and generation draws. Across three seeds on Qwen2.5-1.5B, much of the apparent attribution signal comes from confounds. A control that ranks training steps by gradient size alone, with no behavior target, reaches 4.2 to 4.5 times chance and matches or beats the best targeted estimator on two of three seeds. At saturated checkpoints, model fluency predicts the behavior label at least as well as every gradient method we compared it with. Once fluency is controlled, the per-rollout results change from seed to seed and from one generation draw to the next, so a single run cannot settle the question. One signal does hold on all three seeds. The gradient of the trigger tokens aligns with a target built where the behavior actually occurs. We turn these findings into a checklist for evaluating attribution in RL. We test existing estimators, including GAS (renormalized TracInCP) and a TRAK-style estimator, and do not propose a new one.

Reinforcement learning (RL) fine-tuning improves vision-language-action (VLA) policies through closed-loop experience, yet generalization beyond the fine-tuning distribution remains limited. Our analysis reveals a selective reshaping of exploration: RL contracts behavior globally, yet diversifies successful trajectories, elicits success with fewer rollouts, and covers more of the latent task-valid solution space than supervised fine-tuning. Broader successful-mode coverage may provide alternative strategies under distribution shifts. Inspired by this, we introduce DRIVE (Diversity-driven RL fIne-tuning for VLA gEneralization), which turns successful-behavior diversity into an explicit RL objective. DRIVE groups rollouts under matched task conditions, compares their trajectories with temporal alignment, and derives a success-conditioned intrinsic reward from relative behavioral diversity. This design encourages broader coverage of feasible solutions without rewarding diverse failures or superficial timing differences. Across LIBERO-Plus, ManiSkill3, and RoboTwin 2.0, DRIVE improves the average out-of-domain (OOD) performance over vanilla RL fine-tuning by 5.3 points on $π_0$ and 2.0 points on $π_{0.5}$. On a dual-arm AgileX PiPER-X platform, DRIVE further increases average OOD success from 64.1% to 73.3% (+9.2 points), demonstrating gains that persist under physical deployment.

Multi-turn LLM agents often receive sparse task feedback across several interactions, while generating each response token by token. This creates two related credit-assignment questions: which responses helped achieve the outcome, and which generation decisions mattered within each response? Existing methods typically focus on only one level: turn-level methods evaluate complete responses but do not distinguish the decisions within them; token-level methods can propagate feedback across turns but do not explicitly model credit for each response. These complementary limitations motivate learning credit at both levels and coordinating it in a single policy update. We introduce VETTA, a credit assignment method that jointly learns turn- and token-level values through separate heads on a shared lightweight critic. VETTA computes advantages along both temporal sequences and combines each turn advantage with a within-response-centered token residual for PPO updates. Furthermore, to reduce value-learning cost, the critic retains only early Transformer blocks from the pretrained checkpoint used to initialize the actor. On two challenging agent benchmarks, ALFWorld and WebShop, VETTA improves success rates over PPO by 37.5% and 22.3%, respectively, with Qwen2.5-1.5B-Instruct and achieves success rates of 95.5% and 76.0%, respectively, with Qwen2.5-7B-Instruct. Critic-depth comparisons further show strong task performance with substantially lower critic-side computation. These results suggest that a compact shared critic can coordinate turn- and token-level credit to improve agent performance while keeping value estimation efficient. Code is available at https://github.com/Jiaju-Chen/VETTA-official.

Roadside traffic reasoning requires every free-form textual claim to be backed by visual evidence. Existing grounded multimodal large language models (MLLMs) frequently exhibit say-point mismatch, in which the textual answer contradicts the bounding boxes the model localizes. Evaluation metrics that score answers and boxes separately leave this failure unpenalized. We trace the mismatch to the conventional answer-then-ground factorization, which commits to a numerical claim before any object is enumerated. To measure it, we build RoadSceneVQA-G, a benchmark of 34.7K question-answer pairs in which every free-form answer is linked to the set of boxes that witnesses it, and we propose the Answer-Grounding Consistency (AGC) evaluation suite. To address it, we introduce Enumerate-then-Answer (EtA), which reverses the generation order so that answer-evidence agreement becomes a property of the output structure, and Enumeration-Consistent Policy Optimization (ECPO), a reinforcement learning stage that uses the union of multiple rollouts as a recall teacher without ground-truth boxes. EtA raises say-point consistency from 26.6% to 93.7% and grounding F1 from 52.2% to 73.0%, and ECPO further increases F1 to 75.6% without per-box supervision. On gRefCOCO, the same framework outperforms the strongest compared method, indicating that it transfers beyond traffic scenes. The project is available at https://github.com/GuanRunwei/RoadSceneVQA-G.

Reinforcement learning enables large language model (LLM) agents to learn multi-step behaviors through interaction with their environments. However, rewards in many interactive tasks reflect only the final outcome, providing limited guidance on which intermediate decisions advance the task. Successful training trajectories contain intermediate states that can provide supervision for subsequent interactions. We introduce Trajectory-to-Step Policy Optimization (T2SPO), a method that uses past interaction trajectories to provide step-level feedback for policy learning. T2SPO derives remaining-distance targets from successful trajectories and pairs them with representations of the states visited along the way. Conditioned on these examples, a pretrained TabPFN regressor estimates the remaining distance to success at each state of a new rollout. Changes in this distance estimate across consecutive states yield auxiliary credit for agent steps alongside task-level supervision. As training proceeds, newly completed trajectories refresh the estimator's context, incorporating new experience without updating its parameters. Experiments with 1.5B and 7B language models on ALFWorld and WebShop show that T2SPO consistently improves overall task success over GRPO.

Group Relative Policy Optimization (GRPO) has emerged as a memory-efficient reinforcement fine-tuning (RFT) technique for reasoning-intensive tasks. How- ever, GRPO training dynamics on small language models (SLMs) remain poorly understood, limiting its reliable adoption and reproducibility in open and resource- constrained environments. In this work, we present a systematic study of GRPO fine-tuning for SLMs ranging from 1.5B to 7B parameters under a practical single- node 8xA100 compute budget. Our study spans multiple model families and reasoning domains, including mathematics, coding, and multiple-choice question answering (MCQ) in science. Across these settings, we analyze how group size affects policy convergence, training stability, and downstream benchmark per- formance. We further characterize tensor-level update dynamics during GRPO training and investigate whether the choice of LoRA target modules and layers can improve the performance of GRPO-tuned models. While our initial GRPO-tuned models outperform their base counterparts on approximately 80% of mathematical benchmark evaluations, they demonstrate limited capability on MCQ and code reasoning tasks. Guided by our mechanistic evaluations, we refined our LoRA and reward-shaping configurations to improve performance in latter domains. These findings provide practical guidance for GRPO training for SLMs.

We study reinforcement learning (RL) methods for cooperative multi-agent Vision-Language-Action (VLA) models. This problem is challenging because VLAs are pretrained on large-scale single-agent data and therefore lack the fine-grained coordination skills required for inter-robot collaboration. Supervised fine-tuning (SFT) on multi-robot demonstrations partially bridges this gap, but its performance is bounded by the demonstration data and cannot improve from its own experience. We present a three-stage reinforced fine-tuning (RFT) pipeline for multi-agent VLAs. First, initialization-aware data collection sweeps over initial configurations and invokes human demonstrations only when the pretrained VLA repeatedly fails, yielding robustness to initialization shift with reduced human cost. Second, offline credit-filtered tuning assigns credit to individual agents and fine-tunes on per-agent trajectories with positive advantage rather than on entire joint rollouts. Third, we find existing online RL for VLAs are less effective for hard multi-agent tasks, which we attribute to noisy co-exploration and unstable updates. We instead use online latent-space fine tuning, which freeze the VLA and perform RL in its latent noise space. We evaluate our multi-agent VLA with both $π_0$ and $π_{0.5}$ backbones across 11 tasks in RoboTwin, RoboFactory and real-world manipulation with two Franka robots. Our multi-agent VLA improves the average success rate by $+23.1\%$, $+16.4\%$, and $+44\%$ on RoboTwin, RoboFactory, and real-world tasks, respectively. Code available at https://anonymous.4open.science/r/mavla_rft-2BC0/.

Domain experts trained from a shared checkpoint can transfer their specialized capabilities to a single student through on-policy distillation (OPD). Existing research primarily focuses on improving this merging process, while the algorithms used to train the experts have received limited systematic comparison. We investigate which training algorithm produces teachers better suited to OPD through controlled single-teacher comparisons of supervised fine-tuning (SFT) and reinforcement learning (RL) across Agentic, Reasoning, and Perception. Teachers and students share the same Qwen3.5-9B initialization, and the two teacher types are compared at similar task performance. Our experiments show that RL teachers yield stronger students and higher recovery of teacher performance gains across all three domains. At their best checkpoints, RL-guided students outperform SFT-guided students by 4.27, 1.50, and 0.86 percentage points, respectively. In Agentic, the best SFT-guided student recovers only 44.44% of its teacher's performance gain over the base model, whereas the best RL-guided student recovers 115.00%, surpassing its teacher. Our further analysis shows that RL teachers undergo smaller parameter displacements from the shared initialization than SFT teachers. These findings support the hypothesis that RL teachers' smaller departures from the student's starting point facilitate learning through OPD, resulting in stronger students.

Online reinforcement learning fine-tuning of pretrained flow-matching vision-language-action (VLA) policies promises robots that keep learning after deployment, but continued updates often destroy competence on individual tasks while the aggregate still looks healthy. We study this failure mode, which we call task collapse, under a matched small-compute budget on LIBERO-10 with a 450M-parameter SmolVLA policy trained by PPO with stochastic (SDE) sampling. Three exploration-noise policies differ in one live variable: a fixed noise scale, a ReinFlow-style learned noise network, and an uncertainty-gated controller that redistributes exploration across task streams from task-agnostic novelty and competence signals, without task labels or episode boundaries. Under the pooled definition, fixed noise collapses tasks in two of three seeds and learned noise in every seed measured to iteration 200, while the controller collapses none in any of its three seeds. Measured parameter displacement shows the controller's action expert keeps changing, while its mean applied noise is close to the fixed scale in the available logs. The matched comparison supports the controller's effect on task preservation; the separate contributions of its adaptation across states and over time are not disentangled. A lower fixed scale slows the decline but does not stop it. No arm improves on the behavior-cloning baseline in this budget. Two properties of that regime are measured beside this result, not offered as its cause: following the reference recipe, training runs in bfloat16 with no fp32 master copy, under which 96.02% of the action expert's elements stay bit-identical across three consecutive iterations, and an fp32 master copy at the reference learning rate collapses both arms in a single-seed observation. We release tools measuring per-task collapse under four definitions, rescoring noise and instrument tares.

Reward fine-tuning aims to update a pre-trained flow-based generative model to improve the downstream reward of its generated samples. Existing methods typically formulate this problem as sampling from a reward-tilted distribution, the solution to a KL-regularized reward-maximization problem. Here, we introduce an optimal transport regularizer built directly from the pre-trained drift. Unlike KL reward tilting, the resulting objective transports individual samples toward higher reward rather than reweighting the base distribution. We show that the resulting problem is equivalent to a deterministic optimal control problem on the flow. Given a pre-trained flow map, this equivalence yields a simulation-free reinforcement learning algorithm for fine-tuning generative flows. We call the resulting framework Wasserstein-Tilted Flow Maps (WTF), the first end-to-end fine-tuning recipe native to flow maps. The output is a fine-tuned flow map that retains strong reward-aligned performance at few-step inference budgets without post-hoc distillation. Experiments on ImageNet-256 and text-to-image show that WTF achieves higher reward with comparable or higher diversity than baselines, while requiring up to $280\times$ less training compute. More broadly, we argue that accelerated samplers such as flow maps are essential infrastructure for efficient post-training, and that the dominant KL-regularized formulation is only one of many choices worth revisiting.

A pretrained robot foundation policy may execute most of a long-horizon task yet repeatedly fail at a few critical subtasks. Collecting additional full-task demonstrations for supervised fine-tuning (SFT) requires operators to repeat behaviors the policy already performs well. Reinforcement learning (RL) fine-tuning offers a promising path to bridge this gap, but existing approaches struggle to solve long-horizon tasks using only sparse rewards. We present PARTS (Policy Adaptation with RL on Targeted Subtasks), a real-world subtask RL framework that concentrates practice at these bottlenecks while allowing training rollouts to proceed with minimal human intervention. The frozen pretrained policy supplies nominal actions throughout execution, while agent-generated selectors and success verifiers activate residual corrections and provide local outcome rewards. These rewards support learning from successful subtasks even when complete-task successes are scarce. Training combines online RL with success-reweighted retraining, and each retrained residual policy is redeployed to collect further experience. Humans identify bottlenecks during setup and perform physical resets when needed. On bimanual YAM and single-arm Franka tasks, PARTS improves complete-task success from 32% to 61% and from 50% to 95%, respectively, using tens of minutes of real-world RL rollouts per task on average. Compared with existing real-world RL fine-tuning methods, PARTS raises full-task success by more than 25% under the same robot-rollout budget while requiring less human involvement.

As scaling pre-training data alone yields diminishing returns, post-training is becoming increasingly important across physical AI domains such as autonomous driving. End-to-end driving policies are pre-trained in open loop with behavior cloning on human demonstrations. However, compounding errors during closed-loop deployment can take the vehicle outside the training data distribution, increasing the risk of safety-critical incidents. Closed-loop post-training can mitigate this risk but requires costly simulation for sensor-based policies. We propose OPTED (on-policy fine-tuning for end-to-end driving) which decouples reinforcement learning from the post-training of the end-to-end policy: a privileged teacher is trained using RL on vectorized inputs (HD-map and bounding boxes). This teacher then provides supervision to the pre-trained student during closed-loop post-training. We apply OPTED to two camera-based models, TransFuser and VaVAM, and fine-tune them in AlpaSim, using neural reconstructions (3DGS) of real driving logs. Driving scores increase by factors of 1.6$\times$ and 9.5$\times$, respectively. In controlled experiments OPTED matches closed-loop performance with approximately three orders of magnitude fewer simulator interactions than direct RL post-training, while staying closer to the human prior. Project page: https://01dami23.github.io/opted/

Agent trajectories record what an agent does and what happens next. Yet standard supervised fine-tuning (SFT) applies loss only to agent-authored action tokens, using environment observations as context but not as prediction targets. We ask whether this convention provides the best initialization for subsequent reinforcement learning. We introduce ActObs, which also supervises the observation tokens already present in each trajectory. Although deployed agents never generate observations, learning to predict them encourages the policy to model action consequences without adding data, parameters, sequence tokens, or forward passes. The methods perform similarly after SFT but diverge after GRPO. On Qwen3-4B, GRPO from ActObs achieves higher pass@k at every evaluated sampling budget than its action-only counterpart on Terminal-Bench 2.0. On Qwen3-8B, it trades some pass@1 reliability for higher pass@k (+3.4 pp at pass@16) and solves more distinct tasks. The advantage extends to cross-domain code editing on aider-polyglot (+4.2 pp at pass@1 at 4B), whose tasks are unseen during SFT and RL. ActObs retains more entropy during RL while requiring less policy movement, leaving the final policy closer to its SFT initialization. Our analysis traces this difference to SFT: action and observation gradients rapidly become orthogonal, while action-only training leaves a large residual observation gradient and degrades environment prediction below the base model. Joint supervision prevents this one-sided specialization, preserving consequence prediction and preparing the policy for downstream exploration.

Predicting cellular responses to perturbations supports the study of gene function, disease mechanisms, and therapeutic strategies. Despite advances in single-cell perturbation modeling, existing models typically optimize surrogate losses that do not directly reflect the biological criteria used for evaluation, so better data fitting need not yield better biological predictions. To address this mismatch, we introduce \textbf{CellRFT}, a reinforcement fine-tuning framework that uses biological evaluation as direct training feedback. CellRFT uses policy-gradient optimization to learn from non-differentiable evaluations of generated cell populations and integrates multiple biological rewards through hierarchical reward aggregation. Comprehensive experiments demonstrate CellRFT's applicability across different pretrained models and effectiveness in improving perturbation prediction, reveal that optimizing one biological criterion can help or hinder others, and show that complementary rewards can improve criteria beyond those directly optimized, offering a way to probe how biological metrics shape model behavior, with the potential to inform evaluation design. Code will be made available.

Training embodied foundation models typically requires massive-scale datasets and extensive computational resources, yet often suffers from three critical limitations: (1) inefficient sample utilization due to low-informative samples; (2) imbalanced gradient contributions across heterogeneous tasks; and (3) severe credit assignment problem in long-horizon planning, where trajectory-level rewards indiscriminately penalize all tokens. To address these issues, we propose an efficient training paradigm that achieves state-of-the-art average performance through strategic data selection and hierarchical policy optimization. Our approach consists of three synergistic stages. First, Rejection Sampling-based Fine-Tuning (RSFT) filters out low-informative samples to establish robust behavioral priors while preventing distributional collapse. Second, Iterative Rejection GRPO (IR-GRPO) employs task-specific queues stratified by difficulty to keep datasets balanced across reinforcement learning iterations, coupled with a hybrid reward mechanism for precise cross-task feedback. Third, to enhance long-horizon task planning, we introduce Trie-GRPO, a novel reinforcement learning algorithm based on action prefix trees, which enables step-level advantage estimation. This resolves the credit assignment problem by isolating intermediate correct decisions from downstream errors, while effectively balancing exploration efficiency and depth compared to conventional search trees. As a result, EmbodiedMind achieves a state-of-the-art average performance of 70.02% across 18 benchmarks, and significantly outperforms other embodied foundation models in long-horizon task planning accuracy. Our project will be released for reproducibility.

Reinforcement learning fine-tuning on top of large, pretrained Vision-Language-Action (VLA) models offers promise for highly reliable robot deployment. However, because of their scale, modern VLA models suffer from high inference latency, so the observation used to select an action is often stale by execution time, creating a distribution shift that can substantially degrade reliability and performance. Prior work has explored asynchronous policy execution to reduce the effect of latency, but these methods are mostly built on imitation learning and offer no mechanism for moving beyond the training distribution toward higher reliability. We close this gap by enabling RL fine-tuning that meets the real-time control requirements of dynamic real-world manipulation. Our approach builds on EXPO-FT, a framework for sample-efficient, reliable VLA fine-tuning with reinforcement learning, and decouples slow, expressive action generation from fast, reactive action edits: a large pretrained VLA proposes action chunks using its strong behavior prior, while a lightweight edit policy performs fast, reactive decision-making by editing actions in response to changes in state, conditioned on the latest observation. We instantiate this as Real-Time EXPO-FT, an RL framework for finetuning real-time VLA policies. On the Kinetix benchmark, Real-Time EXPO-FT enables a delayed policy to achieve the best performance among delayed and non-delayed methods in 10 out of 10 environments. On four dynamic real-world tasks, robot object passing, ball balancing, table soccer kicking, and dynamic object picking, with online robot data capped at 10 minutes, Real-Time EXPO-FT improves average policy performance from 42% to 97%, all without human intervention, demonstrating rapid, sample-efficient adaptation to challenging real-world dynamics. Website: https://pd-perry.github.io/real-time-expo-ft

Text summarization requires models to condense content while preserving key qualities such as consistency and coherence. Large language models (LLMs) have shown strong performance on this task and can be further improved through reinforcement learning (RL). However, most existing methods apply reward signals directly to undifferentiated token sequences, overlooking the varying importance of individual tokens to word and sentence level quality in summarization. In this paper, we propose Token Importance-Aware Policy Optimization (TIAO), a novel reinforcement learning strategy that explicitly leverages token-importance awareness. Specifically, TIAO identifies core tokens based on token dependency and reweights a trajectory's advantage according to its overall dependencies. Experiments on the real world dataset show that our TIAO achieves highly competitive results, and that a 7B foundation model enhanced by TIAO performs comparably to GPT-4 and GPT-5-nano. Code is available at https://github.com/TechCloud-x/TIAO

Electronic health record (EHR) foundation models trained on longitudinal patient trajectories have demonstrated strong performance across diverse clinical prediction tasks. However, their clinical reasoning capabilities remain constrained by next-token prediction on limited and incomplete EHR data. To address this, we propose a reinforcement learning (RL) fine-tuning framework that treats EHR foundation models as generative policies over patient trajectories. We formulate common clinical prediction problems (e.g., hospital readmission) as event-conditioned, time-windowed reasoning tasks. We then design time-aware, rollout-sensitive rewards to account for finite rollout lengths and temporally inconclusive outcomes. We find that RL fine-tuning consistently improves over pre-trained backbones and strong baselines. Notably, it enables smaller models to surpass larger pre-trained models in data-limited regimes and induces positive transfer across tasks. Further analysis shows that RL fine-tuned models generate trajectories with stronger structural and semantic alignment to ground truth and greater downstream utility.

Large Language Models demonstrate remarkable proficiency in static reasoning, yet training them as autonomous agents through Reinforcement Learning (RL) for long-horizon tasks is often hindered by severe reward sparsity. While conventional \textit{agent-side warming} up via supervised fine-tuning (SFT) can alleviate this, it is frequently limited by data scarcity and constrained exploration. To address this, we propose a paradigm shift to \textit{environment-side adaptation} by constructing \textbf{F}eedback-\textbf{E}nriched \textbf{E}nvironments (\textbf{FEEs}). Through a pilot study, we establish a feedback design strategy that reformulates environments by transitioning from action guidance to observation enrichment during the later stages of both intra-episode exploration and inter-episode evolution. Large-scale experiments on SciWorld and BFCL benchmarks using various Qwen3 model scales and RL algorithms such as GRPO, GSPO, and DAPO demonstrate that FEEs consistently yield performance improvements over standard settings. Furthermore, our analysis reveals that training with FEEs \textbf{(1)} stabilizes training dynamics by reducing entropy volatility, \textbf{(2)} facilitates proactive state-space exploration in difficult tasks, \textbf{(3) }ensures the internalization of environmental guidance into policy weights rather than acting as a mere inference-time prior, and \textbf{(4) }identifies intra-group feedback consistency as a critical boundary for stable optimization.

Large Language Models (LLMs) are increasingly deployed in settings where rare but severe harmful generations can have significant consequences. Existing Risk-Averse RLHF addresses this issue by optimizing Conditional Value-at-Risk (CVaR), but it trains policies for fixed risk levels and therefore cannot adjust the desired degree of risk aversion at inference time. In this paper, we propose risk-conditioned RLHF, a framework that trains a single policy that provides a continuous risk-control interface, enabling users to select different degrees of risk aversion without retraining or deploying multiple risk-specific models. Experiments across multiple benchmarks demonstrate that a single risk-conditioned policy can adapt to different risk levels at inference time, enabling more flexible and risk-aware LLM deployment.

While Reinforcement Learning (RL) effectively incentivizes reasoning in Large Language Models, current pipelines are hindered by training instability and rapid entropy collapse. These limitations often stem from "Rollout Silencing" and low-quality gradient signals in standard sampling procedures. In this work, we propose a robust, data-centric framework to stabilize RL training. We first introduce Potential-Aware Query Mining (PAQM), which filters data dynamically to focus on the "Distillation Zone"---samples with high potential for capability elicitation. Furthermore, we present Hybrid Stratified Replay (HSR), a novel mechanism that restructures batches by stratifying rollouts based on Path Entropy, a rollout-level confidence proxy, and outcome reward. Within each optimization step, HSR reuses current-policy "Stability Anchors" and "Hard Negatives" to construct high-contrast optimization groups, then clears its buffers before the next step. This approach mitigates entropy collapse while improving the utilization of learning signals under limited compute. Our method outperforms strong baselines on complex reasoning tasks, offering a principled solution for stable and efficient RL fine-tuning.

How to divide a fixed annotation budget between supervised fine-tuning (SFT) and reinforcement learning (RL) during LLM post-training remains an open problem. Existing work characterizes only broad trends (e.g., SFT dominates in low-data regimes), lacks a principled allocation framework, and does not examine whether the optimal ratio transfers across model sizes. We frame this problem in terms of near-optimality: rather than seeking a single optimal SFT-RL ratio, we characterize the near-optimal region, the set of allocations within a specified tolerance of peak performance. Empirically, this region is wide even for small tolerances (2-10%), widens with model scale, and transfers reliably from small proxy models to large target models. This yields a practical strategy: small proxy-model experiments suffice to identify a transferable near-optimal region, eliminating the need for exhaustive large-scale search. Our results hold consistently across tasks, model families, and both preference-based off-policy and reward-supervision on-policy RL methods. We further analyze how the asymmetry in annotation costs between SFT and RL data shifts the near-optimal region.

Reinforcement learning fine-tuning of large language models increasingly adopts multiple reward dimensions, including verifiable rules, task-specific evaluators, and learned reward models, to provide richer supervision across diverse capabilities. These dimensions are commonly scalarized with fixed aggregation weights. We identify a failure mode in which aggregation itself induces reward hacking: static projection aliases qualitatively different reward profiles into a single scalar, steering optimization toward whichever dimensions are easiest, densest, or systematically favored by the reward signal. Over training, this traps the policy in suboptimal profiles and prevents convergence to better-balanced ones that would yield higher task performance. To address this, we propose Adaptive Multi-Reward Projection (AMRP), a lightweight online method that reallocates aggregation weights using three signals, relative shortfall, reward volatility, and recent progress, increasing pressure on lagging, unstable, or stagnant dimensions while relieving saturated ones. Across structured reasoning, citation-grounded generation, and open-ended alignment under GRPO, AMRP consistently improves reward-profile balance and downstream performance over fixed and dynamic weighting baselines; it also remains effective with GDPO and PPO, supporting compatibility across RL algorithms. Our code is available at https://github.com/yyhappier/AMRP.git.

In large-scale e-commerce retrieval, dual-encoder retrievers are op- timized for contrastive similarity, whereas downstream rerankers capture finer-grained relevance preferences; this objective mis- match limits end-to-end retrieval quality. Reinforcement Learning offers a way to use reward-model feedback for retriever adaptation, but we observe that standard policy-gradient updates can degrade embedding geometry, especially when the document index must remain frozen due to industrial constraints. To address this, we propose PAO (Positive-Advantage-Only), a selective RL optimization method. Our analysis reveals that in- discriminate penalization of negative samples (pushing away) in a frozen high-dimensional space disrupts pre-trained semantic man- ifolds. PAO selectively applies gradient updates only to retrieved items with positive advantages, effectively pulling query embed- dings toward high-reward regions while preserving global topo- logical stability. Experiments on both a massive industrial dataset and public benchmarks demonstrate that PAO significantly outper- forms standard RL and distillation baselines.

To mitigate the time complexity of generative models, one-step generative models have recently emerged through direct mapping from noise to data in a single forward pass. However, the reward-guided fine-tuning method of one-step generative models remains largely unexplored. To address this, we consider one-step generators from an optimal transport view, investigating Wasserstein Gradient Flow (WGF) for modeling smooth and controlled distributional evolution in probability space. We then propose a novel reward-guided fine-tuning of a one-step generative model via WGF. We derive a practical training method that requires no reward gradients, thereby handling both non-differentiable and differentiable rewards. Moreover, our method provides smooth and stable reward-guided distributional updates while mitigating reward hacking and mode collapse. Experiments on 2D synthetic data, CIFAR-10, and ImageNet 256$\times$256 with diverse rewards, including JPEG (in)compressibility, class probability, Black-and-White and CLIP alignment, show that our method achieves better reward alignment compared to baselines.

Offline-to-online (O2O) reinforcement learning aims to leverage policies pretrained on static datasets while improving them through online interaction. However, directly reusing an offline-trained critic can hinder online fine-tuning: as the policy and data distribution change rapidly, value estimates inherited from offline training may become misaligned with the online environment, leading to inaccurate policy improvement and inefficient exploration. To address this problem, we introduce Critic-Free Pretraining: an efficient paradigm that completely abandons the approach of offline critic training, allowing a freshly initialized critic to adapt without inheriting biased estimates. CFP is compatible with various mainstream O2O algorithms and consistently matches or improves upon conventional O2O algorithms across a diverse set of tasks, with particularly pronounced gains on several challenging tasks.

Foundation model(FM) for recommendation has shown strong ability to model long-horizon sequential user behavior. In practice, a single pretrained foundation model is often adapted to diverse downstream serving surfaces through Supervised Fine-Tuning(SFT). However, optimizing task-specific objectives such as clicks or likes does not necessarily align the serving policy with the business metrics that determine recommendation quality. We propose a three-phase progressive post-training framework that explicitly separates downstream adaptation from business-metric alignment. The adaptation stage is decomposed into Linear Probing(LP) and Full Fine-Tuning(FFT): LP first stabilizes randomly initialized downstream heads within a frozen pretrained representation space, and FFT then jointly specializes the full model for the target task. On top of this stabilized policy, Reinforcement Fine-Tuning(RFT) aligns the model with practical business objectives using a learned reward model. Rather than directly optimizing the serving policy on sparse business targets, we train the policy on dense implicit feedback and use business-metric supervision only for reward modeling. Offline experiments show that the progressive LP-FFT-RFT framework outperforms single-phase alternatives, and that reward-based alignment yields a stronger serving policy than directly using the reward model itself for ranking. Large-scale online A/B tests further show that the proposed framework improves production recommendation quality over a conventional non-foundation baseline. A reference implementation is available at https://github.com/webtoon/rec-fm-progressive-alignment

Tool-Integrated Reasoning (TIR) enables LLMs to solve complex tasks through iterative tool interactions. However, existing reinforcement learning methods often rely on trajectory-level supervision, limiting fine-grained credit assignment in long-horizon TIR scenarios. On-policy self-distillation offers denser signals through teacher branches with privileged context, but existing approaches typically derive such context from ground-truth answers or retrieved skills, which may not reflect the states actually visited by the agent. Moreover, token-level supervision fails to capture the turn-level structure of tool interactions. To address this, we propose TurnSight, a turn-level hindsight self-distillation framework that derives supervision directly from execution-conditioned hindsight. It then constructs multiple hindsight views with different lookahead horizons and selects reliable supervision through cross-horizon directional agreement. Finally, the selected hindsight signal is normalized across sibling rollouts and used to adaptively modulate RL advantages while preserving their original optimization direction. Extensive experiments on three benchmarks demonstrate the effectiveness of TurnSight. Our codes are available at https://github.com/quchangle1/TurnSight.

Designing effective reward functions remains a major bottleneck in Reinforcement Learning (RL). Recent work uses large foundation Vision-Language Models (VLMs) as reward models, computing text-observation similarity to bypass manual reward engineering. Although promising, these rewards are often noisy and unreliable, limiting their direct utility during deployment. We present Structure-Aware Fine-Tuning (SAFT), a simple, self-supervised method that refines these imperfect reward signals online without access to ground-truth supervision. SAFT leverages intrinsic structural priors to regularize the VLM's latent space via LoRA adapters. We rigorously evaluate SAFT across a spectrum of base model capabilities to demonstrate its versatility. Our results show that SAFT consistently denoises the reward landscape, yielding faster policy convergence and substantially improved alignment (EPIC distance) relative to the underlying base model, suggesting that failures can often be attributed to structural brittleness rather than semantic misunderstanding. By replacing extensive human preference annotation with structural inductive biases inherent to the task, SAFT offers a scalable path for stabilizing text-conditioned RL and underscores the broader value of incorporating task structure as a general inductive bias.

Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) exhibit fundamentally different behaviors in enhancing multi-task reasoning for large language models (LLMs). Our preliminary experiments revealed a phenomenon: SFT suffers from severe task conflicts under multi-stage training, whereas RL enables stable coexistence across diverse tasks. Empirically, we trace this to the parameter level, observing that RL induces sparse and approximately orthogonal updates across tasks. We provide a theoretical explanation for this mechanism by analyzing multi-task gradient interference. Our results reveal a distinction: interference in SFT is norm-limited, scaling with the absolute gradient magnitude, whereas interference in RL is variance-limited, bounded by the gradient variance induced by advantage normalization and on-policy optimization. This small variance bound yields near-orthogonal optimization directions across tasks. Leveraging this insight, we propose Parallel-RL, a paradigm that decouples multi-task training, significantly improving efficiency and flexibility.