RL Fine-Tuning

Modern large language models (LLMs) are increasingly fine-tuned via reinforcement learning from human feedback (RLHF) or related reward optimisation schemes. While such procedures improve perceived helpfulness, we investigate whether sycophantic reward signals degrade calibration -- a property essential for reliable uncertainty quantification. We fine-tune Qwen3-8B under three regimes: no fine-tuning (base), neutral supervised fine-tuning (SFT) on TriviaQA, and sycophancy-inducing Group Relative Policy Optimisation (GRPO) that rewards agreement with planted wrong answers. Evaluating on $1{,}000$ MMLU items across five subject domains with bootstrap confidence intervals and permutation testing, we find that sycophantic GRPO produces consistent directional calibration degradation -- ECE rises by $+0.006$ relative to the base model and MCE increases by $+0.010$ relative to neutral SFT -- though the effect does not reach statistical significance ($p = 0.41$) at this training budget. Post-hoc matrix scaling applied to all three models reduces ECE by $40$--$64\%$ and improves accuracy by $1.5$--$3.0$ percentage points. However, the sycophantic model retains the highest post-scaling ECE relative to the neutral SFT control ($0.042$ vs.\ $0.037$), suggesting that reward-induced miscalibration leaves a structured residual even after affine correction. These findings establish a methodology for evaluating the calibration impact of reward hacking and motivate calibration-aware training objectives.

While Large Language Models (LLMs) have demonstrated significant potential in Tool-Integrated Reasoning (TIR), existing training paradigms face significant limitations: Zero-RL suffers from inefficient exploration and mode degradation due to a lack of prior guidance, while SFT-then-RL is limited by high data costs and capability plateaus caused by low-entropy collapse. To address these challenges, we propose E3-TIR (Enhanced Experience Exploitation), a warm-up paradigm for the early stages of agent training. Specifically, we formulate training as the dynamic integration of three experience types: Expert Prefixes, Expert Guided, and Self-Exploration. By executing diverse branching exploration around expert "anchors" and employing a mix policy optimization mechanism, we effectively mitigate distribution shifts and resolve optimization conflicts arising from shared prefixes. Our method dynamically adapts the model's knowledge boundaries, effectively balancing exploration diversity with training efficiency.Experimental results demonstrate that E3-TIR achieves a 6 performance improvement over traditional paradigms on tool-use tasks, while requiring less than 10 of the synthetic data. Furthermore, in terms of ROI, a comprehensive metric integrating performance, data cost, and training efficiency we achieve a 1.46x gain compared to baselines. Code is available at https://github.com/yuki-younai/E3-TIR.

Post-training paradigms for Large Language Models (LLMs), primarily Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), face a fundamental dilemma: SFT provides stability (low variance) but suffers from high fitting bias, while RL enables exploration (low bias) but grapples with high gradient variance. Existing unified optimization strategies often employ naive loss weighting, overlooking the statistical conflict between these distinct gradient signals. In this paper, we provide a rigorous theoretical analysis of this bias-variance trade-off and propose DYPO (Dynamic Policy Optimization), a unified framework designed to structurally mitigate this conflict. DYPO integrates three core components: (1) a Group Alignment Loss (GAL) that leverages intrinsic group dynamics to significantly reduce RL gradient variance; (2) a Multi-Teacher Distillation mechanism that corrects SFT fitting bias via diverse reasoning paths; and (3) a Dynamic Exploitation-Exploration Gating mechanism that adaptively arbitrates between stable SFT and exploratory RL based on reward feedback. Theoretical analysis confirms that DYPO linearly reduces fitting bias and minimizes overall variance. Extensive experiments demonstrate that DYPO significantly outperforms traditional sequential pipelines, achieving an average improvement of 4.8% on complex reasoning benchmarks and 13.3% on out-of-distribution tasks. Our code is publicly available at https://github.com/Tocci-Zhu/DYPO.

Large language model (LLM) reinforcement learning (RL) pipelines are often bottlenecked by rollout generation, making end-to-end training slow. Recent work mitigates this by running rollouts with quantization to accelerate decoding, which is the most expensive stage of the RL loop. However, these setups destabilize optimization by amplifying the training-inference gap: rollouts are operated at low precision, while learning updates are computed at full precision. To address this challenge, we propose QaRL (Rollout Alignment Quantization-Aware RL), which aligns training-side forward with the quantized rollout to minimize mismatch. We further identify a failure mode in quantized rollouts: long-form responses tend to produce repetitive, garbled tokens (error tokens). To mitigate these problems, we introduce TBPO (Trust-Band Policy Optimization), a sequence-level objective with dual clipping for negative samples, aimed at keeping updates within the trust region. On Qwen3-30B-A3B MoE for math problems, QaRL outperforms quantized-rollout training by +5.5 while improving stability and preserving low-bit throughput benefits.

With the rise of LLMs, there is an increasing need for intelligent recommendation assistants that can handle complex queries and provide personalized, reasoning-driven recommendations. LLM-based recommenders show potential but face challenges in multi-step reasoning, underscoring the need for reasoning-augmented systems. To address this gap, we propose ReRec, a novel reinforcement fine-tuning (RFT) framework designed to improve LLM reasoning in complex recommendation tasks. Our framework introduces three key components: (1) Dual-Graph Enhanced Reward Shaping, integrating recommendation metrics like NDCG@K with Query Alignment and Preference Alignment Scores to provide fine-grained reward signals for LLM optimization; (2) Reasoning-aware Advantage Estimation, which decomposes LLM outputs into reasoning segments and penalizes incorrect steps to enhance reasoning of recommendation; and (3) Online Curriculum Scheduler, dynamically assess query difficulty and organize training curriculum to ensure stable learning during RFT. Experiments demonstrate that ReRec outperforms state-of-the-art baselines and preserves core abilities like instruction-following and general knowledge. Our codes are available at https://github.com/jiani-huang/ReRec.

Generative reward models (GRMs) have emerged as a promising approach for aligning Large Language Models (LLMs) with human preferences by offering greater representational capacity and flexibility than traditional scalar reward models. However, GRMs face two major challenges: reliance on costly human-annotated data restricts scalability, and self-training approaches often suffer from instability and vulnerability to reward hacking. To address these issues, we propose ConsistRM, a self-training framework that enables effective and stable GRM training without human annotations. ConsistRM incorporates the Consistency-Aware Answer Reward, which produces reliable pseudo-labels with temporal consistency, thereby providing more stable model optimization. Moreover, the Consistency-Aware Critique Reward is introduced to assess semantic consistency across multiple critiques and allocates fine-grained and differentiated rewards. Experiments on five benchmark datasets across four base models demonstrate that ConsistRM outperforms vanilla Reinforcement Fine-Tuning (RFT) by an average of 1.5%. Further analysis shows that ConsistRM enhances output consistency and mitigates position bias caused by input order, highlighting the effectiveness of consistency-aware rewards in improving GRMs. Our implementation is available at https://github.com/yuliangCarmelo/ConsistRM.

We introduce Discrete flow Matching policy Optimization (DoMinO), a unified framework for Reinforcement Learning (RL) fine-tuning Discrete Flow Matching (DFM) models under a broad class of policy gradient methods. Our key idea is to view the DFM sampling procedure as a multi-step Markov Decision Process. This perspective provides a simple and transparent reformulation of fine-tuning reward maximization as a robust RL objective. Consequently, it not only preserves the original DFM samplers but also avoids biased auxiliary estimators and likelihood surrogates used by many prior RL fine-tuning methods. To prevent policy collapse, we also introduce new total-variation regularizers to keep the fine-tuned distribution close to the pretrained one. Theoretically, we establish an upper bound on the discretization error of DoMinO and tractable upper bounds for the regularizers. Experimentally, we evaluate DoMinO on regulatory DNA sequence design. DoMinO achieves stronger predicted enhancer activity and better sequence naturalness than the previous best reward-driven baselines. The regularization further improves alignment with the natural sequence distribution while preserving strong functional performance. These results establish DoMinO as an useful framework for controllable discrete sequence generation.

We present an innovative multi-stage optimization strategy combining reinforcement learning (RL) and supervised fine-tuning (SFT) to enhance the pedagogical knowledge of large language models (LLMs), as illustrated by EduQwen 32B-RL1, EduQwen 32B-SFT, and an optional third-stage model EduQwen 32B-SFT-RL2: (1) RL optimization that implements progressive difficulty training, focuses on challenging examples, and employs extended reasoning rollouts; (2) a subsequent SFT phase that leverages the RL-trained model to synthesize high-quality training data with difficulty-weighted sampling; and (3) an optional second round of RL optimization. EduQwen 32B-RL1, EduQwen 32B-SFT, and EduQwen 32B-SFT-RL2 are an application-driven family of open-source pedagogical LLMs built on a dense Qwen3-32B backbone. These models remarkably achieve high enough accuracy on the Cross-Domain Pedagogical Knowledge (CDPK) Benchmark to establish new state-of-the-art (SOTA) results across the interactive Pedagogy Benchmark Leaderboard and surpass significantly larger proprietary systems such as the previous benchmark leader Gemini-3 Pro. These dense 32-billion-parameter models demonstrate that domain-specialized optimization can transform mid-sized open-source LLMs into true pedagogical domain experts that outperform much larger general-purpose systems, while preserving the transparency, customizability, and cost-efficiency required for responsible educational AI deployment.

Reinforcement learning for LLMs is vulnerable to reward hacking, where models exploit shortcuts to maximize reward without solving the intended task. We systematically study this phenomenon in coding tasks using an environment-manipulation setting, where models can rewrite evaluator code to trivially pass tests without solving the task, as a controlled testbed. Across both studied models, we identify a reproducible three-phase rebound pattern: models first attempt to rewrite the evaluator but fail, as their rewrites embed test cases their own solutions cannot pass. They then temporarily retreat to legitimate solving. When legitimate reward remains scarce, they rebound into successful hacking with qualitatively different strategies. Using representation engineering, we extract concept directions for shortcut, deception, and evaluation awareness from domain-general contrastive pairs and find that the shortcut direction tracks hacking behavior most closely, making it an effective representational proxy for detection. Motivated by this finding, we propose Advantage Modification, which integrates shortcut concept scores into GRPO advantage computation to penalize hacking rollouts before policy updates. Because the penalty is internalized into the training signal rather than applied only at inference time, Advantage Modification provides more robust suppression of hacking compared with generation-time activation steering.

While large language models (LLMs) have demonstrated strong performance on complex reasoning tasks such as competitive programming (CP), existing methods predominantly focus on single-attempt settings, overlooking their capacity for iterative refinement. In this paper, we present RefineRL, a novel approach designed to unleash the self-refinement capabilities of LLMs for CP problem solving. RefineRL introduces two key innovations: (1) Skeptical-Agent, an iterative self-refinement agent equipped with local execution tools to validate generated solutions against public test cases of CP problems. This agent always remains skeptical of its own outputs and thereby enforces rigorous self-refinement even when validation suggests correctness. (2) A reinforcement learning (RL) solution to incentivize LLMs to self-refine with only standard RLVR data (i.e., problems paired with their verifiable answers). Extensive experiments on Qwen3-4B and Qwen3-4B-2507 demonstrate that our method yields substantial gains: after our RL training, these compact 4B models integrated with the Skeptical-Agent not only outperform much larger 32B models but also approach the single-attempt performance of 235B models. Notably, although our models are trained solely on CP coding data, RefineRL also substantially improves their performance on mathematical reasoning benchmarks, suggesting that the learned self-refinement behavior can transfer beyond coding-specific tasks. These findings highlight self-refinement as a promising direction for scaling LLMs' general reasoning ability.

Text-to-motion generation has advanced with diffusion- and flow-based generative models, yet supervised pre-training remains insufficient to align models with high-level objectives such as semantic consistency, realism, and human preference. Existing post-training methods have key limitations: they (1) target a specific motion representation, such as joints, (2) optimize a particular aspect, such as text-motion alignment, and may compromise other factors; and (3) incur substantial computational overhead, data dependence, and coarse-grained optimization. We present a reinforcement fine-tuning framework that comprises a heterogeneous-representation, multi-dimensional reward model, MotionReward, and an efficient, fine-grained fine-tuning method, EasyTune. To obtain a unified semantics representation, MotionReward maps heterogeneous motions into a shared semantic space anchored by text, enabling multi-dimensional reward learning; Self-refinement Preference Learning further enhances semantics without additional annotations. For efficient and effective fine-tuning, we identify the recursive gradient dependence across denoising steps as the key bottleneck, and propose EasyTune, which optimizes step-wise rather than over the full trajectory, yielding dense, fine-grained, and memory-efficient updates. Extensive experiments validate the effectiveness of our framework, achieving FID 0.132 at 22.10 GB peak memory for MLD model and saving up to 15.22 GB over DRaFT. It reduces FID by 22.9% on joint-based ACMDM, and achieves a 12.1% R-Precision gain and 21.7% FID improvement on rotation-based HY-Motion. Our project page with code is publicly available.

Understanding spatial dynamics and semantics in point cloud is fundamental for comprehensive 3D comprehension. While reinforcement learning algorithms such as Group Relative Policy Optimization (GRPO) have recently achieved remarkable breakthroughs in large language models by incentivizing reasoning capabilities through strategic reward design, their potential remains largely unexplored in the 3D perception domain. This naturally raises a pivotal question: Can RL-based methods effectively empower 3D point cloud fine-tuning? In this paper, we propose PointRFT, the first reinforcement fine-tuning paradigm tailored specifically for point cloud representation learning. We select three prevalent 3D foundation models and devise specialized accuracy reward and dispersion reward functions to stabilize training and mitigate distribution shifts. Through comprehensive few-shot classification experiments comparing distinct training paradigms, we demonstrate that PointRFT consistently outperforms vanilla supervised fine-tuning (SFT) across diverse benchmarks. Furthermore, when organically integrated into a hybrid Pretraining-SFT-RFT paradigm, the representational capacity of point cloud foundation models is substantially unleashed, achieving state-of-the-art performance particularly under data-scarce scenarios.

Improving data utilization efficiency is critical for scaling reinforcement learning (RL) for long-horizon tasks where generating trajectories is expensive. However, the dominant RL methods for LLMs are largely on-policy: they update each batch of data only once, discard it, and then collect fresh samples, resulting in poor sample efficiency. In this work, we explore an alternative value-based RL framework for LLMs that naturally enables off-policy learning. We propose ReVal, a Bellman-update-based method that combines stepwise signals capturing internal consistency with trajectory-level signals derived from outcome verification. ReVal naturally supports replay-buffer-based training, allowing efficient reuse of past trajectories. Experiments on standard mathematical reasoning benchmarks show that ReVal not only converges faster but also outperforms GRPO in final performance. On DeepSeek-R1-Distill-1.5B, ReVal improves training efficiency and achieves improvement of 2.7% in AIME24 and 4.5% in out-of-domain benchmark GPQA over GRPO. These results suggest that value-based RL is a practical alternative to policy-based methods for LLM training.

Reinforcement learning with verifiable rewards (RLVR) has significantly improved reasoning in large language models (LLMs), yet the token-level mechanisms underlying these improvements remain unclear. We present a systematic empirical study of RLVR's distributional effects organized around three main analyses: (1) token-level characterization of distributional shifts between base and RL models, (2) the impact of token-level distributional shifts on sequence-level reasoning performance through cross-sampling interventions, and (3) fine-grained mechanics of these shifts at the token level. We find that RL fine-tuning induces highly sparse and targeted changes, with only a small fraction of token distributions exhibiting meaningful divergence between the base and RL policies. We further characterize the structure and evolution of these shifts through analyses of token entropy, positional concentration, and reallocation of probability mass. To assess the functional importance of these sparse changes, we conduct cross-sampling experiments that selectively swap token choices between the base and RL models with varying intervention budgets. We show that inserting only a small fraction of RL-sampled tokens into base generations progressively recovers RL performance gains, while injecting a similarly small number of base token choices into otherwise RL-generated sequences collapses performance to base levels, isolating a small set of token-level decisions directly responsible for RLVR's performance gains. Finally, we explore divergence-weighted variants of the advantage signal as a diagnostic intervention, finding that they can yield improvements over baselines. Together, our results shed light on the distributional changes induced by RLVR and provide a fine-grained, token-level lens for understanding RLVR fine-tuning as a targeted refinement process.

This chapter presents three major reinforcement learning algorithms used for fine-tuning financial forecasters. We propose a clear implementation plan for backpropagating the loss of a reinforcement learning task to a model trained using supervised learning, and compare the performance before and after the fine-tuning. We find an increase in performance after fine-tuning, and transfer learning properties to the models, indicating the benefits of fine-tuning. We also highlight the tuning process and empirical results for future implementation by practitioners.

Chain-of-thought reasoning, where language models expend additional computation by producing thinking tokens prior to final responses, has driven significant advances in model capabilities. However, training these reasoning models is extremely costly in terms of both data and compute, as it involves collecting long traces of reasoning behavior from humans or synthetic generators and further post-training the model via reinforcement learning. Are these costs fundamental, or can they be reduced through better algorithmic design? We show that autocurriculum, where the model uses its own performance to decide which problems to focus training on, provably improves upon standard training recipes for both supervised fine-tuning (SFT) and reinforcement learning (RL). For SFT, we show that autocurriculum requires exponentially fewer reasoning demonstrations than non-adaptive fine-tuning, by focusing teacher supervision on prompts where the current model struggles. For RL fine-tuning, autocurriculum decouples the computational cost from the quality of the reference model, reducing the latter to a burn-in cost that is nearly independent of the target accuracy. These improvements arise purely from adaptive data selection, drawing on classical techniques from boosting and learning from counterexamples, and requiring no assumption on the distribution or difficulty of prompts.

Constraints are essential for stabilizing reinforcement learning fine-tuning (RFT) and preventing degenerate outputs, yet they inherently conflict with the optimization objective because stronger constraints limit the ability of a fine-tuned model to discover better solutions. We propose dynamic constraints that resolve this tension by adapting to the evolving capabilities of the fine-tuned model based on the insight that constraints should only intervene when degenerate outputs occur. We implement this by using a reference model as an online refiner that takes the response from the fine-tuned model and generates a minimally corrected version which preserves correct content verbatim while fixing errors. A supervised fine-tuning loss then trains the fine-tuned model to produce the refined output. This mechanism yields a constraint that automatically strengthens or relaxes based on output quality. Experiments on dialogue and code generation show that dynamic constraints outperform both KL regularization and unconstrained baselines, achieving substantially higher task rewards while maintaining training stability.

Sim-to-real transfer of locomotion policies often leads to performance degradation due to the inevitable sim-to-real gap. Naively fine-tuning these policies directly on hardware is problematic, as it poses risks of mechanical failure and suffers from high sample inefficiency. In this paper, we address the challenge of safely and efficiently fine-tuning reinforcement learning (RL) policies for dynamic locomotion tasks. Specifically, we focus on fine-tuning policies learned in simulation directly on hardware, while explicitly enforcing safety constraints. In doing so, we introduce SLowRL, a framework that combines Low-Rank Adaptation (LoRA) with training-time safety enforcement via a recovery policy. We evaluate our method both in simulation and on a real Unitree Go2 quadruped robot for jump and trot tasks. Experimental results show that our method achieves a $46.5\%$ reduction in fine-tuning time and near-zero safety violations compared to standard proximal policy optimization (PPO) baselines. Notably, we find that a rank-1 adaptation alone is sufficient to recover pre-trained performance in the real world, while maintaining stable and safe real-world fine-tuning. These results demonstrate the practicality of safe, efficient fine-tuning for dynamic real-world robotic applications.

Through encouraging self-exploration, reinforcement learning from verifiable rewards (RLVR) has significantly advanced the mathematical reasoning capabilities of large language models. As the starting point for RLVR, the capacity of supervised fine-tuning (SFT) to memorize new chain-of-thought trajectories provides a crucial initialization that shapes the subsequent exploration landscape. However, existing research primarily focuses on facilitating exploration during RLVR training, leaving exploration-aware SFT under-explored. To bridge this gap, we propose Offline eXploration-Aware (OXA) fine-tuning. Specifically, OXA optimizes two objectives: promoting low-confidence verified teacher-distillation data to internalize previously uncaptured reasoning patterns, and suppressing high-confidence incorrect self-distillation data to redistribute probability mass of incorrect patterns toward potentially correct candidates. Experimental results across 6 benchmarks show that OXA consistently improves mathematical reasoning performance, especially achieving an average gain of $+6$ Pass@1 and $+5$ Pass@$k$ points compared to conventional SFT on the Qwen2.5-1.5B-Math. Crucially, OXA elevates initial policy entropy, and performance gains persist throughout extensive RLVR training, demonstrating the long-term value of OXA.

Reinforcement learning has emerged as a powerful paradigm for improving large language model (LLM) reasoning, where rollouts are sampled from the policy and reward signals computed on those rollouts are used to update the policy. However, in data-scarce scenarios, obtaining ground-truth labels to verify rollouts at scale often requires expensive human annotation or labor-intensive expert verification. For instance, evaluating mathematical proofs demands expert review, and open-ended question answering lacks definitive ground truth. When ground-truth labels are scarce, the effectiveness of reinforcement learning fine-tuning is constrained. Inspired by the success of semi-supervised learning in propagating labels from labeled to unlabeled samples, we propose MemReward, a graph-based experience memory framework that integrates reward propagation directly into online policy optimization. MemReward stores rollouts (thinking processes and final answers) from an initial LLM policy as nodes in a heterogeneous graph connected by similarity and structural edges, over which a GNN propagates rewards from labeled to unlabeled rollouts. To train such a framework, we first warm up the GNN on labeled rollouts to predict rewards via heterogeneous aggregation over query, thinking, and answer nodes. During online RL fine-tuning, unlabeled rollouts are attached to the graph by query similarity, and the GNN predicts their rewards, yielding a hybrid reward acquisition strategy that combines ground-truth and GNN-predicted rewards. Experiments on Qwen2.5-1.5B and 3B in mathematics, question answering, and code generation demonstrate that MemReward, with ground-truth rewards on only 20% of rollouts, achieves 96.6% of Oracle performance on 1.5B and 97.3% on 3B, and closely approaches Oracle on out-of-domain tasks.

Diffusion-based robot navigation policies trained on large-scale imitation learning datasets, can generate multi-modal trajectories directly from the robot's visual observations, bypassing the traditional localization-mapping-planning pipeline and achieving strong zero-shot generalization. However, their performance remains constrained by the coverage of offline datasets, and when deployed in unseen settings, distribution shift often leads to accumulated trajectory errors and safety-critical failures. Adapting diffusion policies with reinforcement learning is challenging because their iterative denoising structure hinders effective gradient backpropagation, while also making the training of an additional value network computationally expensive and less stable. To address these issues, we propose a reinforcement learning fine-tuning framework tailored for diffusion-based navigation. The method leverages the inherent multi-trajectory sampling mechanism of diffusion models and adopts Group Relative Policy Optimization (GRPO), which estimates relative advantages across sampled trajectories without requiring a separate value network. To preserve pretrained representations while enabling adaptation, we freeze the visual encoder and selectively update the higher decoder layers and action head, enhancing safety-aware behaviors through online environmental feedback. On the PointGoal task in Isaac Sim, our approach improves the Success Rate from 52.0% to 58.7% and SPL from 0.49 to 0.54 on unseen scenes, while reducing collision frequency. Additional experiments show that the fine-tuned policy transfers zero-shot to a real quadruped platform and maintains stable performance in geometrically out-of-distribution environments, suggesting improved adaptability and safe generalization to new domains.

Despite recent progress in Large Language Model (LLM) Agents for Software Engineering (SWE) tasks, end-to-end fine-tuning typically relies on verifiable terminal rewards such as whether all unit tests pass. While these binary signals reflect whether the final solution is correct, they provide little guidance for shaping intermediate behaviors during multi-step interactions, thereby limiting improvements in the overall quality of the resolution process. To address this, we introduce a rubric-based Generative Reward Model (GRM) that provides richer learning signals. The GRM is equipped with human-designed rubrics that indicate criteria for encouraging or discouraging specific behavioral patterns, and we leverage this feedback for high-quality training data collection via trajectory filtration. When used for Reinforced Fine-Tuning (RFT) on SWE Tasks, our approach outperforms terminal-score-only rejection sampling: it more effectively suppresses undesirable patterns while promoting beneficial ones, as confirmed by case analyses, and it ultimately improves final test accuracy.

Cross-entropy (CE) training provides dense and scalable supervision for language models, but it optimizes next-token prediction under teacher forcing rather than sequence-level behavior under model rollouts. We introduce a feature-matching objective for language-model fine-tuning that targets sequence-level statistics of the completion distribution, providing dense semantic feedback without requiring a task-specific verifier or preference model. To optimize this objective efficiently, we propose energy-based fine-tuning (EBFT), which uses strided block-parallel sampling to generate multiple rollouts from nested prefixes concurrently, batches feature extraction over these rollouts, and uses the resulting embeddings to perform an on-policy policy-gradient update. We present a theoretical perspective connecting EBFT to KL-regularized feature-matching and energy-based modeling. Empirically, across Q&A coding, unstructured coding, and translation, EBFT matches RLVR and outperforms SFT on downstream accuracy while achieving a lower validation cross-entropy than both methods.

Reinforcement fine-tuning (RFT) has shown promise for training LLM agents to perform multi-turn decision-making based on environment feedback. However, most existing evaluations remain largely in-domain: training and testing are conducted in the same environment or even on the same tasks. In real-world deployment, agents may operate in unseen environments with different background knowledge, observation spaces, and action interfaces. To characterize the generalization profile of RFT under such shifts, we conduct a systematic study along three axes: (1) within-environment generalization across task difficulty, (2) cross-environment transfer to unseen environments, and (3) sequential multi-environment training to quantify transfer and forgetting. Our results show that RFT generalizes well across task difficulty within an environment, but exhibits weaker transfer to unseen environments, which correlates with shifts in both semantic priors and observation/action interfaces. In contrast, sequential training yields promising downstream gains with minimal upstream forgetting, and mixture training across environments improves the overall balance. We further provide detailed analyses and deeper insights, and hope our work helps the community develop and deploy generalizable LLM agents.

Reinforcement learning with Verifiable Rewards (RLVR) has emerged as a powerful paradigm for eliciting reasoning capabilities in large language models, particularly in mathematics and coding. While recent efforts have extended this paradigm to broader general scientific (STEM) domains, the complex interplay between supervised fine-tuning (SFT) and RL in these contexts remains underexplored. In this paper, we conduct controlled experiments revealing a critical challenge: for general STEM domains, RL applied directly to base models is highly sample-inefficient and is consistently surpassed by supervised fine-tuning (SFT) on moderate-quality responses. Yet sequential SFT followed by RL can further improve performance, suggesting that the two stages play complementary roles, and that how training data is allocated between them matters. Therefore, we propose DeReason, a difficulty-based data decoupling strategy for general reasoning. DeReason partitions training data by reasoning intensity estimated via LLM-based scoring into reasoning-intensive and non-reasoning-intensive subsets. It allocates broad-coverage, non-reasoning-intensive problems to SFT to establish foundational domain knowledge, and reserves a focused subset of difficult problems for RL to cultivate complex reasoning. We demonstrate that this principled decoupling yields better performance than randomly splitting the data for sequential SFT and RL. Extensive experiments on general STEM and mathematical benchmarks demonstrate that our decoupled curriculum training significantly outperforms SFT-only, RL-only, and random-split baselines. Our work provides a systematic study of the interplay between SFT and RL for general reasoning, offering a highly effective and generalized post-training recipe.

Reinforcement learning (RL) finetuning has become a key technique for enhancing the reasoning abilities of large language models (LLMs). However, its effectiveness critically depends on the selection of training data. Recent advances underscore the importance of online prompt selection methods, which typically concentrate training on partially solved or moderately challenging examples under the current policy, thereby yielding more effective model updates. While significantly accelerating RL finetuning in terms of training steps, they also incur substantial computational overhead by requiring extensive LLM rollouts over large candidate batches to identify informative samples, an expense that can outweigh the finetuning process itself. To address this challenge, this work proposes Dynamics-Predictive Sampling (DPS), which online predicts and selects informative prompts by inferring their learning dynamics prior to costly rollouts. Specifically, we introduce a new perspective by modeling each prompt's solving progress during RL finetuning as a dynamical system, where the extent of solving is represented as the state and the transition is characterized by a hidden Markov model. Using historical rollout reward signals, we perform online Bayesian inference to estimate evolving state distributions, and the inference outcome provides a predictive prior for efficient prompt selection without rollout-intensive filtering. Empirical results across diverse reasoning tasks, including mathematics, planning, and visual geometry, demonstrate that DPS substantially reduces redundant rollouts, accelerates the training process, and achieves superior reasoning performance.

Large language models are increasingly trained as heterogeneous systems spanning multiple domains, expert partitions, and agentic pipelines, yet prevalent proximal objectives operate at a single scale and lack a principled mechanism for coupling token-level, trajectory-level, and higher-level hierarchical stability control. To bridge this gap, we derive the Aggregational Policy Censoring Objective (APC-Obj), the first exact unconstrained reformulation of sample-based TV-TRPO, establishing that clipping-based surrogate design and trust-region optimization are dual formulations of the same problem. Building on this foundation, we develop Fiber Bundle Gating (FBG), an algebraic framework that organizes sampled RL data as a fiber bundle and decomposes ratio gating into a base-level gate on trajectory aggregates and a fiber-level gate on per-token residuals, with provable first-order agreement with the true RL objective near on-policy. From APC-Obj and FBG we derive Fibration Policy Optimization (or simply, FiberPO), a concrete objective whose Jacobian is block-diagonal over trajectories, reduces to identity at on-policy, and provides better update direction thus improving token efficiency. The compositional nature of the framework extends beyond the trajectory-token case: fibrations compose algebraically into a Fibration Gating Hierarchy (FGH) that scales the same gating mechanism to arbitrary hierarchical depth without new primitives, as demonstrated by FiberPO-Domain, a four-level instantiation with independent trust-region budgets at the domain, prompt group, trajectory, and token levels. Together, these results connect the trust-region theory, a compositional algebraic structure, and practical multi-scale stability control into a unified framework for LLM policy optimization.

Nucleus detection in histopathology is pivotal for a wide range of clinical applications. Existing approaches either regress nuclear proxy maps that require complex post-processing, or employ dense anchors or queries that introduce severe foreground-background imbalance. In this work, we reformulate nucleus detection as next-point prediction, wherein a multimodal large language model is developed to directly output foreground nucleus centroids from the input image. The model is trained in two stages. In the supervised learning stage, we propose spatial-aware soft supervision to relax strict centroid matching and a chain-of-visual-thought strategy to incorporate visual priors that facilitate coordinate prediction. In the reinforcement fine-tuning stage, we design distribution matching reward, low-variance group filtering, and fine-grained advantage shaping to further improve the model's detection quality. Extensive experiments on nine widely used benchmarks demonstrate the superiority of our method. Code will be released soon.

Agentic systems operating over large tool ecosystems must plan and execute long-horizon workflows under weak or non-verifiable supervision. While frontier models mitigate these challenges through scale and large context budgets, small language models (SLMs) remain brittle: eager tool loading saturates context, execution errors compound over time, and sparse rewards limit learning. We introduce ATLAS, a reinforcement finetuning framework that enables SLMs to operate effectively in large-scale toolspace environments by learning how to acquire context and how to execute actions. Our approach makes two key contributions. First, we treat context control and execution structure as learnable decisions, combining iterative tool loading with programmatic tool orchestration to bound context growth and stabilize long-horizon trajectories. Second, we propose rubric-based reinforcement finetuning, which decomposes task success into structured, task-aligned criteria and enables scalable training using small judge models. Across MCP benchmarks, these design choices yield large and consistent gains over generic RL baselines, allowing a 4B SLM to approach frontier-agent performance under far tighter parameter and context budgets.

Graphical User Interface (GUI) Agents, benefiting from recent advances in multimodal large language models (MLLM), have achieved significant development. However, due to the frequent updates of GUI applications, adapting to new tasks without forgetting old tasks in GUI continual learning remains an open problem. In this work, we reveal that while Supervised Fine-Tuning (SFT) facilitates fast adaptation, it often triggers knowledge overwriting, whereas Reinforcement Learning (RL) demonstrates an inherent resilience that shields prior interaction logic from erasure. Based on this insight, we propose a Continual GUI Learning (CGL) framework that dynamically balances adaptation efficiency and skill retention by enhancing the synergy between SFT and RL. Specifically, we introduce an SFT proportion adjustment mechanism guided by policy entropy to dynamically control the weight allocation between the SFT and RL training phases. To resolve explicit gradient interference, we further develop a specialized gradient surgery strategy. By projecting exploratory SFT gradients onto GRPO-based anchor gradients, our method explicitly clips the components of SFT gradients that conflict with GRPO. On top of that, we establish an AndroidControl-CL benchmark, which divides GUI applications into distinct task groups to effectively simulate and evaluate the performance of continual GUI learning. Experimental results demonstrate the effectiveness of our proposed CGL framework across continual learning scenarios. The benchmark, code, and model will be made publicly available.