Reward Gradients

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.

Reinforcement learning is increasingly used to align image generators with reward signals, and Flow-GRPO recently extended this paradigm to flow-matching models by treating the denoising sampler as a stochastic policy that can be optimized from reward feedback. Training in this setting is unstable in a way specific to multi-step denoising: the policy update changes systematically across denoising steps, with importance ratios drifting below one, becoming increasingly dispersed, clipping at different rates, and leaving fewer usable samples late in training. Prior work treats these effects as separate failure modes and addresses each with a hand-tuned stabilizer. We show instead that they arise from a single per-step quantity, which we call path variance. This quantity is determined exactly by the sampler's Gaussian transition kernel and can be estimated cheaply during training. This reframes instability as a resource that can be measured and budgeted rather than a collection of symptoms to repair. Our method, $λ$-Controlled GRPO, calibrates importance-ratio behavior from this predicted law rather than from noisy empirical statistics, and allocates gradient effort across denoising steps according to their predicted cost. The two scales governing the update are fixed by standard policy choices rather than introduced as free tuning parameters. On a text-to-image model under two reward settings, rendering difficult target text scored by optical character recognition and matching human preferences scored by a preference model, $λ$-Controlled GRPO improves both text accuracy and preference reward over the strongest empirical stabilizer. It also keeps late-step path variance within its intended budget, precisely where the baseline systematically overshoots. The result is a Flow-GRPO update calibrated by its own transition law rather than stabilized after instability appears.

Reinforcement learning from verifiable rewards (RLVR) drives chain-of-thought reasoning in large language models, yet its binary outcome reward cannot distinguish among correct trajectories. Existing dense reward alternatives, from surface heuristics to process reward models, either ignore the expert solutions already present in training corpora or require expensive offline annotation. We propose Gradient-Aligned Reward (GAR), which operates in the policy's own gradient space: truncated backpropagation through the output projection layer extracts a compact gradient vector for each rollout, and cosine similarity with an expert-anchor gradient yields a dense, reasoning-aware reward with less than 9% wall-clock overhead. We prove that this cosine admits a multiplicative decomposition into prediction-error and activation-pattern factors, providing a concrete characterization of what the alignment signal measures. On Qwen3-4B and Qwen3-8B, GAR consistently improves over GRPO and other baselines on competition-level math benchmarks and transfers to GPQA Diamond and MMLU-Pro without domain-specific data. Code and data are available at https://github.com/LQgdwind/GAR.

Diffusion models have recently advanced text-to-video (T2V) generation, yet they still struggle with fine-grained compositional alignment, such as attribute binding, spatial relations, and object interactions. While reward-based fine-tuning improves alignment, it is susceptible to reward hacking and adapts poorly to new prompt distributions. In this work, we propose NoisEasier, a test-time scaling framework that improves T2V generation through differentiable reward-guided noise optimization without modifying the underlying model. By combining efficient short-step generators with a multi-objective reward formulation, NoisEasier enables stable and practical test-time optimization under realistic inference budgets. Our key insight is that jointly optimizing the entire stochastic trajectory accelerates reward convergence and improves compositional alignment over optimizing only the initial latent, with negligible additional computational and time cost. Experiments on VBench and T2V-CompBench demonstrate consistent improvements across multiple backbones, achieving over 10% average gains on challenging dimensions such as attribute binding, object interaction, and numeracy. Overall, NoisEasier serves as both a flexible alternative and a complementary enhancement to reward-based fine-tuning, establishing test-time scaling as an effective paradigm for controllable text-to-video generation.

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.

On-policy distillation (OPD) trains a student on its own trajectories under dense token-level supervision from a teacher. Reward-extrapolation methods such as ExOPD amplify the teacher-reference log-likelihood ratio to move beyond direct imitation, but apply a single global coefficient $λ$ to every token. This can drive the student to fit extreme peaks in the implicit reward, causing reward hacking and unstable training, and the optimal $λ$ varies across domains, requiring costly sweeps. We propose REOPD, a reliability-adaptive reward extrapolation framework for OPD. REOPD combines a token-level compatibility weight with a batch-level adaptive budget, yielding a token-wise coefficient $λ_{b,t}=1+γ_b q_t$ that preserves teacher alignment while selectively extrapolating along reliable teacher-reference directions. It requires no verifier, reward model, value model, or extra rollout beyond standard OPD. REOPD outperforms G-OPD on single-teacher mathematics and on both domains in the multi-teacher setting, while matching G-OPD on single-teacher code, demonstrating effective fine-grained reliability adaptation across domains and teacher configurations.

Assortment optimization is a fundamental problem in revenue management, typically addressed using parametric choice models such as the multinomial logit (MNL) and its variants. While these models enable tractable formulations, their performance is sensitive to model misspecification and often struggles to capture complex customer behavior. In this paper, we propose a model-agnostic framework for assortment optimization based on guided discrete diffusion. We represent assortments as binary vectors and perform stochastic search via a learned reverse diffusion process, avoiding explicit combinatorial enumeration. To incorporate decision objectives, we introduce a reward-guided mechanism that biases local transitions using estimates of expected revenue. This allows the method to effectively balance exploration and exploitation during generation. Empirically, we show that the proposed approach consistently identifies high-quality assortments and remains robust under model misspecification, often recovering near-optimal solutions in high-dimensional settings. Moreover, the generative nature of diffusion enables the production of diverse high-performing assortments, offering flexibility beyond a single deterministic solution. These results highlight the potential of generative modeling as a scalable and robust paradigm for combinatorial optimization in data-driven decision-making.

Sparse, delayed, and weakly informative rewards remain central obstacles to efficient reinforcement learning. Reward shaping addresses these limitations by supplementing the task reward with an auxiliary signal that can accelerate learning while, in the classical setting, the original objective remains the evaluation criterion. Established theory guarantees safety for fixed shaping signals: potential-based reward shaping preserves optimal policies when the auxiliary term is the discounted difference of a time-invariant potential. In contemporary reinforcement learning systems, however, both the learner and the information available for guidance evolve during training: value estimates improve, novelty diminishes, feedback shifts, and predictive models are refined. Adaptive reward mechanisms occur across exploration, Bayesian inference, human-in-the-loop learning, automated reward design, and foundation-model-based approaches. This study introduces a unified analytical framework for comparing dynamic reward shaping and neighbouring adaptive reward mechanisms. The proposed framework distinguishes parametric revision from state-dependent variation, separates additive shaping from reward replacement and reward-adjacent guidance, and organises existing methods along temporal, informational, and theoretical dimensions. Using this framework, twelve method families are comparatively analysed. The framework further highlights the conditions under which optimality guarantees survive contemporary deep reinforcement learning pipelines, replay buffers, bootstrapped critics, and reward normalisation, while exposing the unresolved relationship between adaptation rate and learner stability.

Latent reward models can supervise visual diffusion models without decoding intermediate states into pixel space. This makes alignment with human preferences more efficient. However, existing latent reward models output only scalar scores. They do not estimate the uncertainty of each prediction. The generator therefore cannot determine which feedback is reliable. This can drive optimization in the wrong direction and lead to reward hacking. We propose \textsc{SURE}, a unified latent-space framework for image and video diffusion models. It learns reward distributions and directly uses their reliability to guide dense post-training. First, we propose sample-adaptive latent reward model (\textsc{SURE-LRM}). It predicts a Gaussian utility for each noisy latent. Its mean predicts the reward score. Its variance reflect the uncertainty of prediction without human annotation. The learned distribution then guides post-training through uncertainty-guided reward feedback learning (\textsc{SURE-REFL}). This method provides uncertainty-guided dense feedback along the denoising trajectory. At selected transitions, \textsc{SURE-REFL} queries the frozen \textsc{SURE-LRM}. It converts detached variance into reliability weights for samples at the same transition. Each weighted reward is backpropagated only through its local transition. The entire process remains in latent space and requires neither pixel-space decoding nor the full denoising graph. Experiments show that \textsc{SURE-LRM} improves preference prediction over strong baselines. \textsc{SURE-REFL} achieves the sota performance among various metrics and further improves optimization stability. It also achieves the highest VBench quality, semantic, and total scores among the evaluated methods.

Flow-based generative models are typically sampled by solving a deterministic ordinary differential equation (ODE), whereas online reinforcement learning requires stochastic rollouts for policy exploration and optimization. Existing GRPO methods for flow models therefore replace the inference-time ODE with a stochastic differential equation (SDE) during training. Although the ODE and SDE share the same marginal distributions in continuous time, their finite-step discretizations can differ substantially. In particular, SDE rollouts often become blurry as the exploration noise increases, creating a mismatch between the samples used for reinforcement learning and those generated by the test-time ODE sampler. We introduce LC-GRPO, a flow-based GRPO framework with Langevin correction. Each rollout transition first takes an inference-aligned ODE Euler step and then applies a stochastic Langevin correction targeting the marginal distribution at the resulting timestep. The required score is recovered directly from the flow velocity, requiring no additional score model, while the resulting transition remains an isotropic Gaussian with a tractable likelihood for policy optimization. We theoretically show that, under suitable conditions, one Langevin correction step reduces the Wasserstein error of an imperfect ODE Euler step. At a matched randomness level, we further show that the proposed transition can be more accurate than the standard Euler--Maruyama discretization of the reverse SDE. Experiments on SD3.5-Medium, FLUX.1-Dev, and HunyuanVideo demonstrate that LC-GRPO consistently improves reward optimization across text-to-image and text-to-video tasks, preserves generation quality, and substantially narrows the gap between stochastic training rollouts and deterministic test-time ODE inference.

Aligning diffusion models with human preferences usually relies on a sparse terminal reward evaluated on the final generated samples, presenting a severe temporal credit-assignment challenge across the multi-step denoising process. We propose Latent Reward Registers, a mechanism that estimates terminal preference directly from intermediate noisy latents by prepending learnable, position-free register tokens to the input sequence of a frozen Diffusion Transformer (DiT). This independent readout mechanism extracts latent reward evidence without altering the generator's hidden states or velocity field. The resulting dense, differentiable reward signal throughout the full denoising process facilitates two alignment strategies. For training, Reward-Gradient On-Policy Distillation (RG-OPD) distills reward-guided updates along on-policy trajectories, bypassing the computationally expensive rollouts of standard policy gradients. For inference, Reward-Guided Sampling (RGS) steers trajectories via magnitude-matched reward gradients without parameter updates. Empirically, at high noise levels (u = 0.8), the registers reach the highest pairwise accuracy among the evaluated latent reward models. Furthermore, RG-OPD outperforms online reinforcement learning baselines while reducing GPU hours by up to 33x, and RGS establishes a new state-of-the-art among training-free methods, strictly enhancing both alignment and perceptual metrics. Code and weights are available at https://github.com/Guanys-dar/latent-reward-register

A broad range of models face the mismatch where they are updated through trajectory losses but are evaluated by downstream task reward. Here, a trajectory is a training instance that induces a surrogate loss whose reduction might not track the model's decision utility update. Theoretically, we ask when one step of trajectory training reduces both population surrogate loss and decision risk, and how transfer accumulates along repeated updates. To formalize this, we first fix a checkpoint and a restricted update space, and define the reductions in population surrogate risk and decision risk induced by a trajectory as its learnability and decision utility, respectively. On this basis, our theory yields four main results. First, a one-step transfer bound separates their discrepancy into first-order gradient misalignment after nonnegative calibration and second-order curvature; and a pathwise extension accumulates the same terms over repeated updates. Second, when the accessible surrogate gradient is nonzero, universal first-order transfer over every accessible direction holds exactly when the accessible surrogate and decision gradients are positively collinear. Third, the calibration gap bounds the decision regret of learnability-based trajectory selection, while a candidate-difference refinement tightens this guarantee by retaining only directions that affect pairwise rankings. Finally, we establish an approximation--calibration trade-off across nested update spaces. Controlled gridworld and LLM post-training experiments yield results consistent with our predictions.

Error-penalized scoring rules ($+1$ for a correct answer, $-λ$ for a wrong one, $0$ for abstaining) are increasingly prescribed against hallucination: a rational agent facing such a rule answers exactly when its correctness probability exceeds Chow's threshold $t^\ast=λ/(1+λ)$. We prove that a KL-anchored gradient learner can do the opposite. When abstention is a discrete action, the reward gradient and the anchor's restoring force are throttled by the same gate-saturation factor and die together: under explicit conditions (among them, blanket answering loses score in expectation and prompts share a bounded readout) the model drifts toward refusing everything, its mean training reward rising to zero like $1/t$ in training time $t$, so the curve reads as improvement while coverage collapses. The advantage estimator compounds the failure: in its sparse-answer regime, group normalization silently replaces every designed penalty with an effective penalty of one, moving the learned threshold from $t^\ast$ to $1/2$. The repair is structural: train a mandatory confidence report with a strictly proper score plus a correctness reward, and abstain only at deployment by thresholding the report. The always-emitted report has no gate to saturate, so no shared factor can kill its reward gradient and its anchor together, and its calibrated optimum is attracting. Simulations confirm every prediction, and experiments on language models at two scales confirm the mechanism live: the rule silences questions the models demonstrably still solve within ten optimizer steps, an ablation isolates the cause, and report-level training raises coverage, accuracy, and calibration together.

Reinforcement learning from human feedback (RLHF) has emerged as a powerful paradigm for aligning generative models with human preferences. However, applying RLHF to diffusion models remains highly feedback inefficient, as existing approaches typically require large amounts of human or reward model evaluations. This limitation reduces the practicality of diffusion RLHF in realworld settings where feedback is the primary bottleneck. In this paper, we propose two complementary strategies that substantially improve the feedback efficiency of diffusion RLHF while preserving generalization to unseen prompts. Our key observation is that reward information in diffusion trajectories is unevenly distributed: not all denoising timesteps or trajectories contribute equally to learning from a reward signal. By emphasizing informative timesteps and trajectories during optimization, we obtain more effective gradient updates. First, we introduce a per-timestep weighting scheme that reweights denoising steps during policy optimization. We theoretically connect this weighting to the optimal convergence properties of proximal policy optimization (PPO) and approximate the resulting weighting trend empirically. Second, we introduce a replay mechanism that prioritizes informative trajectories, enabling the model to reuse past samples instead of repeatedly querying new rewards. Together, these strategies significantly improve the feedback efficiency of diffusion RLHF. Under identical hyperparameter settings, our approach achieves up to a 6$\times$ improvement in sample efficiency compared to widely used diffusion RLHF baselines.

Behavior-cloned diffusion policies are expressive but remain vulnerable to covariate shift: small deviations from demonstrated states can compound into task failure. Existing methods address this either by expanding the training distribution through expert corrections or synthetic augmentation, or by steering a frozen policy at test time with guidance from a learned model. The former can be expensive or assumption-dependent, while the latter discards the corrected trajectories after execution. We introduce ReGuide, a self-improving framework that treats guided rollouts as reusable on-policy recovery data. ReGuide first uses Phase-Conditioned Guidance (PCG) to generate corrective rollouts: it constructs phase-specific latent targets, applies guidance only in the drifted-but-recoverable regime, and guides through the estimated clean action to match the dynamics model's training distribution. Successful guided rollouts are then absorbed back into the policy through ReGuide-FT, which fine-tunes the current checkpoint, or ReGuide-FS, which retrains from scratch on the augmented dataset; the two can also be composed and iterated. On Robomimic Can, Square, Transport, and Tool Hang, ReGuide improves base-policy success by $1.3$--$7.7\times$, outperforms LPB in the test-time-only setting, and matched-data ablations show that the gains come from guided recovery data rather than additional rollouts alone.

Reinforcement learning (RL) post-training improves the reward alignment of flow-based generators, but often degrades perceptual quality in ways that are not captured by the reward proxy. We identify a simple structural signature of this drift: across three post-training methods (NFT, AWM, DPO), RL fine-tuning inflates the per-step velocity norm $\|v_θ\|$ by $5\%$ to $15\%$ relative to the reference. A form of norm inflation has been studied in classifier-free guidance (CFG), where rescaling the velocity back to a reference norm at inference time can mitigate the resulting artifacts. However, this inference-time correction does not transfer cleanly to RL: rescaling $v_θ$ to match $\|v_{\text{ref}}\|$ at inference time neither improves reward nor fixes the quality degradation, because the inflation is co-adapted into the model weights. Furthermore, an adjoint sensitivity analysis shows that velocity magnitude rescaling carries no coherent first-order reward signal at the batch level, indicating that suppressing norm inflation is unlikely to remove a consistently reward-carrying component. Since inference-time renormalization fails while norm suppression carries no reward cost, training-time intervention is the appropriate strategy. Together, these findings motivate NormGuard, a hinge penalty that activates only when $\|v_θ\|$ exceeds $\|v_{\text{ref}}\|$ and composes additively with any velocity-local base loss. Across two base models, three post-training methods, and two reward proxies, NormGuard consistently improves MLLM-judged image quality and forensic realism while preserving reward, with gains that amplify under few-step inference and are not explained by early stopping.

This paper considers stochastic linear contextual bandits (SLCB) with bounded reward noise. Existing works typically assume sub-Gaussian reward noise and bounded expected rewards, under which the optimal regret bound scales as $\tilde{O}(\sqrt{T})$ in terms of horizon $T$. However, in many applications, realized/observed rewards are also naturally bounded, implying bounded reward noise. Bounded noise is more informative than the sub-Gaussian condition but has not been leveraged explicitly in the SLCB literature. In this paper, we propose a novel algorithm SME-OFU by utilizing an uncertainty quantification method called set-membership estimation (SME) and applying the principle of optimism in the face of uncertainty (OFU). Our algorithm enjoys an improved regret bound $O(\log T)$. Notice that this does not contradict the existing optimal bound $\tilde{O}(\sqrt{T})$ for sub-Gaussian noise because bounded noise is a stronger condition. Finally, simulations show empirical improvements of SME-OFU over a benchmark algorithm designed for sub-Gaussian noise when the reward noise is bounded.

We introduce the Noise-Tilted Reverse Kernel (NTRK), a reward-guided diffusion sampler that injects reward gradients through the noise term, leaving the pretrained reverse kernel unchanged and requiring only a single sample per step. Reward-guided sampling at inference time has greatly expanded the versatility of pretrained diffusion models. Yet existing methods face a trade-off. Gradient-based guidance shifts the reverse mean, steering generation but pushing intermediate states outside the region that the model was trained on and degrading quality. Search-based methods preserve quality but gain no gradient signal. No prior method achieves both. NTRK resolves this by keeping the reverse mean fixed and biasing the noise term toward high reward. This is enabled by a whitening operator, the central mechanism behind NTRK, which converts reward gradients into noise-compatible perturbations without losing their guiding signal. Across various reward alignment tasks, NTRK outperforms recent state-of-the-art baselines without losing sample quality. Remarkably, on aesthetic generation, NTRK surpasses the reward of the best baseline at 500 NFEs using only 25 NFEs, a 20 times reduction in compute.

Ensuring the reliability of Large Language Models (LLMs) under distribution drift requires inference-time adaptation. While inference-time alignment methods such as Best-of-$N$ and rejection sampling are widely used, they frame the task as a sampling-intensive, reward-guided search, leading to two key limitations: their performance is bounded by the base model's generation quality, and their reliance on imperfect reward models makes them vulnerable to reward hacking. To address these challenges, we introduce Gradient-Guided Reward Optimization (GGRO), a lightweight inference-time method that performs targeted, minimal intervention during decoding via gradient guidance. Specifically, GGRO monitors token-level entropy to identify high-uncertainty regions indicative of drift or misalignment. Upon detection, it responds by injecting nudging tokens, generated using gradient signals from an off-the-shelf reward model, to steer the generation trajectory rather than merely re-ranking samples. Experiments show that GGRO consistently improves inference-time alignment across safety, helpfulness, and reasoning benchmarks. It also increases coverage of high-quality responses and robustness to reward hacking, with minimal computational overhead. Code is available at https://github.com/lhk2004/GGRO.

Reinforcement learning (RL) presents a promising avenue for enhancing generative recommendation beyond supervised imitation, leveraging reward signals to guide policy improvement. However, its efficacy is critically contingent on the trustworthiness of the reward model for the samples it evaluates. In practice, production rankers, the widely adopted reward models, are trained on exposure-biased logs, leading to sample-dependent inaccuracies that violate this assumption. Our stratified analysis uncovers a consistent pattern: reward guidance is most beneficial when the policy exhibits uncertainty and the ranker can effectively discriminate the ground-truth item from rollout negatives. On other samples, the reward signal is either negligible or detrimental, highlighting the risk of uniform RL application. To address such an issue, we introduce AdaGRPO, a novel framework that treats reward-guided optimization as selective admission rather than uniform pressure. Training is anchored in supervised negative log-likelihood, while the GRPO objective is gated by a binary, per-sample clip determined by two rollout diagnostics: policy-side difficulty and reward discriminability. Instances failing either diagnostic default to pure supervision, ensuring stability and mitigating the amplification of noisy gradients. We validate AdaGRPO on a large-scale e-commerce dataset. At the best intermediate checkpoint, it elevates HR@10 from 11.01% to 12.18% while constraining hallucination below 0.22%, and maintains robustness at the final checkpoint (HR@10 11.63%, hallucination 0.27%), outperforming fixed NLL--GRPO mixtures across the retrieval--validity frontier. In production A/B tests, AdaGRPO achieves statistically significant gains in click-through rate and dwell time, confirming its practical utility.

LLM-driven software engineering agents have become a central testbed for real-world language-model capability, yet their training remains limited by the availability of high-quality SWE tasks. Existing synthetic data methods typically create tasks through fixed mutation or bug-injection procedures, making the resulting distributions largely independent of the agent's own weaknesses and training progress. We introduce Socratic-SWE, a closed-loop self-evolution framework that reuses the agent's historical solving traces as a source of training signal. Rather than treating traces only as evidence for reward computation, Socratic-SWE distills them into structured agent skills that summarize recurring failures and effective repair patterns. These skills then guide the generation of targeted repair tasks in real repositories. Candidate tasks are checked through execution-based validation and scored with a solver-gradient alignment reward, so that the retained tasks are both verifiable and useful for improving the Solver. The updated Solver produces new traces, enabling the task curriculum to adapt over successive rounds. Across SWE-bench Verified, SWE-bench Lite, SWE-bench Pro, and Terminal-Bench 2.0, Socratic-SWE consistently improves over self-evolving baselines under the same compute budget, reaching 50.40% on SWE-bench Verified after three iterations. These results suggest that solving traces can serve as a scalable substrate for self-evolving SWE agents.

A reinforcement-learning agent maximises its reward, which can diverge from the outcome its designer intended. In physical control the reward rarely closes that gap, and drag reduction in wall turbulence makes it concrete. A mass-conservation projection couples agents' outputs and erases the per-agent credit the policy gradient needs; a memoryless policy cannot resolve the slow near-wall cycle it acts on; and a pressure-gradient reward pays for nominal drag reduction by pumping power through the wall. Two degenerate controllers achieve large drag reductions while total dissipation rises, so the reported figure can mask a more wasteful flow. We trace each fault to its cause and fix it: a differentiable projection that restores credit, a recurrent policy with a widened sensing stencil, and a reward scored on the true wall power. The corrected controller acts on the flow within a closed energy budget, earning a conservative $17\%$ under honest accounting.

Reward guidance algorithms steer a learned generative process toward the reward-tilted measure at inference time. While empirically powerful, these methods are prone to reward hacking: the guided model over-optimizes the reward at the cost of fidelity to the learned distribution. Prior work has attributed this to the complexity of neural reward functions or implicit biases in diffusion training, but its fundamental origins remain poorly understood. We show that reward hacking arises from an approximation made in most practical implementations of reward-guided diffusion -- finite-particle plug-in estimation of the Doob h-function -- even in the simplest non-trivial settings of Gaussian and Gaussian mixture targets with quadratic rewards. In closed form, we isolate two distinct failure modes of the plug-in estimator: it leads to reward hacking within each mode and it cannot select high-reward modes. We propose a closed-form reward damping schedule that corrects the within-mode bias with no additional compute, and clarify the role of best-of-n sampling in compensating for the mode selection failure. Experiments on Gaussian mixture targets, a 2D checkerboard, and FLUX.1 text-to-image generation confirm that our theoretical insights carry over to practical settings.

Reinforcement learning (RL) can be used to improve the policy (denoiser) of diffusion large language models (dLLMs), while being hindered by the intractability of the policy likelihood. A dominant and efficient family of methods replaces the likelihood in standard RL with its evidence lower bound (ELBO), estimated from randomly masked sequences. Despite being well aligned with pre-training, these approaches introduce bias through training--inference mismatch by using the ELBO as a likelihood surrogate, which can degrade performance. In this work, we propose Guided Denoiser Self-Distillation (GDSD) to directly distill the denoiser of dLLMs from an advantage-guided self-teacher, derived from the closed-form optimum of reverse-KL regularized RL. GDSD matches the dLLM's denoiser logits to the teacher's via a normalization-free objective, which reduces RL to likelihood-free self-distillation and thus bypasses the TIM biases. Recent ELBO-based methods emerge as instances of applying different distillation divergences, but with diagnosable pathologies that GDSD avoids. On planning, math, and coding benchmarks with LLaDA-8B and Dream-7B, GDSD consistently outperforms prior state-of-the-art ELBO-based methods with a more stable training reward dynamics, achieving test-accuracy improvements of up to $+19.6\%$. These results suggest that direct denoiser self-distillation, without relying on an ELBO likelihood surrogate, can provide a more stable and effective RL procedure for dLLMs. Code is available at https://github.com/GaryBall/GDSD.

We introduce AdvantageFlow, a forward-process reinforcement learning algorithm for rectified flow models. Unlike Flow-GRPO, which optimizes the reverse process, we optimize an advantage-weighted forward-process prediction loss. This optimization problem is unstable when advantages are negative and the loss becomes non-convex. We stabilize it by rollout policy regularization, which reduces variance and arises from fitting a local reward-improving target distribution. We evaluate AdvantageFlow on image generation tasks with Stable Diffusion 3.5 Medium. It outperforms both Flow-GRPO and a state-of-the-art forward-process RL baseline based on negative-aware fine-tuning.

Off-dynamics offline reinforcement learning seeks to learn a target-domain policy from a large source dataset and a limited target dataset under mismatched transition dynamics. Existing approaches such as reward augmentation and data filtering are constrained to the source dataset and cannot synthesize new target behavior to improve coverage beyond the collected source trajectories. While recent model-based methods attempt to address this by learning target-aware dynamics, the generated experience is constructed only at the transition level, which leads to accumulated errors over long horizons. These limitations necessitate a shift toward trajectory-level generation for off-dynamics offline RL. We propose CEDGE, a Cross-domain Energy-guided Diffusion GEneration framework. CEDGE trains a trajectory diffusion model on source-domain trajectories and adapts the generated samples to the target domain through energy guidance. This guidance is derived by minimizing the distribution mismatch between the source and desired target-domain trajectories and is decomposed into return, domain, and behavior energy components. The resulting energy-guided trajectories are useful both for direct planning and as synthetic data for policy learning. Since target adaptation is achieved via energy guidance rather than retraining the diffusion model, CEDGE can be efficiently adapted to new target dynamics compared to previous methods. Experiments on the ODRL benchmark demonstrate that trajectory-level energy-guided generation improves diffusion planning under dynamics shifts and produces synthetic data that improves downstream target policy learning.

The efficient Test-Time Scaling (TTS) paradigm offers a promising perspective for enhancing the generation performance of diffusion models. However, current solutions are limited to a static, pre-defined noise pool and suffer from inflexible noise exploration across the denoising trajectory. To bridge this gap, we propose RTS, a novel Reward-guided Trajectory Scaling method to fully unlock the generative potential of diffusion models. Unlike existing methods, RTS facilitates the synthesis of refined, high-fidelity images via two core innovations: 1) a reward-guided noise optimization strategy to actively direct the search towards promising regions; and 2) a sparse test-time scaling framework together with a PCA-driven curvature analysis scheme to prioritize key intermediate steps in the entire denoising space, effectively compressing the search space. Experiments show our approach outperforms baselines by 15.6% across GenEval Score, and a 60.4% enhancement in ImageReward score, setting a new SOTA while providing a practical guideline for more effective test-time scaling across diffusion-specific architectures.

Reinforcement learning with verifiable rewards (RLVR) has become a dominant paradigm for improving reasoning in large language models (LLMs), yet the underlying geometry of the resulting parameter trajectories remains underexplored. In this work, we demonstrate that RLVR weight trajectories are extremely low-rank and highly predictable. Specifically, we find that the majority of downstream performance gains are captured by a rank-1 approximation of the parameter deltas, where the magnitude of this projection evolves near-linearly with training steps. Motivated by this, we propose a simple and compute-efficient method RELEX (REinforcement Learning EXtrapolation), which estimates the rank-1 subspace from a short observation window and extrapolates future checkpoints via linear regression, with no learned model required. Across three models (i.e., Qwen2.5-Math-1.5B, Qwen3-4B-Base, and Qwen3-8B-Base), RELEX produces checkpoints that match or exceed RLVR performance on both in-domain and out-of-domain benchmarks, requiring as few as 15% steps of full RLVR training. Remarkably, RELEX is able to extrapolate far beyond the observation window at no training cost, predicting checkpoints up to 10-20$\times$ beyond the observed prefix with continued improvement (e.g., observe only the first 50 steps and extrapolate to 1000 steps). Our ablation analysis confirms the minimalist sufficiency of RELEX: neither increasing the subspace rank nor employing non-linear modeling yields further gains in extrapolation. Finally, we show that RELEX's success stems from a "denoising" effect: by projecting updates onto the rank-1 subspace, the model discards stochastic optimization noise that would otherwise degrade performance during extrapolation. Our code is available at https://github.com/weizhepei/RELEX.

Reward fine-tuning has become a common approach for aligning pretrained diffusion and flow models with human preferences in text-to-image generation. Among reward-gradient-based methods, Adjoint Matching (AM) provides a principled formulation by casting reward fine-tuning as a stochastic optimal control (SOC) problem. However, AM inevitably requires a substantial computational cost: it requires (i) stochastic simulation of full generative trajectories under memoryless dynamics, resulting in a large number of function evaluations, and (ii) backward ODE simulation of the adjoint state along each sampled trajectory. In this work, we observe that both bottlenecks are closely tied to the \textit{non-trivial base drift} inherited from the pretrained model. Motivated by this observation, we propose \textbf{Efficient Adjoint Matching (EAM)}, which substantially improves training efficiency by reformulating the SOC problem with a \textit{linear base drift} and a correspondingly modified \textit{terminal cost}. This reformulation removes both sources of inefficiency; it enables training-time sampling with a few-step deterministic ODE solver and yields a closed-form adjoint solution that eliminates backward adjoint simulation. On standard text-to-image reward fine-tuning benchmarks, EAM converges up to 4x faster than AM and matches or surpasses it across various metrics including PickScore, ImageReward, HPSv2.1, CLIPScore and Aesthetics.

Existing reward alignment methods for diffusion and flow models rely on multi-step stochastic trajectories, making them difficult to extend to deterministic generators. A natural alternative is noise-space optimization, but existing approaches require backpropagation through the generator and reward pipeline, limiting applicability to differentiable settings. To address this, here we present ZeNO (Zeroth-order Noise Optimization), a gradient-free framework that formulates noise optimization as a path-integral control problem, estimable from zeroth-order reward evaluations alone. When instantiated with an Ornstein--Uhlenbeck reference process, the update connects to Langevin dynamics implicitly targeting a reward-tilted distribution. ZeNO enables effective inference-time scaling and demonstrates strong performance across diverse generators and reward functions, including a protein structure generation task where backpropagation is infeasible.

Diffusion and flow-matching models scale because pretraining is supervised regression: a clean sample is noised analytically, and a model regresses against a closed-form target. RL post-training aligns the model with a reward. In image generation, this makes samples compose objects correctly, render text legibly, and match human preferences. Existing methods rely on costly SDE rollouts, reward gradients, or surrogate losses, sacrificing pretraining's regression structure. We show that the structure extends to RL post-training. Under KL-regularized reward maximization, the optimal generative process tilts the clean-endpoint distribution towards samples with higher reward and leaves the noising law unchanged. Combining this with the adjoint-matching optimality condition and a REINFORCE identity, we derive Reinforce Adjoint Matching (RAM): a consistency loss that corrects the pretraining target with the reward. At each step, we draw a clean endpoint from the current model, evaluate its reward, noise it as in pretraining, and regress. No SDE rollouts, backward adjoint sweeps, or reward gradients are required. Like the pretraining objective, RAM is simple and scales. On Stable Diffusion 3.5M, RAM achieves the highest reward on composability, text rendering, and human preference, reaching Flow-GRPO's peak reward in up to $50\times$ fewer training steps.

While Reinforcement Learning with Verifiable Rewards (RLVR) has recently emerged as a promising post-training paradigm for Large Language Models (LLMs), its dependency on the gold label or domain-specific verifiers limits its scalability to new tasks and domains. In this work, we propose Verifier-free Intrinsic Gradient-Norm Reward (VIGOR), a simple reward that uses only the policy model itself. Given a prompt, VIGOR samples a group of completions and assigns higher within-group rewards to outputs that induce smaller $\ell_2$ norms of the teacher-forced negative log-likelihood gradients under the current parameters. Intuitively, lower gradient norms suggest the completion aligns better with the current policy, serving as an intrinsic preference signal for policy optimization. To make this intrinsic signal practical for RL, we correct the systematic length bias of averaged token-level gradients with a $\sqrt{T}$ scaling, and apply group-wise rank shaping to stabilize reward scales across prompts. Across mathematical reasoning benchmarks, VIGOR outperforms the state-of-the-art Reinforcement Learning from Internal Feedback (RLIF) baseline, and it also exhibits cross-domain transfer to code benchmarks when trained only on math data. For instance, on Qwen2.5-7B-Base post-trained on MATH, VIGOR improves the average math accuracy by +3.31% and the average code accuracy by +1.91% over this baseline, while exhibiting more stable training dynamics. The code is available at https://github.com/ZJUSCL/VIGOR.

On-policy distillation (OPD) is widely used for LLM post-training. When pushed with a reward-extrapolation coefficient lambda > 1, the student can lift past the teacher in domain, but past a threshold lambda* the same step violates the output contract on structured-output tasks. In a single-position Bernoulli reduction, we derive a closed-form base-relative clip-safety threshold lambda*(p,b,c) determined by three measurable quantities: the teacher modal probability, the warm-start mass, and the importance-sampling clip strength. Above lambda*, the extrapolated fixed point exits the clip-safe region, changing training from format-preserving to format-collapsing. We extend the rule to calibrated K-ary listwise JSON tasks where a single binding equivalence class dominates the output contract and SFT retains parse headroom. On Amazon Fashion, three pre-registered tests--a fine-grid cliff interval, a budget-extension test, and a small-clip cross-prediction--fall within their locked prediction windows, with the small-clip value matching the closed-form prediction below grid resolution. Operating just below lambda*, ListOPD brings a 1.7B Qwen3 student to in-domain parity with an 8B-SFT baseline at one-fifth the parameters. The gain is driven primarily by format adherence: NDCG@1 on parsed outputs remains flat across lambda, while parse validity sharply changes at the predicted boundary. The cliff diagnostic is rubric-independent, whereas the parity claim uses a Gemini-graded rubric and inherits that evaluator's exposure.

State-of-the-art model-based reinforcement learning methods train policies on imagined rollouts. These rollouts are trajectories generated by a learned dynamics model and are scored by a learned reward model, but without querying the true environment during policy updates. We study this training paradigm by quantifying how errors in learned dynamics and reward models affect returns and policy optimization. First, we extend the analysis of Asadi et al. (2018) to MDPs with learned reward models, and derive the optimal sample allocation--the ratio of dynamics samples to reward samples that minimizes a bound on return error under power-law scaling assumptions. We identify lower Lipschitz constants of the learned dynamics, reward, and policy as a representation desideratum that tightens this bound, and we connect this perspective to the temporal-straightening objective of Wang et al. (2026). Second, we examine how policy optimization with REINFORCE tolerates noisy rewards, which are often cheaper to obtain. We show that zero-mean reward noise leaves the gradient estimator unbiased and adds at most a variance term that decreases with the number of rollouts. This introduces a practical tradeoff: given a fixed budget, should one buy more rollouts with cheaper but noisier rewards, or fewer rollouts with more expensive but less noisy rewards? We reduce this choice to a one-dimensional optimization problem and characterize the optimum.

Offline safe reinforcement learning often requires policies to adapt at deployment time to safety budgets that vary across episodes or change within a single episode. While diffusion-based planners enable flexible trajectory generation, existing guidance schemes often treat reward improvement and constraint satisfaction as competing gradient objectives, which can lead to unreliable safety compliance under cost limits. We reinterpret adaptive safe trajectory generation as sampling from a constrained trajectory distribution, where the budget restricts the trajectory region, and reward shapes preferences within that region. This perspective motivates Safe Decoupled Guidance Diffusion (SDGD), which conditions classifier-free guidance on the cost limit to bias sampling toward trajectories satisfying the specified limit, while using reward-gradient guidance to refine trajectories for higher return. Because direct reward guidance can increase return while also steering samples toward trajectories with higher cumulative cost, we introduce Feasible Trajectory Relabeling (FTR) to reshape reward targets and discourage such directions. We further provide a first-order sampling-time analysis showing that FTR suppresses reward-induced cost drift under a prefix-restorative alignment condition. Extensive evaluations on the DSRL benchmark show that SDGD achieves the strongest safety compliance among baselines, satisfying the constraint on 94.7% of tasks (36/38), while obtaining the highest reward among safe methods on 21 tasks.

Generative Flow Networks (GFlowNets) learn to sample states proportional to an unnormalized reward. Despite their theoretical promise, practical training is often unstable, exhibiting severe loss spikes and mode collapse. To tackle this, we first assess the sensitivity of GFlowNet objectives, demonstrating that a small Total Variation (TV) distance between the learned and target distributions does not preclude unbounded training loss. Motivated by this mismatch, we establish converse guarantees by deriving loss-to-TV bounds that certify global fidelity from bounded trajectory balance losses. Lastly, we propose Stable GFlowNets, an algorithm that leverages our theoretical results to stabilize training, and empirically demonstrate improved training behavior and superior distributional fidelity.

In generative modeling, we often wish to produce samples that maximize a user-specified reward such as aesthetic quality or alignment with human preferences, a problem known as \textit{guidance}. Despite their widespread use, existing guidance methods either require expensive multi-particle, many-step schemes or rely on poorly understood approximations. We reformulate guidance as a \textit{deterministic optimal control problem}, yielding a hierarchy of algorithms that subsumes existing approaches at the coarsest level. We show that the \textit{flow map}, an object of significant recent interest for its role in fast inference, arises naturally in the optimal solution. Based on this observation, we propose \textbf{Flow Map Reward Guidance (FMRG)}: a training-free, \textit{single-trajectory} framework that uses the flow map to both integrate and guide the flow. At text-to-image scale, FMRG matches or surpasses baselines across inverse problems and reward-guided generation with \textbf{as few as 3 NFEs}, giving at least an order-of-magnitude speedup in comparison to prior state of the art. Code is available at https://github.com/jrrhuang/fmrg.

This paper focuses on the alignment of flow matching models with human preferences. A promising way is fine-tuning by directly backpropagating reward gradients through the differentiable generation process of flow matching. However, backpropagating through long trajectories results in prohibitive memory costs and gradient explosion. Therefore, direct-gradient methods struggle to update early generation steps, which are crucial for determining the global structure of the final image. To address this issue, we introduce LeapAlign, a fine-tuning method that reduces computational cost and enables direct gradient propagation from reward to early generation steps. Specifically, we shorten the long trajectory into only two steps by designing two consecutive leaps, each skipping multiple ODE sampling steps and predicting future latents in a single step. By randomizing the start and end timesteps of the leaps, LeapAlign leads to efficient and stable model updates at any generation step. To better use such shortened trajectories, we assign higher training weights to those that are more consistent with the long generation path. To further enhance gradient stability, we reduce the weights of gradient terms with large magnitude, instead of completely removing them as done in previous works. When fine-tuning the Flux model, LeapAlign consistently outperforms state-of-the-art GRPO-based and direct-gradient methods across various metrics, achieving superior image quality and image-text alignment.

Reinforcement learning has been widely applied to diffusion and flow models for visual tasks such as text-to-image generation. However, these tasks remain challenging because diffusion models have intractable likelihoods, which creates a barrier for directly applying popular policy-gradient type methods. Existing approaches primarily focus on crafting new objectives built on already heavily engineered LLM objectives, using ad hoc estimators for likelihood, without a thorough investigation into how such estimation affects overall algorithmic performance. In this work, we provide a systematic analysis of the RL design space by disentangling three factors: i) policy-gradient objectives, ii) likelihood estimators, and iii) rollout sampling schemes. We show that adopting an evidence lower bound (ELBO) based model likelihood estimator, computed only from the final generated sample, is the dominant factor enabling effective, efficient, and stable RL optimization, outweighing the impact of the specific policy-gradient loss functional. We validate our findings across multiple reward benchmarks using SD 3.5 Medium, and observe consistent trends across all tasks. Our method improves the GenEval score from 0.24 to 0.95 in 90 GPU hours, which is 4.6 times more efficient than FlowGRPO and $2\times$ more efficient than the SOTA method without reward hacking.

Amortized inference is the task of training a parametric model, such as a neural network, to approximate a distribution with a given unnormalized density where exact sampling is intractable. When sampling is implemented as a sequential decision-making process, reinforcement learning (RL) methods, such as generative flow networks, can be used to train the sampling policy. Off-policy RL training facilitates the discovery of diverse, high-reward candidates, but existing methods still face challenges in efficient exploration. We propose to use an adaptive training distribution (the \teacher) to guide the training of the primary amortized sampler (the \student). The \teacher, an auxiliary behavior model, is trained to sample high-loss regions of the \student and can generalize across unexplored modes, thereby enhancing mode coverage by providing an efficient training curriculum. We validate the effectiveness of this approach in a synthetic environment designed to present an exploration challenge, two diffusion-based sampling tasks, and four biochemical discovery tasks demonstrating its ability to improve sample efficiency and mode coverage. Source code is available at https://github.com/alstn12088/adaptive-teacher.