Stabilizing Reinforcement Learning for Diffusion Language Models

• URL: http://arxiv.org/abs/2603.06743v1
• Date: Fri, 06 Mar 2026 08:55:37 GMT
• Title: Stabilizing Reinforcement Learning for Diffusion Language Models
• Abstract summary: Group Relative Policy Optimization (GRPO) is highly effective for post-training autoregressive (AR) language models.<n>But its direct application to diffusion large language models (dLLMs) often triggers reward collapse.<n>We show these effects form a self-reinforcing instability loop that drives policy drift and further increases ratio variance.<n>We propose StableDRL, a reformulation of GRPO tailored for dLLMs.
• Score: 27.22260417844743
• License: http://creativecommons.org/licenses/by-nc-sa/4.0/
• Abstract: Group Relative Policy Optimization (GRPO) is highly effective for post-training autoregressive (AR) language models, yet its direct application to diffusion large language models (dLLMs) often triggers reward collapse. We identify two sources of incompatibility. First, GRPO relies on importance ratios defined by sequence probabilities, which are intractable in dLLMs and must be estimated (e.g., via ELBO-based or mean-field likelihood proxies), yielding inherently noisy ratios. Second, standard GRPO's formulation is not designed for estimated ratios: its conditional clipping can be anomalously bypassed by model-agnostic estimation noise, producing gradient spikes, while its fixed group-size normalization amplifies gradient-magnitude fluctuations under high-variance ratio estimates. We show these effects form a self-reinforcing instability loop that drives policy drift and further increases ratio variance. To break this loop, we propose StableDRL, a reformulation of GRPO tailored for dLLMs that uses (i) unconditional clipping to suppress outlier-induced spikes and (ii) self-normalization to constrain updates within the convex hull of per-sample gradients. We further extend StableDRL to block-wise diffusion models via a staircase attention mechanism.

Related papers

• Hacking Generative Perplexity: Why Unconditional Text Evaluation Needs Distributional Metrics [49.443264461057645]
  Diffusion and continuous flow-based language models have emerged as the leading non-autoregressive alternatives to language modeling.<n>Progress in both paradigms is overwhelmingly tracked by generative perplexity (gen-PPL): the per-token negative log-likelihood of samples under a frozen autoregressive (AR) scorer.<n>We argue that this metric is unsound. By construction, gen-PPL measures only predictability under the scoring AR, not grammaticality or semantic coherence.
  arXiv  Detail & Related papers  (2026-06-07T02:35:56Z)
• One-Way Policy Optimization for Self-Evolving LLMs [63.8638342097375]
  Reinforcement Learning with Verifiable Rewards (RLVR) has become a promising paradigm for scaling reasoning capabilities of Large Language Models (LLMs)<n>We propose One-Way Policy Optimization (OWPO), a method based on the principle of decoupling optimization direction from update magnitude.<n> Experimental results demonstrate that OWPO outperforms strong baselines, including DAPO, OPD, and MOPD.
  arXiv  Detail & Related papers  (2026-05-21T08:25:27Z)
• KL for a KL: On-Policy Distillation with Control Variate Baseline [13.281263788199219]
  On-Policy Distillation (OPD) has emerged as a dominant post-training paradigm for large language models.<n>We propose vOPD, which casts OPD as policy-gradient RL.<n>We show that vOPD consistently outperforms vanilla OPD and matches the most expensive full-vocabulary baseline.
  arXiv  Detail & Related papers  (2026-05-08T15:24:51Z)
• Too Correct to Learn: Reinforcement Learning on Saturated Reasoning Data [55.84428098924793]
  We propose Constrained Uniform Top-K Sampling (CUTS), a parameter-free decoding strategy enforcing structure-preserving exploration.<n>We integrate this into Mixed-CUTS, a training framework synergizing exploitative and exploratory rollouts to amplify intra-group advantage variance.<n> Notably, Mixed-CUTS improves Pass@1 accuracy on the challenging AIME25 benchmark by up to 15.1% over standard GRPO.
  arXiv  Detail & Related papers  (2026-04-20T16:43:28Z)
• Improving Search Agent with One Line of Code [68.58667107354253]
  Tool-based Agentic Reinforcement Learning (TARL) has emerged as a promising paradigm for training search agents to interact with external tools.<n>We propose textbfSearch textbfAgent textbfPolicy textbfOptimization (textbfSAPO), which stabilizes training via a conditional token-level KL constraint.
  arXiv  Detail & Related papers  (2026-03-10T04:07:39Z)
• Diffusion Controller: Framework, Algorithms and Parameterization [54.82539154511621]
  We present a unified control-theoretic view that casts reverse diffusion sampling as state-only control within (generalized) linearly-solvable Markov Decision Processes.<n>Under this framework, control acts by reweighting the pretrained reverse-time transition kernels, balancing terminal objectives against an $f$-divergence cost.<n>Experiments on Stable Diffusion v1.4 show consistent gains in preference-alignment win rates and improved quality-efficiency trade-offs.
  arXiv  Detail & Related papers  (2026-03-07T01:49:59Z)
• LFPO: Likelihood-Free Policy Optimization for Masked Diffusion Models [48.68246945083386]
  Likelihood-Free Policy Optimization (LFPO) is a native framework that maps the concept of vector field flow matching to the discrete token space.<n>LFPO formulates alignment as geometric velocity rectification, which directly optimize denoising logits via contrastive updates.<n>Experiments demonstrate that LFPO not only outperforms state-of-the-art baselines on code and reasoning benchmarks but also accelerates inference by approximately 20% through reduced diffusion steps.
  arXiv  Detail & Related papers  (2026-03-02T07:42:55Z)
• Rethinking the Trust Region in LLM Reinforcement Learning [72.25890308541334]
  Proximal Policy Optimization (PPO) serves as the de facto standard algorithm for Large Language Models (LLMs)<n>We propose Divergence Proximal Policy Optimization (DPPO), which substitutes clipping with a more principled constraint.<n>DPPO achieves superior training and efficiency compared to existing methods, offering a more robust foundation for RL-based fine-tuning.
  arXiv  Detail & Related papers  (2026-02-04T18:59:04Z)
• Goal inference with Rao-Blackwellized Particle Filters [5.633221187382381]
  Inferring the eventual goal of a mobile agent from noisy observations of its trajectory is a fundamental estimation problem.<n>We study such intent inference using a variant of a Rao-Blackwellized Particle Filter (RBPF)<n>We quantify how well the adversary can recover the agent's intent using information-theoretic leakage metrics.
  arXiv  Detail & Related papers  (2025-12-10T02:48:55Z)
• Principled RL for Diffusion LLMs Emerges from a Sequence-Level Perspective [85.06838178922791]
  Reinforcement Learning (RL) has proven highly effective for autoregressive language models.<n>But adapting these methods to diffusion large language models (dLLMs) presents fundamental challenges.<n>We propose a principled RL framework that treats entire sequence generation as a single action and uses the ELBO as a tractable sequence-level likelihood proxy.
  arXiv  Detail & Related papers  (2025-12-03T13:05:32Z)
• Improving Reasoning for Diffusion Language Models via Group Diffusion Policy Optimization [25.504794432255306]
  Diffusion language models (DLMs) enable parallel, order-agnostic generation with iterative refinement.<n>Adapting reinforcement learning fine-tuning to DLMs remains an open challenge because of the intractable likelihood.<n>We introduce textbfGroup Diffusion Policy Optimization (GDPO), a new RL algorithm tailored for DLMs.
  arXiv  Detail & Related papers  (2025-10-09T17:58:07Z)
• Reinforce-Ada: An Adaptive Sampling Framework for Reinforce-Style LLM Training [47.26632817047513]
  Reinforcement learning applied to large language models (LLMs) for reasoning tasks is often bottlenecked by unstable gradient estimates.<n>We propose Reinforce-Ada, an adaptive sampling framework for online RL post-training of LLMs.<n>Unlike conventional two-stage allocation methods, Reinforce-Ada interleaves estimation and sampling in an online successive elimination process.
  arXiv  Detail & Related papers  (2025-10-06T16:34:09Z)
• Uncalibrated Reasoning: GRPO Induces Overconfidence for Stochastic Outcomes [55.2480439325792]
  Reinforcement learning (RL) has proven remarkably effective at improving the accuracy of language models in verifiable and deterministic domains like mathematics.<n>Here, we examine if current RL methods are also effective at optimizing language models in verifiable domains with outcomes, like scientific experiments.
  arXiv  Detail & Related papers  (2025-08-15T20:50:53Z)
• Inference-Time Scaling of Diffusion Language Models with Particle Gibbs Sampling [70.8832906871441]
  We study how to steer generation toward desired rewards without retraining the models.<n>Prior methods typically resample or filter within a single denoising trajectory, optimizing rewards step-by-step without trajectory-level refinement.<n>We introduce particle Gibbs sampling for diffusion language models (PG-DLM), a novel inference-time algorithm enabling trajectory-level refinement while preserving generation perplexity.
  arXiv  Detail & Related papers  (2025-07-11T08:00:47Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.