SPG: Sandwiched Policy Gradient for Masked Diffusion Language Models

• URL: http://arxiv.org/abs/2510.09541v2
• Date: Mon, 13 Oct 2025 01:58:46 GMT
• Title: SPG: Sandwiched Policy Gradient for Masked Diffusion Language Models
• Abstract summary: Sandwiched Policy Gradient (SPG)<n>Experiments show that SPG significantly outperforms baselines based on ELBO or one-step estimation.<n>Specifically, SPG improves the accuracy over state-of-the-art RL methods for dLLMs by 3.6% in GSM8K, 2.6% in MATH500, 18.4% in Countdown and 27.0% in Sudoku.
• Score: 43.729576865473625
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Diffusion large language models (dLLMs) are emerging as an efficient alternative to autoregressive models due to their ability to decode multiple tokens in parallel. However, aligning dLLMs with human preferences or task-specific rewards via reinforcement learning (RL) is challenging because their intractable log-likelihood precludes the direct application of standard policy gradient methods. While prior work uses surrogates like the evidence lower bound (ELBO), these one-sided approximations can introduce significant policy gradient bias. To address this, we propose the Sandwiched Policy Gradient (SPG) that leverages both an upper and a lower bound of the true log-likelihood. Experiments show that SPG significantly outperforms baselines based on ELBO or one-step estimation. Specifically, SPG improves the accuracy over state-of-the-art RL methods for dLLMs by 3.6% in GSM8K, 2.6% in MATH500, 18.4% in Countdown and 27.0% in Sudoku.

Related papers

• DISA: Offline Importance Sampling for Distribution-Matching LLM-RL [56.9445657766829]
  We introduce DISA, which moves this calibration problem outside the RL loop.<n> DISA draws proposal trajectories offline, estimates the partition function via importance sampling, and freezes the resulting partition-function estimate.<n>On two open-weight backbones across six math and three code benchmarks, DISA matches or exceeds the online-coupled distribution-matching baseline Flow.
  arXiv  Detail & Related papers  (2026-05-17T07:14:44Z)
• fg-expo: Frontier-guided exploration-prioritized policy optimization via adaptive kl and gaussian curriculum [11.537163059885687]
  We propose FG-ExPO, short for Frontier-Guided Exploration-Prioritized Policy Optimization.<n> Accuracy-Conditioned KL Scaling (AKL) adjusts the KL penalty strength through a smooth nonlinear function of batch average accuracy.<n>We conduct evaluations on DeepSeek-R1-Distill-Qwen-1.5B and Qwen3-8B-Base across six mainstream mathematical reasoning benchmarks.
  arXiv  Detail & Related papers  (2026-05-12T01:48:48Z)
• expo: Exploration-prioritized policy optimization via adaptive kl regulation and gaussian curriculum sampling [11.537163059885687]
  We propose Exploration-Prioritized Policy Optimization (EXPO) with two lightweight plug-in modules.<n>We conduct experiments on DeepSeek-R1-Distill-Qwen-1.5B and Qwen3-8B-Base over six mathematical reasoning benchmarks.<n>It obtains an absolute gain of 13.34 on AIME 2025 pass@32, rising from 63.33 percent to 76.67 percent, and achieves an average pass@32 improvement of 2.66 on the 8B model.
  arXiv  Detail & Related papers  (2026-05-11T03:19:04Z)
• V-GRPO: Online Reinforcement Learning for Denoising Generative Models Is Easier than You Think [90.69263509098948]
  We introduce Variational GRPO, a method that integrates ELBO-based surrogates with the Group Relative Policy Optimization algorithm.<n>V-GRPO achieves state-of-the-art performance in text-to-image synthesis, while delivering a $2times$ speedup over MixGRPO and a $3times$ speedup over DiffusionNFT.
  arXiv  Detail & Related papers  (2026-04-25T17:03:21Z)
• 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)
• 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)
• Boundary-Guided Policy Optimization for Memory-efficient RL of Diffusion Large Language Models [53.339700196282905]
  A key challenge in applying reinforcement learning to large language models (dLLMs) is the intractability of their likelihood functions.<n>We propose a memory-efficient RL algorithm that maximizes a specially constructed lower bound of the ELBO-based objective.<n> Experiments show that BGPO significantly outperforms previous RL algorithms for dLLMs in math problem solving, code generation, and planning tasks.
  arXiv  Detail & Related papers  (2025-10-13T17:47:50Z)
• 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)
• Principled and Tractable RL for Reasoning with Diffusion Language Models [0.0]
  Diffusion large language models (dLLMs) are trained to predict multiple tokens in parallel and generate text via iterative unmasking.<n>Recent works have successfully pretrained dLLMs to parity with autoregressive LLMs at the 8B scale, but dLLMs have yet to benefit from modern post-training techniques.<n>We present Amortized Group Relative Policy Optimization (AGRPO), a principled on-policy RL algorithm designed specifically for dLLMs.
  arXiv  Detail & Related papers  (2025-10-05T03:53:16Z)
• DiffusionNFT: Online Diffusion Reinforcement with Forward Process [99.94852379720153]
  Diffusion Negative-aware FineTuning (DiffusionNFT) is a new online RL paradigm that optimize diffusion models directly on the forward process via flow matching.<n>DiffusionNFT is up to $25times$ more efficient than FlowGRPO in head-to-head comparisons, while being CFG-free.
  arXiv  Detail & Related papers  (2025-09-19T16:09:33Z)
• wd1: Weighted Policy Optimization for Reasoning in Diffusion Language Models [15.638885149395657]
  Intractability of dLLMs likelihood function requires approximating the current, old, and reference policy likelihoods at each policy optimization step.<n>We introduce $mathttwd1$, a novel policy optimization approach that reformulates the objective as a weighted likelihood.<n>Experiments on widely used reasoning benchmarks demonstrate that $mathttwd1$, without supervised fine-tuning (SFT) or any supervised data, outperforms existing RL methods for dLLMs.
  arXiv  Detail & Related papers  (2025-07-07T21:27:25Z)
• Pre-Trained Policy Discriminators are General Reward Models [81.3974586561645]
  We propose a scalable pre-training method named Policy Discriminative Learning (POLAR)<n>POLAR trains a reward model (RM) to discern identical policies and discriminate different ones.<n> Empirical results show that POLAR substantially outperforms traditional non-pre-trained methods.
  arXiv  Detail & Related papers  (2025-07-07T16:56:31Z)
• Token-Efficient RL for LLM Reasoning [0.02488650627593658]
  We propose reinforcement learning strategies tailored for reasoning in large language models (LLMs) under strict memory and compute limits.<n>Building on early policy gradient methods with baseline subtraction, we design critic-free methods that operate on a small, informative subset of output tokens.<n>We show that our methods raise accuracy on the SVAMP benchmark from 46% to over 70% and show strong performance on multi-digit multiplication.
  arXiv  Detail & Related papers  (2025-04-29T14:58:43Z)
• Zeroth-Order Policy Gradient for Reinforcement Learning from Human Feedback without Reward Inference [15.038210624870656]
  Reward inference is a critical intermediate step in the Reinforcement Learning from Human Feedback pipeline.<n>This paper develops two RLHF algorithms without reward inference for general RL problems beyond bandits and deterministic MDP bandit, and general preference models beyond the Bradley-Terry model.
  arXiv  Detail & Related papers  (2024-09-25T22:20:11Z)
• Self-Play Preference Optimization for Language Model Alignment [75.83359213697854]
  Recent advancements suggest that directly working with preference probabilities can yield a more accurate reflection of human preferences. We propose a self-play-based method for language model alignment, which treats the problem as a constant-sum two-player game. Our approach, dubbed Self-Play Preference Optimization (SPPO), utilizes iterative policy updates to provably approximate the Nash equilibrium.
  arXiv  Detail & Related papers  (2024-05-01T17:59:20Z)

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.