Delightful Distributed Policy Gradient

• URL: http://arxiv.org/abs/2603.20521v1
• Date: Fri, 20 Mar 2026 21:45:51 GMT
• Title: Delightful Distributed Policy Gradient
• Abstract summary: Distributed reinforcement learning trains on data from stale, buggy, or mismatched actors.<n>High-surprisal failures can dominate the update direction despite carrying little useful signal.<n>High-surprisal successes reveal opportunities the current policy would otherwise miss.<n>No sign-blind reweighting, including exact importance sampling, can reproduce this effect.
• Score: 1.1886634182318419
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Distributed reinforcement learning trains on data from stale, buggy, or mismatched actors, producing actions with high surprisal (negative log-probability) under the learner's policy. The core difficulty is not surprising data per se, but \emph{negative learning from surprising data}. High-surprisal failures can dominate the update direction despite carrying little useful signal, while high-surprisal successes reveal opportunities the current policy would otherwise miss. The \textit{Delightful Policy Gradient} (DG) separates these cases by gating each update with delight, the product of advantage and surprisal, suppressing rare failures and amplifying rare successes without behavior probabilities. Under contaminated sampling, the cosine similarity between the standard policy gradient and the true gradient collapses, while DG's grows as the policy improves. No sign-blind reweighting, including exact importance sampling, can reproduce this effect. On MNIST with simulated staleness, DG without off-policy correction outperforms importance-weighted PG with exact behavior probabilities. On a transformer sequence task with staleness, actor bugs, reward corruption, and rare discovery, DG achieves roughly $10{\times}$ lower error. When all four frictions act simultaneously, its compute advantage is order-of-magnitude and grows with task complexity.

Related papers

• When to Stop Reusing: Dynamic Gradient Gating for Sample-Efficient RLVR [46.88928401641895]
  Reinforcement Learning with Verifiable Rewards (RLVR) has become the dominant paradigm for advanced reasoning in Large Language Models (LLMs)<n>In RLVR, reuse each rollout batch for multiple gradient updates amplifies policy shift, leading to severe performance degradation.<n>We identify the textitDisproportionate Weight Divergence (DWD) phenomenon: performance degradation is synchronized a sharp surge in the textttlm_head weight change, while intermediate layers remain stable.<n>We propose textitDynamic Gradient Gating (DGG), a lightweight intervention
  arXiv  Detail & Related papers  (2026-05-19T06:23:43Z)
• Trust the Batch, On- or Off-Policy: Adaptive Policy Optimization for RL Post-Training [50.86545293331458]
  Reinforcement learning is structurally harder than supervised learning because the policy changes the data distribution it learns from.<n>We present a simple yet effective batch-adaptive objective that replaces fixed clipping with the normalized effective sample size of the policy ratios.
  arXiv  Detail & Related papers  (2026-05-12T16:44:47Z)
• Delightful Policy Gradient [1.1886634182318419]
  Standard policy weight each sampled action by advantage alone, regardless of how likely that action was under the current policy.<n>We introduce the textitDelightful Policy Gradient (DG) which gates each term with a sigmoid of emphdelight, the product of advantage and action surprisal (negative log-probability)<n>For $K$-armed bandits, DG provably improves directional accuracy in a single context and, across multiple contexts, shifts the expected gradient strictly closer to the supervised cross-entropy oracle.
  arXiv  Detail & Related papers  (2026-03-15T21:06:37Z)
• From $\boldsymbol{\logπ}$ to $\boldsymbolπ$: Taming Divergence in Soft Clipping via Bilateral Decoupled Decay of Probability Gradient Weight [6.07556923512707]
  Reinforcement Learning with Verifiable Rewards (RLVR) has catalyzed a leap in Large Language Model (LLM) reasoning, yet its optimization dynamics remain fragile.<n>Standard algorithms like GRPO enforce stability via hard clipping'', which inadvertently stifles exploration by discarding gradients of tokens outside the trust region.<n>We propose Decoupled Gradient Policy Optimization (DGPO), which employs a decoupled decay mechanism based on importance sampling ratios.
  arXiv  Detail & Related papers  (2026-03-15T14:00:48Z)
• Typicalness-Aware Learning for Failure Detection [26.23185979968123]
  Deep neural networks (DNNs) often suffer from the overconfidence issue, where incorrect predictions are made with high confidence scores. We propose a novel approach called Typicalness-Aware Learning (TAL) to address this issue and improve failure detection performance.
  arXiv  Detail & Related papers  (2024-11-04T11:09:47Z)
• Beyond Myopia: Learning from Positive and Unlabeled Data through Holistic Predictive Trends [26.79150786180822]
  We unveil an intriguing yet long-overlooked observation in PUL. Predictive trends for positive and negative classes display distinctly different patterns. We propose a novel TPP-inspired measure for trend detection and prove its unbiasedness in predicting changes.
  arXiv  Detail & Related papers  (2023-10-06T08:06:15Z)
• Unleashing the Power of Graph Data Augmentation on Covariate Distribution Shift [50.98086766507025]
  We propose a simple-yet-effective data augmentation strategy, Adversarial Invariant Augmentation (AIA) AIA aims to extrapolate and generate new environments, while concurrently preserving the original stable features during the augmentation process.
  arXiv  Detail & Related papers  (2022-11-05T07:55:55Z)
• Clipped Stochastic Methods for Variational Inequalities with Heavy-Tailed Noise [64.85879194013407]
  We prove the first high-probability results with logarithmic dependence on the confidence level for methods for solving monotone and structured non-monotone VIPs. Our results match the best-known ones in the light-tails case and are novel for structured non-monotone problems. In addition, we numerically validate that the gradient noise of many practical formulations is heavy-tailed and show that clipping improves the performance of SEG/SGDA.
  arXiv  Detail & Related papers  (2022-06-02T15:21:55Z)
• Policy Gradient for Continuing Tasks in Non-stationary Markov Decision Processes [112.38662246621969]
  Reinforcement learning considers the problem of finding policies that maximize an expected cumulative reward in a Markov decision process with unknown transition probabilities. We compute unbiased navigation gradients of the value function which we use as ascent directions to update the policy. A major drawback of policy gradient-type algorithms is that they are limited to episodic tasks unless stationarity assumptions are imposed.
  arXiv  Detail & Related papers  (2020-10-16T15:15:42Z)
• Unbiased Risk Estimators Can Mislead: A Case Study of Learning with Complementary Labels [92.98756432746482]
  We study a weakly supervised problem called learning with complementary labels. We show that the quality of gradient estimation matters more in risk minimization. We propose a novel surrogate complementary loss(SCL) framework that trades zero bias with reduced variance.
  arXiv  Detail & Related papers  (2020-07-05T04:19:37Z)
• Evaluating Prediction-Time Batch Normalization for Robustness under Covariate Shift [81.74795324629712]
  We call prediction-time batch normalization, which significantly improves model accuracy and calibration under covariate shift. We show that prediction-time batch normalization provides complementary benefits to existing state-of-the-art approaches for improving robustness. The method has mixed results when used alongside pre-training, and does not seem to perform as well under more natural types of dataset shift.
  arXiv  Detail & Related papers  (2020-06-19T05:08:43Z)
• Zeroth-order Deterministic Policy Gradient [116.87117204825105]
  We introduce Zeroth-order Deterministic Policy Gradient (ZDPG) ZDPG approximates policy-reward gradients via two-point evaluations of the $Q$function. New finite sample complexity bounds for ZDPG improve upon existing results by up to two orders of magnitude.
  arXiv  Detail & Related papers  (2020-06-12T16:52:29Z)

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.