PulseAugur
EN
LIVE 21:09:12

New Proximal KL method improves off-policy RLVR accuracy on Qwen3 models

By PulseAugur Editorial · [1 sources] ·

Researchers have developed Proximal KL, a novel method for off-policy reinforcement learning with verifiable rewards (RLVR). This technique addresses challenges like asynchronous generation and response reuse by using token-wise importance weighting and bounded future KL credit, avoiding the need for full-response weights or a fitted critic. In comparisons against GRPO on Qwen3-1.7B and Qwen3-4B models across eight math benchmarks, Proximal KL demonstrated superior accuracy, outperforming GRPO by up to 2.58% in micro-averaged accuracy. AI

IMPACT This new method could enhance the efficiency and accuracy of training AI models, particularly in scenarios involving off-policy data.

RANK_REASON The cluster contains a research paper detailing a new method for reinforcement learning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on Hugging Face Daily Papers →

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →

New Proximal KL method improves off-policy RLVR accuracy on Qwen3 models

How we ranked this

Signal score
1 / 100
Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket
Tool
The cluster contains a research paper detailing a new method for reinforcement learning. [lever_c_demoted from research: ic=1 ai=1.0]
Source corroboration
Single-source cluster
Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
Topics
paper, model release
Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance
High
Clearly on-topic for AI-industry coverage.
Story freshness
1 days old
Coverage has settled into its steady-state source set.

Full methodology in our editorial standards.

COVERAGE [1]

  1. Hugging Face Daily Papers TIER_1 English(EN) ·

    Enabling Token-wise Importance Sampling in Off-Policy RLVR with Proximal KL Regularization

    Off-policy data in reinforcement learning with verifiable rewards (RLVR) can arise from asynchronous generation, rollout-training mismatch, and response reuse. Token-wise importance weighting avoids multiplying likelihood ratios across a response but generally biases policy gradi…

RELATED ENTITIES

RELATED TOPICS