跳到正文
HuggingFace Daily Papers(社區熱門論文)·· 24 天前AI 評分43

LLM 強化學習中訓練與推理不匹配的起源與演變

On the Origin and Evolution of Training-Inference Mismatch in LLM Reinforcement Learning

AI 導讀

這項研究追蹤 LLM 強化學習中訓練與推理不匹配的來源,並提出差異—滯後校正(DLC)以穩定訓練。研究將不匹配拆分為引擎數值差異和策略滯後,並在有效樣本量低於一半時重新同步取樣器。DLC 在 Qwen3-0.6B 和 Qwen3-1.7B 上,將與 on-policy 訓練的 KL 散度維持在兩次獨立 on-policy 訓練之間 KL 散度的兩倍以內。

正文

Published on Oct 29

·

Submitted by

yu li

on Nov 2

Authors:

,

Abstract

On-policy reinforcement learning for large language models suffers from training-inference mismatch due to numerical discrepancies and policy lag, and a derived discrepancy-lag correction stabilizes training by centering errors, clipping importance ratios, and resynchronizing sampling.

On-policy reinforcement learning for large language models assumes that training responses are sampled from the current policy. In practice, an inference engine samples the responses and a separate trainer computes the gradients, and the training-inference mismatch between the two makes the training data off-policy. The mismatch has two sources, the numerics of the engine and the stale weights of asynchronous sampling, and a correction that stabilizes training under one source can fail under the other. We ask which property of a mismatch source decides the correction it needs. Evaluating the trainer at the sampler's weights splits each sampled-token error exactly into a numerical discrepancy of the engine and a policy lag of the trainer, and we follow both through training. The numerical discrepancy stays approximately constant, so it biases the updates by a nearly fixed amount, and centering the score at each prefix removes its leading bias. The policy lag is produced by the trainer's own updates: the bias it leaves enters the next update, which enlarges the next lag and closes a feedback loop. Importance weighting cuts this loop only while the lag ratio stays concentrated, and on Qwen3-1.7B it fails at the step where the effective sample size of the ratio falls below one half. From this account we derive the discrepancy-lag correction (DLC), which centers the numerical discrepancy, clips the lag ratio against the sampler's weights and resynchronizes the sampler when this effective sample size falls below one half. Under a stale sampler, DLC keeps the KL divergence from on-policy training within twice the KL between two independent on-policy trainings on Qwen3-0.6B and Qwen3-1.7B, and its resynchronization prevents collapses that the per-token corrections leave and that a KL measured on the sampler's prefixes does not detect.

View arXiv page View PDF GitHub 2 Add to collection

Get this paper in your agent:

hf papers read 2610.40945

Don't have the latest CLI?

curl -LsSf https://hf.co/cli/install.sh | bash

Models citing this paper 0

No model linking this paper

Cite arxiv.org/abs/2610.40945 in a model README.md to link it from this page.

Datasets citing this paper 0

No dataset linking this paper

Cite arxiv.org/abs/2610.40945 in a dataset README.md to link it from this page.

Spaces citing this paper 0

No Space linking this paper

Cite arxiv.org/abs/2610.40945 in a Space README.md to link it from this page.

Collections including this paper 0

No Collection including this paper

Add this paper to a collection to link it from this page.

來源:HuggingFace Daily Papers(社區熱門論文) · huggingface.co