html,body{height:auto!important;overflow:auto!important} [data-requires-script]{display:none!important} .reader{position:relative!important;overflow:visible!important;min-height:100vh} .reader-frame{height:auto!important;max-height:none!important;margin:auto} .reader-stage,.lesson-reader-stage{position:relative!important;inset:auto!important;overflow:visible!important} .reader-track,.lesson-track{position:relative!important;inset:auto!important;display:block!important;height:auto!important;transform:none!important} .reader-slide,.reader-panel,.lesson-panel{width:100%!important;height:auto!important;min-height:0!important;overflow:visible!important;content-visibility:visible!important} .reader-scroll,.lesson-panel-scroll{position:relative!important;inset:auto!important;height:auto!important;overflow:visible!important;padding-bottom:32px!important} .reader-tabs,.reader-tap,.reader-detail-dock,.card-dock{display:none!important} .lg-screen,.lg-feed,.lg-page,.dr-screen .lg-page{height:auto!important;overflow:visible!important} .lg-page{min-height:100vh} @media(min-width:1000px){.lg-page{min-height:0!important}}

Author

Tong Zhang

Papers that credit Tong Zhang in their source bylines. Matching names may refer to different researchers.

1 results

PaperMachine Learning
A token-wise update keeps stale-response training aimed at its target
Proximal KL has higher math response accuracy than GRPO in seven tested off-policy comparisons across two Qwen3 sizes.
Added to Chiosa Oct 14, 2026
#cost-performance #off-policy-correction #kl-divergence

Tong Zhang

A token-wise update keeps stale-response training aimed at its target