Search

直接匹配不多,已補上最新動態。

Tag: #gradloc1 results

GradLoc Locates RLVR Crash Tokens

GradLoc Locates RLVR Crash Tokens

Tencent Hunyuan's GradLoc tool pinpoints gradient spikes to specific tokens in RLVR training, ending guesswork for collapses. Provides infrastructure for observable RL dynamics in high-noise systems. Open-sourced GitHub lowers engineering barriers for mechanism research.

机器之心MediaFeb 14#research#hunyuan#gradloc