The authors formalize the connection between GRPO's analytic update structure and activation-state fidelity as policy-update exposure.
HiLoRe allocates GRPO activations between storage, compression, and recomputation using update exposure.
The method achieves up to 13.5% faster actor updates than gradient checkpointing at comparable memory while keeping downstream score differences below 0.6 percentage points.
Research Lab
Xinrui Chen · Mengyang Li · Ou Wu · Ji Zhang
Hangzhou Institute for Advanced Study, University of Chinese Academy of Sciences · Tianjin Normal University · University of Southern Queensland
Research Digest··2 min read
The authors propose HiLoRe, an activation-management method for GRPO training that uses the policy's analytic update coefficients to decide which states to store at high precision, which to compress, and which to recompute.
Why this paper
From Hangzhou Institute for Advanced Study, University of Chinese Academy of Sciences and 2 others
In one line
HiLoRe accelerates GRPO training by allocating activations among precise storage, low-precision compression, and recomputation according to their current update sensitivity.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§