GrowMTP: Can RL Grow Its Own Draft Head?

简体中文

L1 mechanism diagnostic. Depth-coupled acceptance loss, without demonstrated GPU speedup.

Paper information

Field	Value
Paper	GrowMTP: Can RL Grow Its Own Draft Head?
Affiliations	WeChat AI, Tencent / Peking University
First public release	2026-09-15 (arXiv v1)
Authors' code	No public author repository found as of 2026-09-16.
Adapter / method	growmtp
Local code	src/auto_research/post_training/latest_20260916.py

Original paper

Background and contribution

During RL rollouts, train the multi-token prediction head by accepted draft depth and decouple its gradient from the backbone.

flowchart LR
  I[RL draft and verification] --> D[Accepted depth] --> L[Depth-coupled MTP loss] --> O[Head update without backbone gradient]

Key figure from the paper

Original Figure 2: rollout and draft-head training. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism

The reference retains decisive gating, weighting, and state changes, exposing depth and other auditable quantities. Fixed budget and three seeds test invariants, not paper-scale training.

Paper results

Paper results on speed, cost, and capability remain separate from local measurements.

Local reproduction

Three-seed results are diagnostic_only=true and do not enter the public dashboard as full capability gains.

Reproduction limits

Only acceptance loss was validated; no GPU speedup claim.