AI

MEND: RL For Flow Models via Proximal Velocity Matching

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 07/10/2026, 06:18:25 UTC.

Briefings

MEND, a reinforcement learning method for flow models, uses proximal velocity matching to optimize training. It caps rewards per prompt group to prevent overfitting and only accepts moves exceeding a quadratic displacement threshold. In 100 updates, MEND outperformed Flow-GRPO on five of six evaluators and surpassed ReFL and DiffusionNFT under an equal-budget protocol, reaching a PickScore of 24.03. The approach applies to any flow backbone with differentiable reward functions.

Key points

• MEND uses proximal velocity matching to optimize flow models by capping rewards within prompt groups and accepting moves only when the capped reward gain exceeds a quadratic displacement price.
• In 100 updates, MEND outperformed Flow-GRPO on five of six evaluators and surpassed ReFL and DiffusionNFT under an equal-budget protocol, reaching a PickScore of 24.03.

Context

MEND addresses issues with reward post-training in flow models by introducing proximal velocity matching. This method caps rewards to prevent overfitting and only accepts moves that exceed a specific displacement threshold, ensuring efficient optimization without unnecessary KL penalties or frozen reference models.

Limitations

The study covers 100 updates and a 300-update three-reward run. Results are based on five of six evaluators and four training rewards under an equal-budget protocol. No additional background context is supplied beyond the provided evidence.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 25/10/2026, 13:10:55 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeide42qb6pc3sd5casoztlz6l7xlbwcuwbsimnmdam7uoac7l5yyei