大模型后训练，破解多模态分布性遗忘！7B模型七项基准全线提升
Ranking
Observed public metrics from 1 member.
Merged summary
TL;DR - Remember-R1 (Northwestern Polytechnical Univ., HKUST, Zhejiang Univ.) is an RL post-training method that adds three process rewards to counter "distributional visual forgetting," where multimodal models drift from image evidence toward pure-text reasoning over long chains. It matters because outcome-only rewards can't distinguish visually grounded trajectories from lucky language-prior shortcuts.
- Frames the problem as trajectory unidentifiability: modality asymmetry (fixed visual condition vs. growing text state), sparse credit assignment, and multiple correct-answer paths let RL reinforce "right answer, wrong evidence" rollouts.
- Three process rewards on top of answer correctness, trained with GRPO: visual-vocabulary coverage (weighted by how late visual terms persist), visual-memory reward (comparing last-layer attention to visual tokens early vs. late in the chain), and key-region reward (attention share on question-relevant regions, weighted toward later steps).
- Training set of 38,657 filtered samples annotated with visual keywords and key regions — used as reward anchors, not extra visual knowledge injection.
- Reported gains across seven benchmarks for the 7B model: MathVista 62.30→69.80 (+7.50), MMVet 59.44→72.37 (+12.93), RealWorldQA 69.28→69.67, suggesting perception isn't sacrificed for reasoning.
Sources (1)
大模型后训练，破解多模态分布性遗忘！7B模型七项基准全线提升
TL;DR - Remember-R1 (Northwestern Polytechnical Univ., HKUST, Zhejiang Univ.) is an RL post-training method that adds three process rewards to counter "distributional visual forgetting," where multimodal models drift from image evidence toward pure-text reasoning over long chains. It matters because outcome-only rewards can't distinguish visually grounded trajectories from lucky language-prior shortcuts.
- Frames the problem as trajectory unidentifiability: modality asymmetry (fixed visual condition vs. growing text state), sparse credit assignment, and multiple correct-answer paths let RL reinforce "right answer, wrong evidence" rollouts.
- Three process rewards on top of answer correctness, trained with GRPO: visual-vocabulary coverage (weighted by how late visual terms persist), visual-memory reward (comparing last-layer attention to visual tokens early vs. late in the chain), and key-region reward (attention share on question-relevant regions, weighted toward later steps).
- Training set of 38,657 filtered samples annotated with visual keywords and key regions — used as reward anchors, not extra visual knowledge injection.
- Reported gains across seven benchmarks for the 7B model: MathVista 62.30→69.80 (+7.50), MMVet 59.44→72.37 (+12.93), RealWorldQA 69.28→69.67, suggesting perception isn't sacrificed for reasoning.