Reframe: NJU's Two-Stage Post-Training Framework Aligns Visual Agents' Observations, Beats Mix-RL and Vanilla MOPD
NJU · hf · 2026-11-10
Researchers from Nanjing University and CAS present Reframe, a two-stage post-training framework for agentic visual reasoning that treats the observation state—not just the output distribution—as the alignment target. Stage one trains spatial/temporal specialists via Ground-Drop-GRPO (GD-GRPO) with evidence-aware rewards, fixing the failure mode of invoking the right tool but grounding actions on wrong observations. Stage two applies observation-gated on-policy distillation from multiple teachers. Reported to outperform Mix-RL and vanilla MOPD on both image and video benchmarks.
More from Research
- Terence Tao says AI Navier–Stokes proof is correct but mathematicians 'learned almost nothing' — sudoraohacker · 2026-11-10
- Must-read papers of the week: VFlash, GR00T 2, agent guardrails, AI-scientist benchmarks — TheTuringPost · 2026-11-10
- Hugging Face open-sources 14.8B DNA model Carbon-X, second only to GPT-6.1 Sol on GenomeQA — LoubnaBenAllal1 · 2026-11-10
- Agent swarm experiment: prompt name order and model choice decide who leads — bibryam · 2026-11-10
- New preprint frames Bayesian experimental design under misspecification as a max-min game — canaesseth · 2026-11-10
- François Fleuret asks: what are the key modern features of SOTA transformers? — francoisfleuret · 2026-11-10