NV-CoT: predicting continuous coordinates lifts visual grounding IoU by 12.2 points (NeurIPS 2026 Oral)
jiqizhixin · x · 2026-11-11
Numerical Visual Chain-of-Thought (NV-CoT), from NTU, USTC, NUS and Microsoft Research, accepted as a NeurIPS 2026 Oral, targets a weakness in visual reasoning: finding the right region before zooming in to answer.
Instead of spelling out bounding-box coordinates as vocabulary tokens, a lightweight numerical head predicts continuous coordinates directly. Supervised training uses a regression objective reflecting spatial error; reinforcement learning samples coordinates from a Laplace policy and optimizes their probability densities alongside text-token probabilities with GRPO.
Result: localization IoU rises from 47.3 to 59.5 on the Vis-CoT-363K validation set — a 12.2-point gain that concerns supervised localization rather than overall QA accuracy.
More from Research
- Microsoft Research Redmond recruiting PhD interns for self-improving agents research — HildeKuehne · 2026-11-11
- Steering Qwen3-8B with random directions: 61% of runs produce body-sensation claims — Sauers_ · 2026-11-11
- Intelligence Explosion by 2027? Kokotajlo Cites METR Report, Bar Challenges How to Adjudicate — sebkrier · 2026-11-11
- Why batch size decides PPO vs GRPO: a directional bias-variance view — stochasticchasm · 2026-11-11
- Dense process rewards lose to standard GRPO at matched compute, 1,668-run study finds — CWRU · 2026-11-11
- HKU team proposes MC-JEPA, a market-conditioned JEPA for financial time series — OpenFinMind · 2026-11-11