Research map / Vision-language and multimodal / Multimodal LLMs
Beyond Where to Look: Trajectory-Guided Reinforcement Learning for Multimodal RLVR
Jinda Lu, Junkang Wu, Jinghan Li, Kexin Huang, Shuo Yang, Mingzhu Chen, Jiancan Wu, Kuien Liu, Xiang Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add trajectory filtering (keep only correct expert trajectories) (ablation) | TGRL-GRPO · Avg of 5 benchmarks (Geo3K) | average accuracy: 57.50 -> 58.46 | Improved |
| add token reweighting to expert trajectories (ablation) | TGRL-GRPO · Avg of 5 benchmarks (Geo3K) | average accuracy: 57.03 -> 58.46 | Improved |
| add expert trajectories (with reweighting and filtering) (ablation) | GRPO · Avg of 5 benchmarks (Geo3K) | average accuracy: 57.16 -> 58.46 | Improved |
| add expert trajectories without token reweighting (ablation) | GRPO · Avg of 5 benchmarks (Geo3K) | average accuracy: 57.16 -> 57.03 | Hurt |
| add trajectory filtering (ablation) | TGRL-DAPO · Avg of 5 benchmarks (Geo3K) | average accuracy: 57.93 -> 59.51 | Improved |
| add token reweighting (ablation) | TGRL-DAPO · Avg of 5 benchmarks (Geo3K) | average accuracy: 57.50 -> 59.51 | Improved |
| add expert trajectories (with reweighting and filtering) (ablation) | DAPO · Avg of 5 benchmarks (Geo3K) | average accuracy: 56.92 -> 59.51 | Improved |
| set token reweighting beta to 5 instead of 7 (ablation) | TGRL-GRPO · Avg of 5 benchmarks (Geo3K) | average accuracy: 57.36 -> 58.46 | Improved |
| set token reweighting beta to 14 instead of 15 (ablation) | TGRL-DAPO · Avg of 5 benchmarks (Geo3K) | average accuracy: 57.88 -> 59.51 | Improved |
| add TGRL trajectories to GRPO with 39K ViRL39K training data | GRPO · Avg of 5 benchmarks (ViRL39K) | average accuracy: 57.71 -> 59.38 | Improved |
| add TGRL trajectories to DAPO with 39K ViRL39K training data | DAPO · Avg of 5 benchmarks (ViRL39K) | average accuracy: 58.15 -> 60.68 | Improved |
| scale TGRL-GRPO training data from 2.1K Geo3K to 39K ViRL39K | TGRL-GRPO · Avg of 5 benchmarks | average accuracy: 58.46 -> 59.38 | Improved |
Closest papers by meaning
- Reinforcing Multimodal Reasoning Against Visual Degradation (NeurIPS 2026 · similarity 0.91)
- More Thought, Less Accuracy? On the Dual Nature of Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.90)
- Perception-Aware Policy Optimization for Multimodal Reasoning (ICLR 2026 · similarity 0.90)
- See It, Say It, Sorted: An Iterative Training-Free Framework for Visually-Grounded Multimodal Reasoning in LVLMs (CVPR 2026 · similarity 0.90)
- Perception-R1: Advancing Multimodal Reasoning Capabilities of MLLMs via Visual Perception Reward (ICLR 2026 · similarity 0.90)
- Spotlight on Token Perception for Multimodal Reinforcement Learning (ICLR 2026 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
