Research map / Vision-language and multimodal / Multimodal LLMs
Grounding Multimodal Reasoning with Evidence-Ablated Negatives
chenxi liu, Tianyi Xiong, Yanshuo Chen, Heng Huang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes evidence-ablated negatives to ground RLVR-based multimodal reasoning in visual evidence, addressing credit assignment ambiguity in outcome supervision.
Problem
Outcome-based RLVR for multimodal reasoning can reward responses that exploit language priors or dataset shortcuts instead of grounding in visual evidence, creating credit assignment ambiguity.
Approach
Constructs high-quality negative responses by ablating the requisite visual evidence, so that reasoning that ignores the evidence is penalized during RL training. Details beyond the abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add evidence-ablated negatives to DAPO | DAPO · 9 multimodal reasoning benchmarks average | average accuracy: 50.3 -> 51.6 (3B); 56.5 -> 57.9 (7B); 71.9 -> 73.0 (Qwen3-VL-8B) | Improved |
| use RankMix selection instead of global-only scaled-IIG selection (ablation) | EVAN, Qwen2.5-VL-3B · 9 benchmarks average | average accuracy: 50.3 -> 51.6 | Improved |
| use RankMix instead of random selection (ablation) | EVAN, Qwen2.5-VL-3B · 9 benchmarks average | average accuracy: 51.1 -> 51.6 | Improved |
| build negatives from evidence-ablated image instead of clean image (ablation) | EVAN, Qwen2.5-VL-3B · 9 benchmarks average | average accuracy: 50.7 -> 51.6 | Improved |
| use scaled-IIG-guided patch masking instead of FGSM pixel perturbation (ablation) | EVAN, Qwen2.5-VL-3B · 9 benchmarks average | average accuracy: 50.8 -> 51.6 | Improved |
| use scaled-IIG-guided masking instead of random patch masking at 50% (ablation) | EVAN, Qwen2.5-VL-3B · 9 benchmarks average | average accuracy: 50.8 -> 51.6 | Improved |
| mask 50% of patches instead of 25%, 75% or 100% (ablation) | EVAN, Qwen2.5-VL-3B · 9 benchmarks average | average accuracy: 51.1 / 51.4 / 50.9 -> 51.6 | Improved |
| select 20% of data instead of 30% (ablation) | EVAN, Qwen2.5-VL-3B · 9 benchmarks average | average accuracy: 50.8 -> 51.6 | Improved |
| select 20% of data instead of 10% (ablation) | EVAN, Qwen2.5-VL-3B · 9 benchmarks average | average accuracy: 51.5 -> 51.6 | No clear effect |
| add bounded negative advantage to evidence-ablated negatives (ablation) | EVAN, Qwen2.5-VL-3B · training dynamics | entropy and gradient norm stability | Improved |
| add clean-image likelihood and length filtering of negatives (ablation) | EVAN, Qwen2.5-VL-7B · training dynamics | PPO KL and response length | Improved |
| add evidence-ablated negatives to DAPO | Qwen2.5-VL-3B · ChartQA, GQA, Visual CounterFact | accuracy: Visual CounterFact pooled 81.61 -> 85.01; ChartQA 81.74 -> 82.55; GQA 54.46 -> 55.83 | Improved |
Experimental setup
Builds on: RLVR
Closest papers by meaning
- Reinforcing Multimodal Reasoning Against Visual Degradation (NeurIPS 2026 · similarity 0.91)
- Perceptual-Evidence Anchored Reinforced Learning for Multimodal Reasoning (CVPR 2026 · similarity 0.90)
- CARE What Fails: Contrastive Anchored-REflection for Verifiable Multimodal Reasoning (CVPR 2026 · similarity 0.90)
- See It, Say It, Sorted: An Iterative Training-Free Framework for Visually-Grounded Multimodal Reasoning in LVLMs (CVPR 2026 · similarity 0.89)
- Perception-Aware Policy Optimization for Multimodal Reasoning (ICLR 2026 · similarity 0.89)
- Boosting Reasoning in Large Multimodal Models via Activation Replay (CVPR 2026 · similarity 0.89)
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.
