Research map / Vision-language and multimodal / Multimodal LLMs

See Less, See Right: Bi-directional Perceptual Shaping For Multimodal Reasoning

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 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 BiPS, a training-time two-stage GRPO method that applies bidirectional KL constraints (consistency with an evidence-preserving view, separation from an evidence-ablated view) built via programmatic chart-code editing, improving Qwen2.5-VL-7B on chart and general VQA benchmarks.

Problem

VLMs lack fine-grained visual perception and often rely on text-only shortcuts; existing intermediate visual cue methods are shape-rigid, domain-specific, and add inference overhead.

Approach

Use question-conditioned evidence-preserving and evidence-ablated chart views (generated by editing executable plotting code) as training-time signals: a KL-consistency term pulls predictions on the full image toward the preserved view, and a KL-separation term pushes them away from the ablated view, applied in a coarse-to-fine two-stage curriculum on top of GRPO, with no inference-time cost.

Key results

• BiPS-Chart-7B (trained on 13K chart samples) raises Qwen2.5-VL-7B average across eight benchmarks from 44.3 to 51.6 (+7.3), with CharXiv 42.5 to 49.4 and Evochart 52.0 to 68.2.
• BiPS-General-7B, adding 39K math samples with standard GRPO, reaches 52.5 average (+8.2) and 50.6 on CharXiv.
• BiPS-General outperforms a standard GRPO baseline trained on the same data by 5.2 points on CharXiv.

Surprising findings

• Random masking of 60% of image patches as the alternative view underperforms programmatic views, and on Qwen3-VL-8B-Thinking the gains on OOD benchmarks are only moderate.

Limitations

• Evaluated mainly on chart training data; transfer to Qwen3-VL-8B-Thinking and non-chart domains is more limited.
• Training relies on an LLM arbitrator (GPT-5-mini) for question reformulation and code editing, and only about 7K of 13K samples have preserve views.

Experimental setup

Datasets: ECD ViRL39k

Benchmarks: CharXiv ChartQAPro ChartMuseum Evochart MathVista MathVision MathVerse-VO MMStar ECD-Bench

Metrics: accuracy

Baselines: GRPO Chart-R1-7B DeepEyes-7B Vision-R1-7B R1-OneVision-7B BigCharts-R1-7B InternVL3-8B GPT-4o Claude-3.7-Sonnet

Builds on: GRPO ECD ViRL39k Qwen2.5-VL

Closest papers by meaning

• Perception Tokens Enhance Visual Reasoning in Multimodal Language Models (CVPR 2025 · similarity 0.89)
• SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows (NeurIPS 2026 · similarity 0.89)
• VisionPulse: Dynamic Visual Sparsity for Efficient Multimodal Reasoning (ICML 2026 · similarity 0.89)
• Evidence-RL: Towards Evidence-intensive Visual Reasoning (NeurIPS 2026 · similarity 0.89)
• Seeing but Not Believing: Probing the Disconnect Between Visual Attention and Answer Correctness in VLMs (ICLR 2026 · similarity 0.89)
• Look-Back: Implicit Visual Re-focusing in MLLM Reasoning (AAAI 2026 · similarity 0.89)