Research map / Vision-language and multimodal / Multimodal LLMs

See Further, Think Deeper: Advancing VLM's Reasoning Ability with Low-level Visual Cues and Reflection

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual 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 ForeSight, an RL-trained Qwen2.5-VL-7B framework that interleaves low-level visual tools (Canny, zoom-in, color) and mask-based visual reflection for odd-one-out grounding, evaluated on a new CG-SalBench.

Problem

VLM reasoning is largely text-only, lacking low-level visual cues and effective visual feedback, which leads to missed fine-grained visual details and open-loop answers on saliency/odd-one-out grounding tasks.

Approach

Combine RL-learned invocation of low-level visual tools (Canny edges, zoom-in, color enhancement) with a mask-based visual reflection loop, where the draft answer's bounding box is masked out of the image and the model re-examines its answer, all trained with outcome-based composite rewards via GRPO after an SFT cold start.

Key results

• ForeSight-7B reaches 62.24 IoU and 84.24 F1 on CG-SalBench, versus 32.56 IoU and 64.97 F1 for Qwen2.5-VL-7B (+29.68 IoU, +19.27 F1).
• ForeSight-7B's IoU is close to Qwen2.5-VL-72B (66.00) and above the best closed-source model Doubao-Seed-1.6 (47.66).
• Ablation: SFT+RL gives 62.24 IoU vs 57.47 for SFT-only and 52.80 for RL-only; on RefCOCO, ForeSight reaches 91.42 vs 88.91 for reproduced Qwen2.5-VL-7B.

Surprising findings

• Inference-time tools and reflection added to closed-source models without training did not consistently help; Doubao-Seed-1.6 IoU dropped with tools (47.66 to 34.73).
• ForeSight slightly underperforms its base on MMBench (81.50 vs 82.47 reproduced Qwen2.5-VL-7B).

Limitations

• Built on the small Qwen2.5-VL-7B, which the authors say limits results on failure cases.
• Reflection can miss incomplete bounding-box encapsulation; Canny tool sometimes redundant and causes errors.
• Only evaluated on the CG-SalBench odd-one-out task plus a few grounding/VQA benchmarks; gains on general VQA are modest.

Experimental setup

Datasets: SalBench (P3) SalBench (O3) CG-SalBench training split RefCOCO RefCOCO+ RefCOCOg MME MMBench

Benchmarks: CG-SalBench RefCOCO RefCOCO+ RefCOCOg MME MMBench

Metrics: IoU F1 Precision Recall

Baselines: InternVL3-8B Qwen2.5-VL-7B Qwen2.5-VL-72B Qwen3-VL-8B GPT-5 Gemini-2.5-Flash Doubao-Seed-1.6 Claude-Sonnet-4 DeepEyes

Builds on: GRPO DeepEyes SalBench Qwen2.5-VL SWIFT

Closest papers by meaning

• More Thought, Less Accuracy? On the Dual Nature of Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.91)
• From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models (ICML 2026 · similarity 0.91)
• CoFFT: Chain of Foresight-Focus Thought for Visual Language Models (NeurIPS 2025 · similarity 0.91)
• VTool-R1: VLMs Learn to Think with Images via Reinforcement Learning on Multimodal Tool Use (ICLR 2026 · similarity 0.91)
• Empowering Small VLMs to Think with Dynamic Memorization and Exploration (ICLR 2026 · similarity 0.91)
• VGR: Visual Grounded Reasoning (ICLR 2026 · similarity 0.91)