Look Before You Reason: Implicit Visual Thinking for Efficient Multimodal Reasoning
To address these issues, we propose Look Before You Reason (LoRe), an implicit visual thinking framework for efficient visual reasoning within a single-turn VLM interaction.
ProofPaper ↗
Key points
- Vision-language models (VLMs) have advanced rapidly in multimodal reasoning, yet they remain limited on tasks requiring fine-grained understanding of small or easily overlooked image regions.
- While effective, their multi-turn interactions with VLMs introduce redundant token computation and slow inference, and separate region inspection from the reasoning step that depends on it.
- To support fine-grained evaluation of multimodal reasoning, we introduce COCO-ObjQA, a large-scale benchmark for object-centric multimodal reasoning.
- Experiments on three challenging multimodal benchmarks demonstrate that LoRe consistently improves VLM reasoning performance with low additional inference overhead.
Sources (1)
- [1]Look Before You Reason: Implicit Visual Thinking for Efficient Multimodal ReasoningarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 10:11 PM
To address these issues, we propose Look Before You Reason (LoRe), an implicit visual thinking framework for efficient visual reasoning within a single-turn VLM interaction.
Vision-language models (VLMs) have advanced rapidly in multimodal reasoning, yet they remain limited on tasks requiring fine-grained understanding of small or easily overlooked image regions.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 19, 2026Mitigating Reward Hacking in Preference Reward Models with Evolving Penalties
- Oct 18, 2026Finite Interaction Lens: Structured Precision Sharing for Vision-Language Models
- Oct 18, 2026Explore or Confirm? LLM Agents Misallocate Measurements Under Noise
- Oct 18, 2026Conformal Prediction-Enabled Sequence-level Policy Optimization for Reasoning Language Models
- Oct 17, 2026Learning to Program Retrievals: Composing Search Primitives via Reinforcement Learning
- Oct 17, 2026From Views to Worlds: Active Exploration over 3D Worlds for Vision-Language Models