Research map / Vision-language and multimodal / Multimodal LLMs

Mini-o3: Scaling Up Reasoning Patterns and Interaction Turns for Visual Search

ICLR 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.

Mini-o3 builds a recipe for multi-turn image-tool visual search (cold-start SFT on in-context-synthesized trajectories, then GRPO with over-turn masking), enabling test-time scaling to tens of interaction turns on a 7B VLM.

Problem

Open-source thinking-with-image VLMs show monotonous reasoning and very few tool-interaction turns, failing on hard visual search tasks requiring trial-and-error exploration.

Approach

Mask (rather than penalize with zero reward) over-turn and over-context rollouts in GRPO so the model is not pushed to answer early, allowing training with a small 6-turn budget while accuracy scales with up to 32 test-time turns; combined with a harder training dataset (VisualProbe) and cold-start trajectories synthesized by in-context prompting of a VLM.

Key results

• Mini-o3 reaches 48.0% on VisualProbe-Hard, versus 35.1% for DeepEyes, and 88.2% on V* Bench, which is above Chain-of-Focus's 88.0% and Pixel Reasoner's 86.3%.
• Mini-o3 is trained with a 6-turn cap but accuracy keeps rising as test-time max turns increase from 4 to 32.
• Removing over-turn masking drops VisualProbe-Hard from 44.4% to 32.2% in the ablation (1M max pixels).

Surprising findings

• Naively assigning zero reward to over-turn responses causes the model to answer prematurely and suppresses interaction turns.
• The VLM used for cold-start data synthesis need not have native thinking-with-image ability; in-context mimicking suffices.

Limitations

• Performance collapses without cold-start SFT, so RL alone is insufficient for multi-turn tool use.
• Training with a small turn budget is a trade-off: larger max pixels cause early stopping while smaller ones increase perceptual hallucinations.

Experimental setup

Datasets: VisualProbe training split DeepEyes-Datasets-47k VisualProbe test

Benchmarks: VisualProbe V* Bench HR-Bench MME-Realworld

Metrics: Avg@32 accuracy Avg@8 accuracy Avg@1 accuracy average interaction turns

Baselines: GPT-4o LLaVA-OneVision Qwen2.5-VL-Instruct SEAL DyFo Chain-of-Focus Pixel Reasoner DeepEyes

Builds on: GRPO DAPO DeepEyes Qwen2.5-VL

Closest papers by meaning

• InSight-o3: Empowering Multimodal Foundation Models with Generalized Visual Search (ICLR 2026 · similarity 0.91)
• Zebra-CoT: A Dataset for Interleaved Vision-Language Reasoning (ICLR 2026 · similarity 0.90)
• SenseSearch: Empowering Vision-Language Models with High-Resolution Agentic Search-Reasoning via Reinforcement Learning (CVPR 2026 · similarity 0.88)
• MiCo: Multi-image Contrast for Reinforcement Visual Reasoning (NeurIPS 2025 · similarity 0.88)
• DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learning (ICLR 2026 · similarity 0.88)
• Reason-RFT: Reinforcement Fine-Tuning for Visual Reasoning of Vision Language Models (NeurIPS 2025 · similarity 0.88)