Research map / Vision-language and multimodal / Multimodal LLMs

AStar: Boosting Multimodal Reasoning with Automated Structured Thinking

AAAI 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 AStar, a training-free test-time method that distills MCTS-derived reasoning paths from 500 seed problems into "thought cards" and adaptively retrieves five per test query to guide MLLM reasoning.

Problem

Multimodal LLMs struggle with complex visual reasoning; existing remedies are either computationally expensive explicit search or data- and compute-heavy post-training (SFT/RL) that can be unstable.

Approach

Abstract high-level reasoning action sequences ("thought cards") from MCTS on a small seed set, indexed by problem complexity and text-image semantics, are retrieved per test problem and instantiated with the frozen model plus self-consistency and outcome reward verification, with no weight updates.

Key results

• AStar with Qwen2.5-7B reaches 53.9% on MathVerse, above GPT-4o's 50.2%, and 32.7% on MathVision versus GPT-4o's 30.4%.
• Removing thought cards drops the MathVision/MathVerse average from 43.3 to 33.8 (-9.5 points).
• Combining AStar with the PPO-trained LMM-R1 raises MathVerse accuracy from 41.8% to 48.3%.

Surprising findings

• A training-free method with 500 seed samples outperforms GPT-4o and models trained on 260K to 1.1M samples on MathVerse and MathVision.
• Thought cards derived from math problems also improve GPT-4o on MMMU and GAOKAO-MM.

Limitations

• Relies on MCTS for thought card construction, which the authors note is itself a search cost.
• Verification uses text-domain outcome reward models because visual-domain verifiers are scarce.
• Uses five fixed candidate instantiations per query, so inference cost scales with multiple sampled solutions.

Experimental setup

Datasets: MathVista MathVerse MathVision MMMU GAOKAO-MM ChartQA MMStar BLINK

Benchmarks: MathVerse MathVista MathVision MMMU GAOKAO-MM ChartQA MMStar BLINK

Metrics: accuracy

Baselines: GPT-4o Qwen2-VL-7B InternVL2-8B URSA-8B Math-LLaVA-13B R1-VL-7B LMM-R1 MM-Eureka-7B Mulberry-7B

Builds on: Monte Carlo Tree Search UCT self-consistency Mulberry

Closest papers by meaning

• ThinkMorph: Emergent Properties in Multimodal Interleaved Chain-of-Thought Reasoning (ICLR 2026 · similarity 0.94)
• MMFineReason: Closing the Multimodal Reasoning Gap via Open Data-Centric Methods (NeurIPS 2026 · similarity 0.93)
• SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows (NeurIPS 2026 · similarity 0.92)
• VGR: Visual Grounded Reasoning (ICLR 2026 · similarity 0.92)
• UniVLR: Unifying Text and Vision in Visual Latent Reasoning for Multimodal LLMs (NeurIPS 2026 · similarity 0.92)
• LENS: Multi-level Evaluation of Multimodal Reasoning with Large Language Models (ICLR 2026 · similarity 0.92)