Research map / Vision-language and multimodal / Multimodal LLMs

V-ABS: Action-Observer Driven Beam Search for Dynamic Visual Reasoning

ICML 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 V-ABS, a training-free thinker-actor-observer beam search with entropy-based adaptive weighting of prior and observer scores for multi-step visual reasoning, plus an 80k-sample SFT dataset for binary action verification.

Problem

Multi-step visual reasoning with MLLMs suffers from the imagination-action-observer (IAO) bias: prior action scores from the model are misaligned with feedback from the visual state after executing tools such as cropping.

Approach

Closed-loop beam search where each candidate action is executed via a visual tool, the updated visual state is scored by the same MLLM as an observer, and prior and observer scores are fused with an entropy-driven sigmoid weight; an SFT stage sharpens the prior toward correct actions.

Key results

• V-ABS on Qwen2.5-VL-7B reaches 91.1% on V* and 77.5% on HR-Bench-8K, versus 75.9% and 62.0% for the base model.
• Average improvement of 19.7% on Qwen3-VL-8B across benchmarks, as claimed in the abstract.
• On Sudoku, V-ABS raises Qwen3-VL accuracy from 14.4% to 37.8%; SFT lifts Jigsaw Res4 from 7.8% to 81.2% with V-ABS.

Surprising findings

• Removing the task-specific heuristic improves V* accuracy from 87.43% to 90.58%, suggesting the heuristic hurts.
• SFT gives only marginal gains on restricted-action tasks (+0.6% on V*) while being critical for open-ended Jigsaw.

Limitations

• Relies on the same MLLM as observer and thinker, so observer quality is bounded by the base model.
• Accuracy degrades beyond the optimal depth of 3 and inference cost grows with beam size and depth.
• Heuristic and entropy thresholds (mu, beta, delta) are hand-set and the paper reports inconsistent numbers across sections.

Experimental setup

Datasets: V* HR-Bench-4K HR-Bench-8K VisuoThink Frozen Lake (synthesized) TIR-Bench Maze Jigsaw (RefCOCO images) Sudoku SFT verification dataset (~80k, constructed)

Benchmarks: V* HR-Bench-4K HR-Bench-8K VisuoThink Frozen-Lake TIR-Bench Maze Jigsaw Sudoku

Metrics: accuracy success rate time cost API calls

Baselines: ZoomEye DeepEyes v2 Pixel-Reasoner Thyme Beam Search MCTS VisuoThink LLaVA-v1.6-13B LLaVA-HR-X-7B

Builds on: beam search tree of thoughts Monte Carlo tree search ZoomEye VisuoThink

Closest papers by meaning

• Multi-step Visual Reasoning with Visual Tokens Scaling and Verification (NeurIPS 2025 · similarity 0.91)
• Insight-V: Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.90)
• VisuRiddles: Fine-grained Perception is a Primary Bottleneck for Multimodal Large Language Models in Abstract Visual Reasoning (ICLR 2026 · similarity 0.90)
• Act2See: Emergent Active Visual Perception for Video Reasoning (CVPR 2026 · similarity 0.90)
• More Thought, Less Accuracy? On the Dual Nature of Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.90)
• BLINK-Twice: You see, but do you observe? A Reasoning Benchmark on Visual Perception (NeurIPS 2025 · similarity 0.89)