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

2026-05-11

AI Summary

This work addresses the inconsistency between imagined actions and observed feedback in multimodal large language models, which often leads to Imagination-Action-Observation (IAO) bias, undermining reasoning stability and optimality. To mitigate this issue, the authors propose the V-ABS framework, which introduces an action-observer-driven beam search mechanism coupled with an entropy-based adaptive weighting algorithm to dynamically balance policy priors and observational signals. Additionally, they construct a supervised fine-tuning dataset comprising 80,000 samples to guide models toward correct action trajectories. The proposed method achieves state-of-the-art performance across eight benchmarks, yielding an average improvement of 19.7% over the Qwen3-VL-8B baseline and demonstrating consistent effectiveness on both open-source and closed-source models.

Abstract

Multimodal large language models (MLLMs) have achieved remarkable success in general perception, yet complex multi-step visual reasoning remains a persistent challenge. Although recent agentic approaches incorporate tool use, they often neglect critical execution feedback. Consequently, they suffer from the imagination-action-observer (IAO) bias, a misalignment between prior imagination and observer feedback that undermines reasoning stability and optimality. To bridge this gap, we introduce V-ABS, an action-observer driven beam search framework that enables deliberate reasoning through thinker-actor-observer iterations. We also propose an entropy-based adaptive weighting algorithm to mitigate the IAO bias by dynamically balancing the confidence scores between the policy priors and the observational feedback. Moreover, we construct a large-scale supervised fine-tuning (SFT) dataset comprising over 80k samples to guide the model to assign higher prior confidence to correct action paths. Extensive experiments across eight diverse benchmarks show that V-ABS achieves state-of-the-art performance, delivering an average improvement of 19.7% on the Qwen3-VL-8B baseline and consistent gains across both open-source and proprietary models.

Problem

• visual reasoning
• multimodal large language models
• execution feedback
• IAO bias
• dynamic reasoning

Innovation

• action-observer feedback
• beam search
• multimodal reasoning
• adaptive weighting
• IAO bias