Reflective Dialogue between Teacher and Solver Agents for Video Question Answering

2026-05-26

AI Summary

This work addresses the challenge of domain adaptation in few-shot video question answering without requiring model fine-tuning. It proposes a multi-agent reflective dialogue mechanism operating at inference time, wherein a teacher agent and a solver agent engage in iterative interactions: the teacher poses questions based on the support set and provides feedback on answer correctness, while the solver generates answers along with visually grounded explanations. This process constructs an enriched contextual prompt that enhances final reasoning. The framework operates entirely through in-context learning during inference, eliminating the need for parameter updates. Evaluated on the EgoCross benchmark, the method substantially outperforms zero-shot and standard in-context learning baselines and secured third place in the open-track category of the inaugural Cross-Domain EgoCross Challenge at the CVPR 2026 EgoVis Workshop.

Abstract

Various approaches have been proposed to adapt Vision-Language Models (VLMs) to specialized domains for Video Question Answering, including fine-tuning and in-context learning. However, acquiring task-specific knowledge at the inference phase from only a small labeled support set without fine-tuning remains a challenge. In this paper, we propose a method that achieves adaptation solely through inference-time context injection. Our method first constructs a Reflective Dialogue (RD) -- a multi-turn conversation between two agents, in which Teacher poses each support question and delivers correctness feedback, and Solver answers and provides visual grounding explanations (or reflections) for both correct and incorrect answers. This dialogue history is then used as context at the inference phase. Experiments on the EgoCross benchmark demonstrate that our method outperforms both a baseline zero-shot setting and a standard in-context learning approach that passes support set examples directly, achieving 3rd place in the Open-source Track of the 1st Cross-Domain EgoCross Challenge at the CVPR 2026 EgoVis Workshop, for which this paper also serves as a technical report.

Problem

• Video Question Answering
• Vision-Language Models
• In-context Learning
• Few-shot Adaptation
• Inference-time Adaptation

Innovation

• Reflective Dialogue
• Inference-time Adaptation
• Vision-Language Models
• Video Question Answering
• In-context Learning