Companion-style QA Assistance in Ego-Vision

TL;DR

MyBuddy significantly improves performance on the BuddyVQA benchmark through multimodal chain-of-thought reasoning.

cs.CV 🔴 Advanced 2026-09-07 2 views
Hangyu Qin Junbin Xiao Shenglang Zhang Angela Yao
multimodal video QA egocentric chain-of-thought AI assistant

Key Findings

Methodology

MyBuddy employs a multimodal chain-of-thought reasoning mechanism, integrating historical QA and visual content to infer final answers. It features a question filter and multi-level memory to enhance efficiency in streaming QA settings.

Key Results

  • MyBuddy achieved over 50% accuracy improvement on the BuddyVQA benchmark, significantly outperforming existing foundation models.
  • It also performed well on other streaming and common video QA benchmarks, demonstrating the method's generalizability.
  • Ablation studies show that chain-of-thought reasoning and multi-level memory are crucial for performance improvement.

Significance

This study provides a new benchmark and method for egocentric video QA, addressing shortcomings in existing models when handling ego-deictic expressions and interactively chained questions. It holds significant implications for academia and industry in developing real-time intelligent assistants.

Technical Contribution

Introduced a training-free, model-agnostic framework MyBuddy capable of handling ego-deictic and chained questions, offering new engineering possibilities.

Novelty

MyBuddy is the first to implement multimodal chain-of-thought reasoning in an egocentric streaming environment, significantly enhancing the practicality of QA assistants.

Limitations

  • Memory mechanisms may cause performance bottlenecks when processing extremely long videos.
  • There is still a possibility of misjudgment in complex visual scenes.

Future Work

Future research could explore more efficient memory management strategies and more complex visual reasoning capabilities to further enhance the system's real-time performance and accuracy.

AI Executive Summary

With the rapid adoption of wearable cameras and mixed-reality devices, egocentric visual assistants have become a significant research area for enhancing daily human experiences. Existing video QA models mainly focus on offline, third-person video clips, neglecting the interactively chained questions and ego-deictic expressions occurring in shared first-person contexts. To address this, researchers have introduced the BuddyVQA benchmark, focusing on egocentric QA assistants in streaming environments.

The MyBuddy framework employs a multimodal chain-of-thought reasoning mechanism, integrating historical QA and visual content to infer final answers. Its multi-level memory mechanism and question filter efficiently retrieve QA and visual information in streaming settings. Experimental results show that MyBuddy significantly enhances foundation model performance on the BuddyVQA benchmark and performs well on other streaming and common video QA benchmarks.

While MyBuddy has made breakthroughs in handling ego-deictic and chained questions, memory mechanisms may cause performance bottlenecks when processing extremely long videos. Future research could explore more efficient memory management strategies and more complex visual reasoning capabilities to further enhance the system's real-time performance and accuracy.

Deep Analysis

Background

With the proliferation of wearable cameras and mixed-reality devices, egocentric visual assistants have become a research hotspot. Existing video QA models primarily handle offline, third-person video clips, neglecting the interactively chained questions and ego-deictic expressions in egocentric contexts.

Core Problem

Traditional video QA models struggle with ego-deictic expressions and interactively chained questions in egocentric contexts. These issues require models to infer user intent in streaming environments.

Innovation

The MyBuddy framework addresses ego-deictic and chained questions through a multimodal chain-of-thought reasoning mechanism. Its multi-level memory mechanism and question filter enhance efficiency in streaming environments.

Methodology

  • �� Multimodal chain-of-thought reasoning: integrates historical QA and visual content to infer answers.
  • �� Question filter: enhances QA retrieval efficiency.
  • �� Multi-level memory mechanism: efficiently stores and retrieves information in streaming environments.

Experiments

Experiments were conducted on the BuddyVQA benchmark using various streaming and common video QA datasets to validate MyBuddy's performance. Ablation studies show that chain-of-thought reasoning and multi-level memory are crucial for performance improvement.

Results

MyBuddy achieved over 50% accuracy improvement on the BuddyVQA benchmark and performed well on other benchmarks, demonstrating the method's generalizability.

Applications

MyBuddy can be used to develop real-time intelligent assistants, particularly in scenarios requiring egocentric video QA.

Limitations & Outlook

Memory mechanisms may cause performance bottlenecks when processing extremely long videos. Future research could explore more efficient memory management strategies.

Plain Language Accessible to non-experts

Imagine you're cooking in a kitchen, and an assistant is there to help you. You ask, "Is this ready to eat?" The assistant uses previous conversations and current visual information to give you an answer. MyBuddy is like that assistant, understanding your questions in egocentric videos and providing reasonable answers.

ELI14 Explained like you're 14

Imagine you're playing a game, and an assistant is there to help you. You ask, "Where should I go next?" The assistant uses previous conversations and the current game screen to tell you what to do. MyBuddy is like that assistant, understanding your questions in egocentric videos and providing reasonable answers.

Glossary

Multimodal Chain-of-Thought Reasoning

A process that combines multiple information sources for reasoning, especially in video QA.

MyBuddy uses multimodal chain-of-thought reasoning to handle ego-deictic and chained questions.

Ego-Deictic Expressions

Indicative language used in first-person perspectives, such as 'this' and 'there'.

The BuddyVQA benchmark emphasizes the importance of ego-deictic expressions.

Streaming Environment

An environment that processes video streams in real-time, requiring efficient information retrieval and reasoning.

MyBuddy achieves efficient QA processing in streaming environments.

Multi-Level Memory Mechanism

A hierarchical memory system for efficient storage and retrieval of information.

MyBuddy uses a multi-level memory mechanism to enhance efficiency in streaming environments.

Question Filter

A mechanism to enhance QA retrieval efficiency.

MyBuddy features a question filter to improve efficiency in streaming QA.

Open Questions Unanswered questions from this research

  • 1 Efficient memory management in extremely long videos remains a challenge.
  • 2 Misjudgment issues in complex visual scenes require further research.

Applications

Immediate Applications

Real-Time Intelligent Assistants

MyBuddy can be used to develop real-time intelligent assistants, particularly in scenarios requiring egocentric video QA.

Long-term Vision

Augmented Reality Applications

In the future, MyBuddy could be applied to augmented reality devices, providing smarter user interaction experiences.

Abstract

AI companions are envisioned as always-on assistants that support users in daily life. With this regard, we introduce BuddyVQA, a benchmark for companion-style question answering (QA) on egocentric streaming video. BuddyVQA contains 21.6K questions linked to 6K highlight moments across 1,012 long, egocentric videos. It features two key characteristics that are common in daily first-person QA assistance but are largely overlooked in existing VideoQA benchmarks: ego-deictic expressions and interactively chained questions (e.g., "Where is it?", "How to get there?"). These require models to infer a user's in-situation intent by resolving visual pronouns in the context of egocentric visual and QA contents, with both grounded in a long-form streaming setting. To tackle the challenges, we propose MyBuddy, a companion-style QA assistant that highlights a multimodal chain-of-thought reasoning mechanism to infer the final answer based on the historical QA and visual content. An additional question filter and multi-level memory are designed to facilitate efficient QA and visual information retrieval under streaming QA settings. Experiments show that MyBuddy significantly enhances the performance of foundation models on BuddyVQA. Moreover, these gains generalize to other streaming and common video QA benchmarks, demonstrating the applicability and effectiveness of our approach. Our code and dataset are available at https://github.com/QHUni/BuddyVQA

cs.CV cs.AI