InterMT: Multi-Turn Interleaved Preference Alignment with Human Feedback

TL;DR

This paper introduces InterMT, a dataset and benchmark for multi-turn, multimodal interaction aligned with human preferences, aiming to enhance multimodal large models' interactive capabilities.

Contribution

It presents the first preference dataset for multi-turn multimodal interaction, incorporating human feedback and expert annotations, and introduces tools and benchmarks to evaluate and improve MLLMs' interactive abilities.

Findings

InterMT dataset contains 15.6k prompts and 52.6k dialogue instances.

InterMT-Bench evaluates MLLMs' multi-turn multimodal assistance.

Open-source data facilitates future research in multimodal alignment.

Abstract

As multimodal large models (MLLMs) continue to advance across challenging tasks, a key question emerges: What essential capabilities are still missing? A critical aspect of human learning is continuous interaction with the environment -- not limited to language, but also involving multimodal understanding and generation. To move closer to human-level intelligence, models must similarly support multi-turn, multimodal interaction. In particular, they should comprehend interleaved multimodal contexts and respond coherently in ongoing exchanges. In this work, we present an initial exploration through the InterMT -- the first preference dataset for multi-turn multimodal interaction, grounded in real human feedback. In this exploration, we particularly emphasize the importance of human oversight, introducing expert annotations to guide the process, motivated by the fact that current MLLMs lack such complex interactive capabilities. InterMT captures human preferences at both global and local levels into nine sub-dimensions, consists of 15.6k prompts, 52.6k multi-turn dialogue instances, and 32.4k human-labeled preference pairs. To compensate for the lack of capability for multi-modal understanding and generation, we introduce an agentic workflow that leverages tool-augmented MLLMs to construct multi-turn QA instances. To further this goal, we introduce InterMT-Bench to assess the ability of MLLMs in assisting judges with multi-turn, multimodal tasks. We demonstrate the utility of \InterMT through applications such as judge moderation and further reveal the multi-turn scaling law of judge model. We hope the open-source of our data can help facilitate further research on aligning current MLLMs to the next step. Our project website can be found at https://pku-intermt.github.io .