Research library

Multi-model review Research note

An answer can benefit from another model's objections

Early evidence for review between language-model instances.

The finding

Du and colleagues ask model instances to propose answers and discuss them over multiple rounds. They report improvements on selected mathematical, strategic-reasoning, and factual tasks. The work demonstrates that interaction between generated answers can be useful, rather than merely adding more text. [1]

What it means for Pingpong

Pingpong uses a sequential review chain, not a round-table debate. Each later model can inspect previous responses and produce a revised answer to the original question. The shared idea is that another assessment can expose a weakness the initial response left intact.

The limit

The study's models, tasks, and debate protocol differ from Pingpong. Its gains do not establish that five different providers, a particular order, or our framing outperform other methods.

Source

Improving Factuality and Reasoning in Language Models through Multiagent Debate

Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mordatch (2023)

ICML 2024
DOI: 10.48550/arXiv.2305.14325

Paper PDF

This is Pingpong's interpretation of external research, not a result from a Pingpong experiment or an endorsement by the authors. Editorial standard