Research library

Evaluation Research note

A strong baseline can be surprisingly simple

Sample several reasoning paths before adding a review protocol.

The finding

Self-consistency samples multiple reasoning paths and selects the answer that is most consistent across them. The authors report improvements on arithmetic and commonsense reasoning benchmarks over single-path decoding. [1]

What it means for Pingpong

Pingpong aims to revise an answer through critique, which is different from selecting the most common answer. To demonstrate that the review framework adds value, it should be evaluated against self-consistency as well as a single response. Otherwise extra samples could explain a gain attributed to the chain.

The limit

Voting is easier to define for questions with a discrete answer than for an open-ended strategy or proposal. That limits direct transfer, but it does not remove the need for a competitive baseline.

Source

Self-Consistency Improves Chain of Thought Reasoning in Language Models

Xuezhi Wang et al. (2022)

ICLR 2023
DOI: 10.48550/arXiv.2203.11171

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