Investigate PG-TD (Planning-Guided Transformer Decoding) sampling

[walking-octopus]

There's been some work going on for beam search #2267, CFG #2135, steering vectors #1472, Honest LLaMA #1799, and other techniques to improve generation quality. So I thought sharing this paper may be of use.

It outlines some sampling tricks to improve code generation, which may be appropriate for StarCoder.cpp, given LLaMA's underwhelming coding performance, but I think these ideas can perhaps be used as inspiration for other types of planners for different tasks.

"Planning with Large Language Models for Code Generation"

Abstract: Existing large language model-based code generation pipelines typically use beam search or sampling algorithms during the decoding process. Although the programs they generate achieve high token-matching-based scores, they often fail to compile or generate incorrect outputs. The main reason is that conventional Transformer decoding algorithms may not be the best choice for code generation. In this work, we propose a novel Transformer decoding algorithm, Planning-Guided Transformer Decoding (PG-TD), that uses a planning algorithm to do lookahead search and guide the Transformer to generate better programs. Specifically, instead of simply optimizing the likelihood of the generated sequences, the Transformer makes use of a planner that generates candidate programs and tests them on public test cases. The Transformer can therefore make more informed decisions and generate tokens that will eventually lead to higher-quality programs. We also design a mechanism that shares information between the Transformer and the planner to make our algorithm computationally efficient. We empirically evaluate our framework with several large language models as backbones on public coding challenge benchmarks, showing that 1) it can generate programs that consistently achieve higher performance compared with competing baseline methods; 2) it enables controllable code generation, such as concise codes and highly-commented codes by optimizing modified objective.

[github-actions]

This issue was closed because it has been inactive for 14 days since being marked as stale.