Chain of Code
Reasoning with a Language Model-Augmented Code Emulator
Chengshu Li*,1,2, Jacky Liang1, Andy Zeng1, Xinyun Chen1, Karol Hausman1,2,
Dorsa Sadigh1,2, Sergey Levine1,3, Li Fei-Fei2, Fei Xia†,1, Brian Ichter†,1
1Google DeepMind, 2Stanford University, 3University of California, Berkeley
*Work done as a student researcher at Google DeepMind. †Equal advising.
[Paper][Code (coming soon)]
Video
Introduction
We propose Chain of Code (CoC), a simple yet surprisingly effective extension that improves Language Model (LM) code-driven reasoning. It broadens the scope of reasoning questions that LMs can correctly answer by "thinking in code".
The key idea is to encourage LMs to format semantic sub-tasks in a program as flexible pseudocode that the interpreter can explicitly catch undefined behaviors and hand off to simulate with an LM (as an "LMulator").
Method
Compared to previous reasoning methods, Chain of Code first (d) generates code or pseudocode to solve the question and then (e) executes the code with a code interpreter if possible, and with an LMulator (language model emulating code) otherwise.
Blue highlight indicates LM generation.
Red highlight indicates LM generated code being executed by an interpreter.
Purple highlight indicates an LMulator simulating the code via a program state in green.
Results on Language Reasoning
On BIG-Bench Hard (BBH), Chain of Code achieves 84%, a gain of 12% over Chain of Thought and a new state of the art. It further, outperforms the average human raters in 18 out of 23 tasks.
Chain of Code performs on par with Chain of Thought for the NLP subset of BBH, and outperforms even the best human raters for the algorithmic subset of BBH.
Chain of Code (Interweave) based on text-davinci-003 even outperforms a much larger instruction tuned model gpt-4, which is instructed to write code to solve the reasoning problems, if it's helpful to do so.
Robotics Applications
Chain of Code is well fit for solving robotics tasks because they require both semantic and algorithmic reasoning.
They also involve interfacing with other APIs through code (e.g., control or perception APIs) and with users through natural language.
Red highlight indicates LM generated code being executed by an interpreter.
Purple highlight indicates an LMulator simulating the code.
Example Model Outputs for Language Reasoning
Example model outputs for some of the most challenging BBH tasks that require both semantic and algorithm reasoning.
Red highlight indicates LM generated code being executed by an interpreter.
Purple highlight indicates an LMulator simulating the code via a program state in green.
Ablation Studies
With careful ablation studies, we confirm all design choices of Chain of Code is essential for its good performance.
Scaling
Unlike Chain of Thought, which only emerges for large models, Chain of Code scales well with large and small models alike.