This repo contains the code, prompts, and model outputs for Deductive Verification of Chain-of-Thought Reasoning (NeurIPS 2023)
Chain-of-Thought prompting in Large Language Models (LLMs) emphasizes intermediate reasoning steps, which can inadvertantly introduce hallucinations and accumulated errors when solving complex tasks. To address this issue, we propose Natural Program, which harnesses the power of natural language and encourages LLMs to engage in explicit and rigorous deductive logical reasoning processes to solve tasks, just like humans do. By adopting Natural Program, we empower LLMs to perform effective self-verification of deductive reasoning processes. This is achieved through the decomposition of complex reasoning chains into step-by-step subprocesses that focus on necessary contexts and premises. As a result, we significantly enhance the rigor, reliability, and interpretability of reasoning steps and answers.
Here are teasers for our methods:
You need to first have an OpenAI API key and store it as the environment variable OPENAI_API_KEY (see here).
Package requirement: pip install openai mmengine
In the file of prompts, we have provided the prompts we used to instruct LLMs to generate reasoning chains in the Natural Program format. The Natural Program reasoning chains generated by ChatGPT are provided here for the 6 tasks we experimented with. In addition, we also provide prompts to verify deductive reasoning chains. The deductive verification results are provided in the verification results.
If you'd like to perform deductive verification using our approach on new tasks, please first generate Natural Program-based reasoning chains following the output format here. Then, you can perform deductive verifications of reasoning chains by following the instructions below.
To run deductive verification, use:
python run_verification.py --data-name gsm8k --input-result NATURAL_PROGRAM_PATH --output-result OUTPUT_FILE_PATH
Some key args:
--model-name: model name, defaults to gpt3.--data-name: dataset name, defaults to gsm8k.--input-result: input data file, following the format here--output-result: output result file, following the format here.--verify-mode: verification mode; supportsnaive(verify the entire reasoning chain at once);simultaneous(Natural Program-based deductive verification with step-by-step decomposition, where we use the prompt in Table 17 of the paper);sequential(Natural Program-based deductive verification, except that we split the grounding, reasoning, and calculation checks into 3 different prompts)--n: number of single-step verification samples on which we perform majority voting to determine its final validity (k'in Sec. 4.3 of our paper).
There is also an argument --ref-end, which assumes that the input Natural Program reasoning chains to be verified have premise references at the end of each reasoning step, rather than at the beginning of each reasoning step as in our paper. These reasoning chains are provided here. We observe that this allows final answer correctness to be improved (see table below), albeit with lower deductive verification accuracy.
| Formats | GSM8K | AQuA | Date |
|---|---|---|---|
| Prefix | 87.05 | 70.34 | 72.49 |
| Suffix | 88.49 | 71.29 | 77.40 |
data/human_annotation contains human annotations for the 6 tasks we experimented with. For each task, we sample 50 valid reasoning chains and 50 reasoning chains exhibiting mistakes. It has the following format:
{
"question", // question
"answer", // Natural Program reasoning chain output (to be verified)
"final_answer", // ground truth solution of this question
"correct", // final answer correctness
"flag": 1, // label given by annotator; flag=1 means the reasoning chain is valid; flag=0 means the reasoning chain has mistakes
}data/instruction_finetuning contains the prompts and responses we used to finetune Vicuna models to perform Deductive Verification of reasoning steps.
results/chatgpt3.5/natural_program/{verify_mode} contains all Natural Program reasoning chains generated by ChatGPT (GPT-3.5-turbo) before deductive verification. For each problem, we sample 10 candidate reasoning chains. The files have the following format:
{
"question", // question
"final_answer", // ground truth solution of this question
"example_idx", // example idx from the original dataset
"model_input", // the Natural Program prompt we used to instruct LLMs to generate Natural Program reasoning chains
"model_outputs", // the 10 candidate reasoning chains generated by LLM
"pred_answers", // the final answers extracted from the 10 candidate reasoning chains
"per_sample_correct", // whether each final answer is correct or not
"majority_results", // the final answer(s) based on majority voting over 10 candidates; note that there can be multiple results after majority voting if they receive the same number of votes
"majority_corrects", // whether each of the majority_results is correct
"majority_count", // number of final answers that are identical to the first majority result
"gt_count", // number of final answers that are identical to ground truth
"mean_expectation", // gt_count / num_candidates
"sample_idx_need_verify", // the ids of the reasoning chains that need to be verified; we verify the reasoning chains whose final answers receive the most and the second-most votes
}results/chatgpt3.5/verification/{verify_mode} contains everything in results/chatgpt3.5/natural_program/{verify_mode}, with the following additional content:
{
"verify_results", // a list of length(sample_idx_need_verify);
// each entry of which is a list of length(n_reasoning_steps) that contains raw verification outputs for each reasoning step of the current reasoning chain;
// each entry of which is a dict {'verify_model_inputs': n verification inputs for current step, 'verify_model_outputs': n validity outputs for current step};
// here we sample ``n=3`` validity prediction results for each reasoning step following the Uniformity phase of our Uniformity-Plurality voting strategy
"verify_correct", // a list of length(sample_idx_need_verify), the final verification result of each reasoning chain
"verify_result", // a list of length(sample_idx_need_verify);
// each entry of which is a list of length(n_reasoning_steps) that contains extracted verification results for each reasoning step of the current reasoning chain;
// each entry of which is a list of length (n=3) that contains the final extracted verification result from each validation prediction candidate of the current reasoning step
}Please cite our paper if you find our idea helpful. Thanks a lot!
@article{ling2023deductive,
title={Deductive Verification of Chain-of-Thought Reasoning},
author={Ling, Zhan and Fang, Yunhao and Li, Xuanlin and Huang, Zhiao and Lee, Mingu and Memisevic, Roland and Su, Hao},
journal={arXiv preprint arXiv:2306.03872},
year={2023}
}
This project is licensed under the CC-BY-4.0 License.