🟦 Reversing Chain-of-Thought (RCoT) Prompting

Last updated on August 19, 2024 by Bhuwan Bhatt

What is Reversing Chain-of-Thought (RCoT) Prompting?

Chain-of-Thought (CoT) prompting¹ helps to improve the accuracy of Large Language Models (LLMs) on arithmetic reasoning tasks by forcing the model to think step by step before answering the question. However, LLMs still struggle to maintain factual consistency during reasoning and seem to overlook conditions, misinterpret the original questions, and may hallucinate. This is especially true when the original question contains multiple conditions. The example below shows an instance where ChatGPT overlooks the condition that the provided date in the question is tomorrow's date.

Reversing Chain-of-Thought (RCoT) prompting² aims to enhance LLMs' reasoning abilities by detecting and rectifying condition hallucinations. RCoT employs a 3 step process to detect and rectify hallucinations:

1. RCoT uses the original incorrect solution to construct a new problem (say Q')
2. RCoT conducts a fine-grained comparison between the original problem (Q) and reconstructed problem (Q') and detects hallucinations, overlookings, and question misinterpretation.
3. RCoT provides the detected shortcomings to the LLM in the form of feedback and guides it to the correct output.

How to Use Reversing Chain-of-Thought?

Let's use RCoT to solve a math word problem. The dialog below shows the question and initial LLM response:

Step 1: Problem Reconstruction
Reconstruct the problem by providing the question and initial response to the LLM.

As you can see, the reconstructed problem has different facts than the ones stated in the original problem:

Step 2 (i): Problem Decomposition
Now, let's decompose both problems into a list of conditions and facts stated in the problem.

Step 2 (ii): Condition Comparison
Detect hallucinated and overlooked conditions from the condition list.

• overlooked condition:

• hallucinated condition

Step 2 (iii): Question comparison
Compare the questions to check if the LLM can detect the difference in facts within the questions.

Step 3: Fine-grained feedback and revision
Collect all factual inconsistencies, provide feedback to the LLM, and instruct it to revise its solution.

What Are Reversing Chain-of-Thought Results?

• RCoT consistently outperforms standard CoT and the double-check (asks LLMs to check their answers but does not point out whether the answer is correct) methods in the zero-shot setting.

• The performance gain after employing RCoT is greater in challenging tasks than in easier tasks. For instance, ChatGPT's performance improves by 4.1% on AQuA and by 5.0% on the Date dataset. AQuA dataset contains diverse problems, and the Date dataset requires multi-hop reasoning and common sense date knowledge, making them complex compared to other datasets.
• On simpler datasets like the SVAMP dataset, which contains problems that usually only need one-step calculation, the performance improves but is less apparent. ChatGPT's performance improves by 2.8% in the SVAMP dataset.
• The result of employing RCoT with a few-shot setting is similar to the result obtained after employing RCoT using a zero-shot setting.
• RCoT can effectively detect overlooking and misinterpretation errors but struggles to detect hallucination errors.

Limitations of Reversing Chain-of-Thought

Some limitations of RCoT include:

• RCoT cannot detect all possible reasoning errors. For instance, RCoT struggles to detect computational errors.
• There is a significant difference between RCoT-generated feedback and human-generated feedback.
• RCoT involves multiple conversations with the LLM. This may consume additional resources (hardware, money, etc) and can also slow down the inference speed.

Conclusion

RCoT improves the reasoning abilities of LLMs by enabling them to detect and rectify factual inconsistencies in their generated solution. RCoT detects factual inconsistencies by fine-grained comparison between the original question and the reconstructed question using the original incorrect answer. It then reports the detected inconsistencies to the LLM through fine-grained feedback and guides the LLM to the correct solution.

Footnotes

1. Jason Wei. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. ↩

2. Tianci Xue. (2023). RCOT: Detecting and Rectifying Factual Inconsistency in Reasoning by Reversing Chain-of-Thought. ↩ ↩² ↩³