Standard zero-shot chain-of-thought prompting (Kojima et al., 2022) elicits step-by-step reasoning by appending an explicit instruction such as "Let's think step by step" before greedily decoding the response. Wang & Zhou (2024), "Chain-of-Thought Reasoning Without Prompting," report finding reasoning paths in a pretrained model without adding any such instruction to the prompt at all. What do they change instead, and what do they observe as a result?
- Instead of adding any chain-of-thought instruction to the prompt, they change how the very first token of the response is decoded, inspecting the top-k alternative tokens rather than only the single highest-probability greedy token at that first step; branching down some of those alternative paths reveals chain-of-thought reasoning that was already latent in the pretrained model, and when such a path appears, the model tends to show higher confidence in its final answer
- They fine-tune the pretrained model on a small set of chain-of-thought demonstrations, after which greedy decoding alone reproduces step-by-step reasoning without needing the "let's think step by step" instruction
- They replace the pretrained model's tokenizer with one that segments numbers digit by digit, which the paper claims is solely responsible for eliciting latent reasoning at the decoding stage
- They add a hidden system-level instruction that is invisible to the end user but functionally identical to Kojima et al.'s "let's think step by step" phrase, appended before decoding begins
Why A? And why not the others?
Correct answer: A. Instead of adding any chain-of-thought instruction to the prompt, they change how the very first token of the response is decoded, inspecting the top-k alternative tokens rather than only the single highest-probability greedy token at that first step; branching down some of those alternative paths reveals chain-of-thought reasoning that was already latent in the pretrained model, and when such a path appears, the model tends to show higher confidence in its final answer
Wang and Zhou leave the prompt completely untouched and instead intervene at decoding time: at the first decoding step, rather than following only the single highest-probability greedy token, they branch out and inspect the top-k alternative tokens available at that step, then continue decoding down several of those alternative branches to completion. They find that some of these alternative branches naturally contain step-by-step reasoning that greedy decoding alone would never surface, showing that this reasoning ability is already latent in the pretrained model rather than something an explicit instruction creates from nothing; they also observe that when a decoded path does contain such a reasoning chain, the model's own probability assigned to its final answer tends to be higher, giving a usable signal for picking good paths. The option describing fine-tuning is wrong because the paper's central claim is that no additional training is needed; the reasoning is uncovered from an already-pretrained model purely by changing the decoding procedure. The option describing a digit-by-digit tokenizer swap is wrong because the method makes no change to tokenization at all. The option describing a hidden instruction is wrong because the entire point of the paper is that reasoning appears without any instruction, hidden or otherwise, added to the prompt.
Source: Wang & Zhou, 'Chain-of-Thought Reasoning Without Prompting' (arXiv:2402.10200, 2024)