learn.ifs.name

Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

Wei et al., 2022, Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (Advances in Neural Information Processing Systems 35 (NeurIPS 2022))

Subscriber deck — free during early access

15 cards

1 Introduction

  1. What is chain-of-thought prompting?
    Few-shot prompting where each exemplar is a triple ⟨input, chain of thought, output⟩: the demonstrations include a series of intermediate natural-language reasoning steps before the final answer, so the model generates its own reasoning steps for the test question.
    invented exemplar: “A library has 12 shelves with 8 books each and lends out 30 books. How many books are left?”
    standardA: The answer is 66.
    chain of thoughtA: 12 shelves with 8 books each is 12 × 8 = 96 books. After lending 30, 96 − 30 = 66 are left. The answer is 66.
  2. Which two ideas does chain-of-thought prompting combine, and what limits did each have?
    Natural-language rationales help arithmetic reasoning, but training or fine-tuning on rationales needs costly annotated data; few-shot prompting needs no training, but standard prompting does poorly on reasoning and often barely improves with scale. CoT prompting gets rationales through prompting alone.
  3. What was the headline result of chain-of-thought prompting?
    PaLM 540B with eight CoT exemplars reached about 57% on GSM8K math word problems, a new state of the art above fine-tuned GPT-3 with a verifier (55%), whereas standard prompting got about 18%.

2 Chain-of-Thought Prompting

  1. What attractive properties of chain of thought did Wei et al. list?
    It lets models decompose multi-step problems and spend more computation on harder ones; it gives an interpretable (if incomplete) window into how an answer was reached; it applies in principle to any task humans solve via language; and it can be elicited from off-the-shelf models just by adding examples to the prompt.

3.1 Experimental Setup

  1. How were chain-of-thought experiments on math word problems set up?
    Five benchmarks (GSM8K, SVAMP, ASDiv, AQuA, MAWPS); one set of eight manually written CoT exemplars for all except AQuA (four from training); models GPT-3 (InstructGPT variants), LaMDA, PaLM, UL2 and Codex, with greedy decoding.

3.2 Results

  1. In what sense is chain-of-thought reasoning an emergent ability of scale?
    It doesn’t help small models, and only gives gains at around 100B parameters; smaller models produced fluent but illogical chains of thought and did worse than with standard prompting.
  2. For which problems does chain of thought help most?
    More complicated ones: on GSM8K, with the lowest baselines, performance more than doubled for the largest GPT and PaLM models; on MAWPS SingleOp, solvable in one step, gains were negative or very small.
  3. What did manual analysis of LaMDA 137B’s chains of thought on GSM8K show?
    Of 50 correct answers, all but two had logically and mathematically correct chains (two were right by coincidence). Of 50 wrong answers, 46% had nearly correct chains with minor errors (calculator error, symbol mapping, one missing step) and 54% had major semantic or coherence errors; scaling PaLM from 62B to 540B fixed many one-step-missing and semantic errors.

3.3 Ablation Study

  1. Which ablations showed that chain of thought isn’t just an equation, extra computation, or knowledge activation?
    Equation-only prompting barely helped on GSM8K (questions too hard to translate directly into equations), though it helped on 1–2-step data sets. Outputting a matching number of dots (variable compute only) did about as well as the baseline. Putting the chain of thought after the answer also did about as well as the baseline.

3.4 Robustness of Chain of Thought

  1. How robust is chain-of-thought prompting to who writes the exemplars?
    Chains written by three different annotators, a deliberately concise style, and exemplars sampled from GSM8K’s training set all beat standard prompting by a large margin despite some variance, so success doesn’t depend on a particular linguistic style; it was also robust to exemplar order and number.

4 Commonsense Reasoning

  1. How did chain-of-thought prompting do on commonsense reasoning?
    On CSQA, StrategyQA, Date and Sports Understanding (BIG-bench) and SayCan robot planning, it added gains on top of scaling, largest for PaLM 540B: e.g. StrategyQA 75.6% vs a prior best of 69.4%, sports understanding 95.4% vs 84% for an unaided sports enthusiast; gains on CSQA were minimal.

5 Symbolic Reasoning

  1. Which symbolic reasoning tasks did Wei et al. use, and what did they show?
    Last-letter concatenation (“Amy Brown” → “yn”) and coin-flip state tracking. With CoT, PaLM 540B solved in-domain cases almost perfectly, and CoT enabled length generalization to more words or flips than in the exemplars, where standard prompting failed; small models failed even in-domain.

6 Discussion

  1. What does chain-of-thought prompting imply about standard prompting as a measure of model ability?
    Standard prompting gives only a lower bound: on many reasoning tasks it has flat scaling curves, while chain-of-thought prompting turns them into steeply rising ones, expanding the set of tasks large models can do.
  2. What limitations of chain-of-thought prompting did Wei et al. acknowledge?
    It doesn’t answer whether the network is actually “reasoning”; annotation could be costly for fine-tuning; reasoning paths aren’t guaranteed correct, leading to both right and wrong answers; and it emerges only at large scale, which is costly to serve.

7 Related Work

  1. How does chain-of-thought prompting differ from instruction tuning and prompt learning?
    Those augment or learn the input part of the prompt (instructions or soft prompts); chain of thought takes the orthogonal direction of augmenting the model’s outputs with intermediate reasoning steps.

Source