Bbh Prompting Eval

bbh-prompting-eval

Invalid Logic, Equivalent Gains: The Bizarreness of Reasoning in Language Model Prompting — Schaeffer et al. (2023) (arXiv:2307.10573, 2023)

What this evaluates

Tests the impact of prompt structure and logical validity on language model reasoning performance. Specifically, it compares answer-only, standard chain-of-thought, and logically invalid chain-of-thought prompting strategies on complex reasoning tasks.

Datasets

• BIG-Bench Hard — total ?; splits: test (-1)

Metrics

• accuracy (primary) — range: [0, 1]
  • Standard classification accuracy: the proportion of correctly predicted answers out of the total number of questions.

Input / output format

Input: Task description, few-shot examples, and a target question formatted according to one of three prompt types: Answer Only (AO), Chain-of-Thought (CoT), or Logically Invalid Chain-of-Thought (CoT Invalid).

Output: A single predicted answer (e.g., a number, word, or multiple-choice option) corresponding to the target question.

Scoring recipe

correct = 0
for pred, gold in zip(predictions, gold_labels):
    if pred == gold:
        correct += 1
accuracy = correct / len(gold_labels)

Common pitfalls

• Assuming logically invalid CoT prompts are ineffective; the study shows they often achieve performance gains comparable to logically valid CoT prompts.
• Overlooking that some previously published 'valid' CoT prompts already contain logical errors, which confounds claims that performance gains stem from genuine reasoning.

Evidence (verbatim from paper)

The Accuracy of Codex on each of the BIG-Bench Hard (BBH) tasks under each of the three prompt types (Answer Only, Chain-of-Thought, Logically Invalid Chain-of-Thought) is displayed in Fig. 3. We found that Chain-of-Thought prompting beats Answer Only prompting, but Logically Invalid Chain-of-Thought is close behind Chain-of-Thought and better than Answer Only (Fig. 4 top).

Citation

@misc{schaeffer2023invalidlogic,
  title={Invalid Logic, Equivalent Gains: The Bizarreness of Reasoning in Language Model Prompting},
  author={Schaeffer et al. (2023)},
  year={2023},
  note={arXiv:2307.10573}
}

• arXiv: 2307.10573