IC-1159GPT-4 is a strong inductive hypothesis proposer, achieving high accuracy on inductive reasoning benchmarks when its generated rules are applied by a symbolic interpreter

Linlu Qiu, Liwei Jiang, Ximing Lu, Melanie Sclar, Valentina Pyatkin, Chandra Bhagavatula, Bailin Wang, Yoon Kim, Yejin Choi, Nouha Dziri, Xiang Ren

SourcePhenomenal Yet Puzzling: Testing Inductive Reasoning Capabilities of Language Models with Hypothesis Refinement

Using iterative hypothesis refinement (t=3, n=5), GPT-4 achieves 82.5% raw accuracy on ACRE, 93.3% on MiniScan, 71.2% on List Functions, and 18.7% on MiniARC, substantially outperforming standard IO prompting (64.0%, 61.7%, 65.1%, 33.1%) and self-consistency prompting. The approach also enables better OOD generalization: on List Functions OOD, rule prompting achieves 65.7% raw accuracy versus 46.1% for IO prompting. The key component is the symbolic interpreter providing accurate feedback; replacing it with the LM itself (self-refine) degrades performance substantially.

Evidence
correlational
Key metric
GPT-4 t=3,n=5 raw accuracy: ACRE 82.5%, MiniScan 93.3%, List Functions 71.2%, MiniARC 18.7%; task accuracy: 59.0%, 85.0%, 61.2%, 14.6%. IO prompting raw: 64.0%, 61.7%, 65.1%, 33.1%. OOD List Functions: rule 65.7% vs IO 46.1%.
Caveat
The authors note that MiniARC performance is far from optimal for all methods due to limited visual understanding in text-only models, and that some MiniARC tasks are heavily dependent on pattern matching where IO prompting may be more effective. They also note they did not exhaustively tune hyperparameters.
Model
GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report, GPT-3.5 / ChatGPT-3.5, Llama 2 / Llama 2 base Llama 2 70B
Datasets
ACRE [eval], MiniScan [eval], MiniARC [eval]
Methods
Hypothesis Refinement / Iterative hypothesis refinement [primary], Self-Consistency / Self-consistency prompting / Wang et al. 2023 (self-consistency) / Wang et al. 2023b (Self-Consistency) [compared-to], Self-Refine [compared-to]
Related work
Hypothesis search: inductive reasoning with language models [compared-to]
Related findings
IC-1157, IC-1158
Extraction
automatic-extraction