IC-522LLMs perform correct example inference without inducing the correct rule, and this gap is robust to prompting methods, fact count, and scenario form
Jiachun Li, Pengfei Cao, Zhuoran Jin, Yubo Chen, Kang Liu, Jun Zhao
Across five released LLMs evaluated on the MIRAGE dataset, rule induction (RI) accuracy is consistently and substantially lower than example inference (EI) accuracy in nearly all conditions. The gap persists when using chain-of-thought, self-consistency, self-refine, and hypothesis refinement prompting, when varying the number of observed facts, and when transferring rules across scenario formats. A change-rate control experiment confirms the two tasks have comparable intrinsic difficulty, ruling out that EI is simply easier to guess. A transferability test shows that a rule induced in one scenario does not help in another, further indicating the model is not internally representing and applying an abstract rule.
Evidence
correlational
Key metric
GPT-4o RI 0.16–0.44 vs EI 0.17–0.68; GPT-4 RI 0.15–0.58 vs EI 0.21–0.71; Claude-3.5 RI 0.13–0.46 vs EI 0.24–0.79; Llama3-8B RI 0.02–0.23 vs EI 0.09–0.30; Llama2-13B RI 0.00–0.21 vs EI 0.01–0.26 (d=3,5,8; 4 scenarios, n=5). CR (d=3,n=3): GPT-4o RI 0.74 vs EI 0.77; Claude-3.5 RI 0.81 vs EI 0.66. CoT 0-shot on GPT-4o: RI 0.50, EI 0.57 (LT) vs IO 0-shot RI 0.46, EI 0.76.
Caveat
The authors note they do not meticulously design prompts and use 0-shot IO with simple instructions; they also could not evaluate the o1 model due to API frequency limits. The evaluation is on a synthetic dataset with single-digit integer vectors.