IC-1158GPT-4 and other LMs are brittle to noisy exemplars and unfamiliar output representations, with performance degrading sharply even under minimal perturbation

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

On List Functions, GPT-4's raw accuracy drops from 71.2% (0% noise) to 50.9% (12.5% noise), 42.4% (25%), and 29.1% (50%), with task accuracy falling from 61.2% to 10.0%. Explicitly instructing the model to consider noisy examples mitigates but does not eliminate the drop. On MiniScan, replacing English output words with pseudowords reduces raw accuracy from 93.3% to 86.8% and task accuracy from 85.0% to 72.0%. The same brittleness is observed for GPT-3.5 and Claude-2.

Evidence
correlational
Key metric
GPT-4 List Functions raw accuracy: 71.2% (0%), 50.9% (12.5%), 42.4% (25%), 29.1% (50%); task accuracy: 61.2%, 30.4%, 18.8%, 10.0%. GPT-4 MiniScan: raw 93.3% (English) vs 86.8% (pseudowords); task 85.0% vs 72.0%.
Caveat
The authors note that with only 8 exemplars, perturbing 1-2 elements may make the task ill-defined, and humans are not necessarily robust to noisy observations either. However, the relative performance drop of the LM is more significant than that of humans.
Model
GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report, GPT-3.5 / ChatGPT-3.5
Concepts
Failure mode
Datasets
MiniScan [eval]
Methods
Hypothesis Refinement / Iterative hypothesis refinement [primary]
Related findings
IC-1157, IC-1159
Extraction
automatic-extraction