IC-931GPT-4 achieves approximately 95% accuracy on contrast instructions, far exceeding human performance without tools

Lingfeng Shen, Sihao Chen, Linfeng Song, Lifeng Jin, Baolin Peng, Haitao Mi, Daniel Khashabi, Dong Yu

SourceThe Trickle-down Impact of Reward Inconsistency on RLHF

The paper evaluates GPT-4 on the contrast instructions benchmark by prompting it to determine which of two instruction-response pairs would be more preferred by humans. Across the four datasets (Stack, WMT, Twitter, RealSumm), GPT-4 achieves 95.5% average response consistency (cres) and 96.0% average instruction consistency (cins). This substantially exceeds the human performance without tools (82.8% cres, 81.7% cins) and is comparable to human performance with tools (96.3% cres, 96.1% cins). The result validates that contrast instructions is a solvable task and that the near-random performance of standard RMs reflects a genuine limitation of the training procedure rather than task impossibility.

Evidence
correlational
Key metric
GPT-4: cres 91.2 / 98.4 / 96.8 / 95.4 (avg 95.5), cins 91.8 / 98.8 / 97.2 / 96.2 (avg 96.0) across Stack / WMT / Twitter / RealSumm
Caveat
GPT-4 was prompted with a specific instruction to judge preference; the result is sensitive to the exact prompt formulation. The evaluation is a single-turn judgment, not a full reward-modeling task.
Model
GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report
Datasets
StackExchange [source], WMT [source], RealSumm [source]
Related findings
IC-930
Extraction
automatic-extraction