Skip to content
LessWrong· Ziqian Zhong·Indexed by this site · Originally published· Evaluation →· Its place in the scan record →SelectedAttention38

After removing a shortcut, LLMs perform below random level on the TASTE benchmark

LLMs are below chance on TASTE once a shortcut is removed

AI summary

Researcher Ziqian Zhong found that the TASTE benchmark contains a shortcut; after removing that shortcut, more capable models' judgments of AI safety research proposals fall below random level. The benchmark has models judge which of two AI safety research proposals experts prefer. The authors found that experts generally give lower ratings to proposals related to mechanistic interpretability; a baseline that simply selects the side with fewer interpretability-related terms in the proposal achieves 63.0% overall accuracy, and providing only the scoring rubric surpasses all tested LLMs. On the half of questions with no term difference, all 14 models with ECI above 158 fall below 50%, with 8 significantly below random, 7 of which are OpenAI models; the nine OpenAI models average only 29.6%, and performance worsens as ECI rises (r = -0.74). Providing experts' past judgments only partially mitigates this: models with ECI above 159 (including newer Anthropic and OpenAI models) still all fall below 50%, and the OpenAI models' average rises to 39.1%.

Why it was selected

Using a simple word-frequency baseline, the authors found that the TASTE benchmark contains a shortcut; after removing it, strong models' judgments of safety research proposals fall below random.

Read originallesswrong.com