"Contextualized Evaluations: Taking the Guesswork Out of Language Model Evaluations"

Adding synthetic context makes LLM evaluation more reliable and insightful

Context-aware evaluation reveals hidden biases and capabilities in LLM responses

https://arxiv.org/abs/2411.07237

Original Problem 🤔:

LLM users often issue underspecified queries where context (user identity, intent, response criteria) is missing. This leads to arbitrary evaluation judgments and unreliable benchmarking results.

-----

Solution in this Paper 🛠️:

→ Introduces "contextualized evaluations" - a protocol that adds synthetic context during LLM response evaluation

→ Context is represented as follow-up question-answer pairs that clarify the original query

→ Uses GPT-4, Claude-3.5-Sonnet and Gemini-1.5-Pro to generate and validate follow-up QA pairs

→ Implements three evaluation settings: standard (no context), implicit context discovery, and adaptive evaluation

-----

Key Insights from this Paper 💡:

→ 76% of queries in current LLM benchmarks are open-ended and lack proper context

→ Adding context during evaluation increases evaluator agreement by 3-10% absolute

→ Context can flip win rates between model pairs in comparisons

→ Default model responses show bias towards WEIRD (Western, Educated, Industrialized, Rich, Democratic) contexts

→ Models vary significantly in their ability to adapt to different contexts even when provided in prompts

-----

Results 📊:

→ Human validation shows 76% follow-up questions are important, with 90% realistic and 80% diverse answer sets

→ Context-aware evaluation improves inter-evaluator agreement from 64% to 78%

→ Context reduces evaluators' reliance on surface-level criteria like style by 5-7%

→ Models show varying instruction-following abilities, with performance differences of up to 1.64 points in constraint satisfaction