SherlockBench

An LLM Benchmarking System

SherlockBench is a benchmarking system designed to test an LLM's ability to pro-actively investigate a problem. It is given a mystery function, and has to test it to determine what it does.

It tests "pro-active learning" which means it requires the LLM to reason about a strategy for it's own learning (in context).

LeaderBoard

SherlockBench is now saturated by GPT-6 Astra as-of September 2026. However the benchmark is still useful to test lesser models. Here is the scoreboard for the Irene2 problem-set.

Model Provider Average pass@3
GPT-6 Astra gpt-6-astra - high effort OpenAI 98% 100%
GPT-5.6 Sol gpt-5.6-sol - high effort OpenAI 84% 91%
Qwen3.5 9b 6 bit quant, reasoning enabled Qwen 42% 58%

I can't afford to do a pass@3 on Fable 5.1 because it's twice as expensive as Astra, but I did a pass@1 and it scored about the same as GPT-5.6 Sol. In future I'll focus on testing open-weight models unless someone wants to pay for it.