Human-in-the-loop for LLM capability elicitation
Human participants are needed to tailor prompts and refine LLM outputs for accurate capability assessment.
Fully understanding information access requires testing with human participants.
Language models often provide better information when a human is involved to tailor prompts, correct mistakes, and follow up. This contrasts with automated benchmarking, which uses fixed questions and rubrics.
A human user can iteratively refine prompts to explore specific biological threat creation protocols, leading to more accurate and relevant information than a static set of questions.
Human participants are crucial for accurately assessing LLM capabilities, especially in complex domains like biological threat creation. Their ability to tailor prompts and refine outputs provides a more realistic evaluation than automated methods.
Automated benchmarking may suffice for simpler, well-defined tasks with less potential for nuanced misuse.
- Building an early warning system for LLM-aided biological threat creation ↗OpenAI · OpenAI News · Jan 31, 2024 · paraphrased
First seen on Subvolts Jan 31, 2024. Updated Sep 11, 2026. Spotted a problem? Tell us.
- Building an early warning system for LLM-aided biological threat creationOpenAI · OpenAI News · Jan 31, 2024
The Playbook: reusable how-tos, prompts, configs and tips pulled out of the news.
- HeuristicAuxiliary environment signals improve LLM authorization accuracyProviding LLMs with auxiliary environment signals can substantially improve their accuracy in authorization decisions. · 30 min
- HeuristicDecompose LLM vulnerability discovery into sub-tasksBreak down LLM vulnerability discovery into distinct, measurable steps for better evaluation. · 5 min
- Anti-patternRelying on gameable benchmarks for LLM vulnerability discoveryAvoid LLM benchmarks that are susceptible to data contamination or score recall against unknowable sets. · 5 min
- Case studyUsing LLMs to build tools for iOS location data analysisResearchers used an LLM to build a tool that decoded iOS location data, overcoming LLM analysis limitations. · 60 min
- HeuristicLLMs are unreliable for direct privacy analysis but good toolmakersUse LLMs to build analysis tools, not for direct forensic or privacy analysis. · 15 min
- Eval methodCheck LLM state policy recall for cross-jurisdiction substitutionVerify if an LLM provides another state's policy value when asked about a specific state. · 60 min
Questions people ask
- Which AI tools does this work with?
- Any current model. It was seen with in the sources.
- Where did this come from?
- OpenAI (OpenAI News). Subvolts collected, normalized and credited it; the link above opens the original.