Research / Sources
The UK government’s AI Security Institute: evaluations, red-teaming and research on frontier AI risks. aisi.gov.uk
Introduces Transect, an open-source tool that builds interactive reports from agentic evaluation transcripts so reviewers can follow runs and check LLM-judge interpretations against source passages.
Introduces optstop, an open-source Inspect-integrated tool that stops LLM evaluation runs once estimates are precise enough, cutting compute while preserving capability elicitation.
Shows that fixed test-time compute caps underestimate frontier agent capability, especially on long tasks and for newer models, so evaluations should account for compute budgets.
AISI and Irregular find frontier models use 10-50x larger token budgets productively on cyber tasks, so standard capped evaluations may underestimate dangerous cyber capabilities.
Presents Inspect Scout, an open-source tool and seven-step pipeline for systematic transcript analysis of AI evaluations, to surface failure modes and safety-relevant behaviours.
AISI argues that transcript analysis should supplement agent pass rates, using nearly 6,400 transcripts to surface behaviour patterns, refusals and failure modes in safety-relevant evaluations.
AISI blog presenting a standardised protocol for capability elicitation experiments, so dangerous-capability evaluations better avoid underestimating what models can do.
Proposes Bayesian GLMs to use LLM autograders for model evaluation while simultaneously estimating their reliability and systematic biases, from AISI's evaluation-science work.
Introduces HiBayES, a hierarchical Bayesian framework for statistically analysing hierarchical, small-sample, stochastic LLM evaluation results with principled uncertainty quantification.
Presents AISI's Long-Form Task methodology, an automatically graded evaluation of how helpfully LLMs guide users through scientific tasks, as a scalable proxy for human uplift studies.
AISI shares lessons from months of building open-ended question-answer evaluations for frontier models, covering question design, automated grading, and why they complement multiple-choice tests.