OpenAI LASER Targets Rare Safety-Evaluation Cases
The recursive pipeline concentrates expensive reasoning graders on ambiguous policy boundaries, then widens the resulting evaluation set through embedding-based diversity sampling.
Edited by Tyronne Panaino
OpenAI described LASER on October 6, 2026 as a way to build safety-evaluation sets from rare, ambiguous conversations without grading a vast random sample with an expensive reasoning model. The method matters to model-safety teams because the hardest policy cases are often both uncommon and costly to find, yet they can expose whether a system handles sensitive requests consistently.
LASER stands for Logistic Augmented Sampling over Embeddings, Recursively. OpenAI says the pipeline can curate evaluation data within hours and use far less grader compute than random sampling. Those are first-party research and operational claims, not independent benchmark results.
A recursive loop targets policy boundaries
The process begins with synthetic conversations shaped by a safety policy, de-identified conversations similar to those examples and a random sample. A reasoning model labels the initial set as allowed or disallowed. LASER then fits a logistic-regression classifier over conversation embeddings and asks that cheaper model to find more examples near its decision boundary, where the predicted label is uncertain.
The newly selected conversations return to the reasoning grader, their labels update the classifier and the loop repeats. This design assigns the most expensive component to a narrower set of high-information examples instead of applying it uniformly across the entire candidate pool. It is an active-learning pattern: the current classifier helps decide which examples should be labeled next.
Diversity is a separate selection objective
Finding difficult cases does not guarantee a useful evaluation set if the selected conversations all resemble one another. OpenAI says LASER therefore adds greedy diversity sampling after identifying disallowed examples. Starting from one conversation, it repeatedly chooses the example farthest from those already selected in embedding space.
That final ordering is intended to give evaluators broader coverage when they can inspect only a limited number of cases. The distinction is important: uncertainty sampling concentrates on borderline judgments, while the diversity step tries to prevent those judgments from collapsing into one narrow cluster.
The compute claim has a specific denominator
OpenAI reports that roughly half of the conversations selected for grading in a typical LASER run are ultimately labeled disallowed. It contrasts that with a random-sample base rate that may be about one in 20,000, depending on the policy. On that basis, the company says LASER can require roughly 10,000 times less grader compute to find comparable numbers of disallowed examples.
This figure describes the compute needed to locate rare examples under OpenAI's stated setup. It does not by itself show that the resulting evaluation is more complete, that its labels are correct or that the same ratio will transfer to another policy, model or deployment. The page does not publish an external replication, a public comparison dataset or error analysis for the full pipeline.
Privacy limits access to the raw data
OpenAI says LASER operates on synthetic and de-identified conversations and does not access raw user data. That is a meaningful boundary for the described pipeline, but it is not a complete privacy audit. The source does not detail the de-identification procedure, residual re-identification risk or how representative the synthetic and selected conversations are across populations and languages.
The next useful checkpoints are independently reproducible evaluation sets, measured label quality, tests across additional policy domains and clearer evidence that the sampling strategy improves coverage rather than only reducing search cost.
Status
Learning. The architecture, privacy boundary and efficiency figures come from OpenAI's own research blog. Internal confidence is medium because the implementation and results were not independently reproduced in this run.
Sources
Update note: Last reviewed 2026-10-11. We will revise this post if OpenAI releases reproducible data, external validation or material changes to the method.
Sources
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.