← News

Models & toolsSources: arXiv · 20 Nov 2026

PosteriorBench tests how well language models can be steered toward rare behaviors

Researchers at Transluce AI released a benchmark that scores sampling algorithms on cost and accuracy when drawing model outputs showing a chosen behavior.

In plain words

A language model rarely produces certain outputs, so researchers want cheap ways to sample them on purpose. The new benchmark checks whether such methods truly match the model's own odds of producing those outputs. On its tasks, no tested method was both cheap and accurate.

What was announced

The benchmark, described in an arXiv paper, defines the task as follows: given a model and a target behavior, find an algorithm that efficiently draws transcripts showing that behavior from the model's distribution. PosteriorBench includes 21 tasks built on the Qwen3.5-0.8B model, with target behaviors ranging in probability from one in a hundred thousand to one in ten. Each task comes with exact reference samples drawn from a corpus of one billion labeled responses, and methods are scored on compute cost plus how closely their samples match the true distribution in forward and reverse KL divergence, a standard measure of distance between probability distributions.

The authors report a consistent gap: no evaluated method was both efficient and faithful. Prompting is inexpensive but stalls at a large divergence from the target distribution, fine-tuning on reference samples is often too costly, and several earlier training approaches showed no clear gains. The team frames building a cheap, accurate posterior sampler as an open problem and positions the benchmark as a way to track progress. Code and data are posted on GitHub.

What it does not show

  • The evidence is the paper's abstract only; the full paper's methods and results were not reviewed, and the reported findings have not been independently confirmed.

Sources

Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 21 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.