LetTokenBurn
Sign in
Auto86%Harbor: pendingAgentic / Tool UseInstruction Following

AgentPersonaBench

AgentPersonaBench: Benchmarking Persona-Driven User Simulation

A benchmark of 2,460 tasks across 867 traits evaluating latent persona adherence via observable actions in survey, chat, web, and app environments.

PaperRepo

How to run

ltb · agentpersonabench
$ npm i -g lettokenburn-cli
$ ltb bench info agentpersonabench

External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.

Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs

Scores

7 sourced entries
HarnessSourceConfidenceLink
GPT-6 Astra79.7N/AIndependentmediumOct 3, 2026arxiv.org
Claude Opus 5.576.6N/AIndependentmediumOct 3, 2026arxiv.org
GPT-5.6 Sol74.1N/AIndependentmediumOct 3, 2026arxiv.org
GPT-6 Sol71.5N/AIndependentmediumOct 3, 2026arxiv.org
GPT-5.6-terra64.5N/AIndependentmediumOct 3, 2026arxiv.org
GPT-5.6 Luna63.4N/AIndependentmediumOct 3, 2026arxiv.org
GPT-6 Luna61.4N/AIndependentmediumOct 3, 2026arxiv.org

Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.

Reproduced by the community

All runs
Loading community runs…

About

APB embeds target traits into complete synthetic profiles to test behavioral fidelity across four interaction surfaces: survey, chat, interactive web, and desktop apps. Ground truth is verified via deterministic code predicates on environment artifacts.

Scoring measures full-pass adherence rates. Leading models achieve up to 84.7% accuracy, though performance drops significantly when requiring consistency across all four modalities or handling multi-attribute demands.