Discovering Language Model Behaviors with Model-Written Evaluations
Abstract
As language models (LMs) scale, they develop many novel behaviors, good and bad, exacerbating the need to evaluate how they behave. Prior work creates evaluations with crowdwork (which is time-consuming and expensive) or existing data sources (which are not always available). Here, we automatically generate evaluations with LMs. We explore approaches with varying amounts of human effort, from instructing LMs to write yes/no questions to making complex Winogender schemas with multiple stages of LM-based generation and filtering. Crowdworkers rate the examples as highly relevant and agree with 90-100% of labels, sometimes more so than corresponding human-written datasets. We generate 154 datasets and discover new cases of inverse scaling where LMs get worse with size. Larger LMs repeat back a dialog user's preferred answer ("sycophancy") and express greater desire to pursue concerning goals like resource acquisition and goal preservation. We also find some of the first examples of inverse scaling in RL from Human Feedback (RLHF), where more RLHF makes LMs worse. For example, RLHF makes LMs express stronger political views (on gun rights and immigration) and a greater desire to avoid shut down. Overall, LM-written evaluations are high-quality and let us quickly discover many novel LM behaviors.
Cite
@article{arxiv.2212.09251,
title = {Discovering Language Model Behaviors with Model-Written Evaluations},
author = {Ethan Perez and Sam Ringer and Kamilė Lukošiūtė and Karina Nguyen and Edwin Chen and Scott Heiner and Craig Pettit and Catherine Olsson and Sandipan Kundu and Saurav Kadavath and Andy Jones and Anna Chen and Ben Mann and Brian Israel and Bryan Seethor and Cameron McKinnon and Christopher Olah and Da Yan and Daniela Amodei and Dario Amodei and Dawn Drain and Dustin Li and Eli Tran-Johnson and Guro Khundadze and Jackson Kernion and James Landis and Jamie Kerr and Jared Mueller and Jeeyoon Hyun and Joshua Landau and Kamal Ndousse and Landon Goldberg and Liane Lovitt and Martin Lucas and Michael Sellitto and Miranda Zhang and Neerav Kingsland and Nelson Elhage and Nicholas Joseph and Noemí Mercado and Nova DasSarma and Oliver Rausch and Robin Larson and Sam McCandlish and Scott Johnston and Shauna Kravec and Sheer El Showk and Tamera Lanham and Timothy Telleen-Lawton and Tom Brown and Tom Henighan and Tristan Hume and Yuntao Bai and Zac Hatfield-Dodds and Jack Clark and Samuel R. Bowman and Amanda Askell and Roger Grosse and Danny Hernandez and Deep Ganguli and Evan Hubinger and Nicholas Schiefer and Jared Kaplan},
journal= {arXiv preprint arXiv:2212.09251},
year = {2022}
}
Comments
for associated data visualizations, see https://www.evals.anthropic.com/model-written/ for full datasets, see https://github.com/anthropics/evals