Evaluating Language Models as Risk Scores | Social Foundations of Computation – Max Planck Institute for Intelligent Systems

Institute Homepage

Institute Homepage Sign In

Research Overview

Social Prediction

Performative Prediction: Past and Future

Difficult Lessons on Social Prediction from Wisconsin Public Schools

Allocation Requires Prediction Only if Inequality Is Low

Digital Platforms, Power and Work

Performative Power

An Engine Not a Camera: Measuring Performative Power of Online Search

Causal Inference from Competing Treatments

Contesting Algorithmic Systems

Algorithmic Collective Action in Machine Learning

Decline Now: A Combinatorial Model for Algorithmic Collective Action

Algorithmic Collective Action in Recommender Systems: Promoting Songs by Reordering Playlists

Algorithmic Fairness

Fairness and Machine Learning: Limitations and Opportunities

Unprocessing Seven Years of Algorithmic Fairness

Science of Machine Learning Benchmarks

ImageNot: A Contrast with ImageNet Preserves Model Rankings

Don't Label Twice: Quantity Beats Quality when Comparing Binary Classifiers on a Budget

A Theory of Dynamic Benchmarks

Evaluating Large Language Models

Inherent Trade-Offs between Diversity and Stability in Multi-Task Benchmarks

Evaluating Language Models as Risk Scores

Fine-tuning Large Language Models

Training on the Test Task Confounds Evaluation and Emergence

Lawma: The Power of Specialization for Legal Tasks

Social Foundations of Computation Members Publications

Evaluating Language Models as Risk Scores

Folktexts — Folktexts encodes tabular data rows as prompts, runs inference on a model, and decodes a risk score from the language model. This makes it possible to use LLMs on tabular data in much the same way that sklearn would do it.

We created a Python package called folktexts that makes it easy to evaluate language models as risk scores. This allows us to demonstrate that instruction-tuned language models yield strongly miscalibrated risk scores.

Members

Thumb ticker sm andre innsbruck face 2

Social Foundations of Computation

André Cruz

Doctoral Researcher

Thumb ticker sm 20241104 hardt moritz 12 cleaned kleiner

Social Foundations of Computation

Moritz Hardt

Director

Thumb ticker sm portrait celestine

Algorithms and Society

Celestine Mendler-Dünner

Hector Endowed Fellow of the ELLIS Institute

Publications

Social Foundations of Computation Algorithms and Society Conference Paper Evaluating Language Models as Risk Scores Cruz, A. F., Hardt, M., Mendler-Dünner, C. Advances in Neural Information Processing Systems 37 (NeurIPS 2024), The Thirty-Eighth Annual Conference on Neural Information Processing Systems (NeurIPS), December 2024 (Published) ArXiv Code URL BibTeX