Research - Papers
Explore a selection of our published work on a variety of key research challenges in AI.
On Linear Representations and Pretraining Data Frequency in Language Models
Pretraining data has a direct impact on the behaviors and quality of language models (LMs), but we only understand the most basic principles of this relationship. While most work focuses on…
Generalization v.s. Memorization: Tracing Language Models' Capabilities Back to Pretraining Data
The impressive capabilities of large language models (LLMs) have sparked debate over whether these models genuinely generalize to unseen tasks or predominantly rely on memorizing vast amounts of…
Holistically Evaluating the Environmental Impact of Creating Language Models
As the performance of artificial intelligence systems has dramatically increased, so too has the environmental impact of creating these systems. While many model developers release estimates of the…
WildBench: Benchmarking LLMs with Challenging Tasks from Real Users in the Wild
We introduce WildBench, an automated evaluation framework designed to benchmark large language models (LLMs) using challenging, real-world user queries. WildBench consists of 1,024 tasks carefully…
Trust or Escalate: LLM Judges with Provable Guarantees for Human Agreement
We present a principled approach to provide LLM-based evaluation with a rigorous guarantee of human agreement. We first propose that a reliable evaluation method should not uncritically rely on…
OLMoTrace: Tracing Language Model Outputs Back to Trillions of Training Tokens
We present OLMoTrace, the first system that traces the outputs of language models back to their full, multi-trillion-token training data in real time. OLMoTrace finds and shows verbatim matches…
Skilful global seasonal predictions from a machine learning weather model trained on reanalysis data
Machine learning weather models trained on observed atmospheric conditions can outperform conventional physics-based models at short- to medium-range (1-14 day) forecast timescales. Here we take the…
Understanding the Logic of Direct Preference Alignment through Logic
Recent direct preference alignment algorithms (DPA), such as DPO, have shown great promise in aligning large language models to human preferences. While this has motivated the development of many…
CodeScientist: End-to-End Semi-Automated Scientific Discovery with Code-based Experimentation
Despite the surge of interest in autonomous scientific discovery (ASD) of software artifacts (e.g., improved ML algorithms), current ASD systems face two key limitations: (1) they largely explore…
A Little Depth Goes a Long Way: The Expressive Power of Log-Depth Transformers
Recent theoretical results show transformers cannot express sequential reasoning problems over long input lengths, intuitively because their computational depth is bounded. However, prior work…