Studies how model behavior develops during training, using public checkpoints and controlled experiments on memorization and capabilities.
Why read Stella Biderman?
Read Biderman when you want to look beyond the score of a finished model. She coauthored Pythia, a suite whose training checkpoints and data order are available for comparison. Keeping the data sequence consistent across model sizes lets researchers investigate when a behavior appears, while reducing differences caused by unrelated training choices. The paper explains why those controls matter for studying a language model.
Her coauthored memorization study turns the checkpoints into a concrete question: can cheaper experiments predict which training passages a larger model will reproduce? Its results and corrections show why the measurement definition and missed predictions matter. Her 2026 position paper gives an accessible entry to the broader research agenda: study how behavior develops, then test whether an intervention changes that trajectory.
Start with the original
Selected work
Position paper ·
Position: Don't Just "Fix it in Post": A Science of AI Must Study Training Dynamics
Coauthored position paper. Sections 2.1 and 2.2 explain the difference between describing a finished model and predicting behavior during training. Start here without the experimental papers' technical setup.
Research paper ·
Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling
Coauthored research releasing a controlled model suite. Section 2 explains the shared data order and saved checkpoints; section 3 shows what questions that infrastructure makes testable.
Research paper ·
Emergent and Predictable Memorization in Large Language Models
Coauthored research. Inspect the exact-continuation definition before interpreting the predictions. The corrected results and Pythia-only experiments limit how far they can be generalized.
Projects & roles
EleutherAI
The research organization she identifies on her current personal site as its executive director.
Pythia
The team's public hub for model checkpoints, training details and experiments on learning dynamics.
Community rating
Sources & further reading
- Her current site identifies the EleutherAI executive director role and her research interests.
- Her publication list links the selected works and records their coauthors.
- Current X metadata names Stella Biderman and links her personal site.
- Read for the controlled suite's construction, checkpoint availability and case studies.
- Read for the memorization definition, prediction goal, corrected analysis and limitations.
- Read for the position paper's prediction and training-trajectory argument; it is an agenda, rather than proof of a completed theory.