I am a third-year Ph.D. student in Computer Science at the University of Chicago, advised by Victor Veitch and Ari Holtzman. I’m developing ways to do precise causal inference over irreducibly subjective variables — the kind that matter for understanding LLMs but can’t be cleanly formalized. I think this is the biggest missing piece in our toolkit for analyzing the effect of AI on society at large.
News
- Your Causal Variables Are Irreducibly Subjective. Embrace the Shoe Leather. (blog post, 2026) — My musings on why causal mech interp has such a bad track record for reproducibility.
- Multiple Streams of Knowledge Retrieval: Enriching and Recalling in Transformers (Nief, Reber, Richardson, & Holtzman, ICLR 2026) — Dynamic weight grafting is a more surgical version of activation patching, locally patching computations rather than the information stored in the residual stream. Proof of concept: this reveals two separate pathways for retrieving finetuned knowledge.
- RATE: Causal Explainability of Reward Models with Imperfect Counterfactuals (Reber, Richardson, Nief, Garbacea, & Veitch, ICML 2025) — Precise causal inference over subjective-but-reproducible variables, applied to measuring how high-level attributes like helpfulness or sentiment actually drive reward model scores. It appears that “length bias” concerns about reward models were largely an artifact of biased measurement!
Contact
Email: davidpreber@gmail.com