Recent Work
Models and systems I've built over the past couple of years.
Industry
Academia
Forecasting the future tests both knowledge and judgment. These papers ask how well language models judge what is plausible, predict outcomes and state their confidence, and revise their forecasts when new information arrives.
Language models can sound convincing without really reasoning. These papers explore how they represent meaning and relations, handle symmetry and time, and chain deductions from premise to conclusion. The open question is what it takes to make those abilities reliable.
- paperEvaluating the Semantic Specificity of Representation Steering in Language Models
- paperTCP: a Benchmark for Temporal Constraint-Based Planning
- paperCapturing Symmetry and Antisymmetry in Language Models through Symmetry-Aware Training Objectives
- paperCan Pretrained Language Models (Yet) Reason Deductively?
Checking a claim is rarely a simple yes or no lookup. AVeriTeC is a Cambridge project that builds systems to find the right evidence, ask the right questions, and show their work.