Research Index
Systematic explorations into the limits of test-time compute and algorithmic patience.
ONGOINGFaithfulness of Chain-of-Thought Explanations in Small Language Models
Chain-of-Thought explanations suggest models reason step by step, but do they really? This study probes their causal role and finds that CoT can mislead about how small models make decisions.
Chain-of-ThoughtFaithfulnessInterpretabilitySmall Language Models
Read full paper →
OCT 4, 2025Scaling Test-Time Compute for Small Language Models: An Information-Theoretic Framework
Small models can't grow infinitely, so how do we get more out of them at inference? We formalize the link between test-time compute and reasoning fidelity, showing predictable gains when compute is allocated wisely.
Test-Time ComputeInformation TheoryScaling LawsSmall Language Models
Read full paper →