Andi Han

Research profile

Research summary

Andi Han is a researcher at The University of Sydney working at the intersection of learning theory, optimization, and mechanistic interpretability. The OpenAlex-resolved profile (ORCID 0000-0003-4655-655X) shows an early-career record — h-index 5 and i10-index 2 — with a steep activity trajectory: recorded works rise from 2 in 2020 to 12 in 2024 and 31 in 2026, with 48 works in the most recent three-year window. Topic modeling concentrates on stochastic gradient optimization techniques (27), advanced optimization algorithms (18), and advanced graph neural networks (14), while the field distribution splits evenly across machine learning (mechanistic interpretability), theoretical computer science (computational learning theory / inductive inference), and statistical learning theory. Three 2026 papers reviewed by Evolve illustrate this range: a Lean-checked characterization of language generation in the limit via finite witnesses with a separation-width hierarchy reaching ω+1; nearly tight Rademacher bounds for sparsely activated networks that remove a √n factor with matching lower bounds; and a survey connecting feature superposition theory to interpretability practice. Evolve's reproducibility review of three papers averages 58, with readiness spanning high to low. With 8 distinct coauthors and a collaboration score of 44, the record describes a theory-driven contributor whose citation footprint (h-index 5) has not yet caught up with an unusually fast output pace.

Research keywords

Highlights

• Lean-checked answer to an open problem in language generation in the limit

  In this 2026 paper (middle author), the work answers Raman et al.'s open characterization of generation in the limit via finite witnesses, attaining every separation-width level up to ω+1; Evolve finds the Lean-checked core rigorous and rates reproducibility high.

• Nearly tight Rademacher bounds for sparsely activated networks

  As last author, this 2026 paper removes ADKM's √n factor up to a log^{3/2}(2m) term with matching lower bounds. Evolve notes the proofs are self-contained, but core revisions were AI-assisted, and rates reproducibility medium.

• Bridging feature superposition theory and interpretability practice

  This 2026 survey (middle author) connects superposition theory to mechanistic interpretability practice — sparse autoencoders, dictionary learning, polysemanticity — via a rigorous L1–L6 claim taxonomy and honest hedging; Evolve praises the care but notes it reports no new experiments (reproducibility low).

Timeline

1. 2020First recorded works appear in the OpenAlex countsByYear series (2 works).
2. 2022Recorded output reaches 7 works as the year-over-year climb begins.
3. 202412 works recorded — the strongest complete year before 2026 in countsByYear.
4. 2025Recorded works dip to 5, likely an indexing artifact (see caveats).
5. 2026Peak year: 31 recorded works, including three Evolve-reviewed arXiv papers spanning inductive inference, statistical learning theory, and mechanistic interpretability.

Papers in the Evolve catalog

• 2026-09-09

  Characterizing Language Generation in the Limit: Finite Witnesses and a Separation-Width Hierarchy

• 2026-09-08

  Nearly Tight Rademacher Bounds for Sparsely Activated Neural Networks

• 2026-09-06

  Feature Superposition in Neural Networks: From Theory to Practice

Frequent co-authors