David S. Berman

Research profile

Research summary

David S. Berman is a researcher at Queen Mary University of London (GB) whose recently reviewed work sits at the interface of machine-learning theory and quantum many-body physics. Two first-author arXiv preprints dated 2026 anchor the catalog. 'Do Quantum Models Scale Like LLMs?' tests whether LLM-style power-law data scaling carries over to RydbergGPT, a transformer trained on Rydberg atom-array samples; Evolve finds the scaling law holds only near quantum criticality, on controlled but single-size, data-axis-only evidence. 'Optimal Pruning for Neural Architectures using Fisher Information Distances' uses information geometry to unify magnitude- and Fisher-based importance measures; Evolve judges it sound and transparent, with toy-scale evidence tempering LLM extrapolation. Record-level OpenAlex bibliometrics report h-index 49, i10-index 167, 24 works in the last three years and a compact recent network of 6 distinct co-authors including Roger G. Melko and Ying-Jer Kao. The same record, however, also indexes paleontology and seventeenth-century political thought and counts reaching back to 1957, so its statistics may conflate same-name authors and are quoted as-is. Both reviewed papers rate high on reproducibility readiness.

Research keywords

Highlights

• LLM-style scaling laws hold for RydbergGPT only near quantum criticality

  Evolve's reviewed verdict on this first-author 2026 preprint: power-law data scaling observed in LLMs carries over to RydbergGPT on Rydberg atom arrays only near quantum criticality. The evidence is controlled but limited to a single system size and the data axis, so it is suggestive rather than conclusive.

• Fisher information distances unify neural-network pruning criteria

  In this first-author 2026 preprint, Fisher-Rao information geometry yields a single distance-based importance measure that recovers magnitude pruning and Fisher-based pruning within one framework. Evolve judges the method sound and transparent with high reproducibility readiness, while noting the experiments are toy-scale and extrapolation to LLMs needs tempering.

• One agenda: information geometry across scaling and compression

  The two reviewed papers share a common thread — applying geometric and statistical measures (mutual information near criticality, Fisher information distances) to questions usually posed about large language models. Both are first-author 2026 arXiv preprints rated high for reproducibility, marking a coherent, early-stage line of work.

Timeline

1. 2016countsByYear records 13 indexed works, the highest annual figure of the late 2010s in the OpenAlex record.
2. 202115 indexed works, tied with 2004 as the annual maximum in the record's countsByYear series.
3. 20245 indexed works, a dip from the 12 works recorded in 2022.
4. 20258 indexed works, a partial rebound from the 2024 figure.
5. 202611 indexed works so far; two first-author arXiv preprints — the RydbergGPT scaling-law study and the Fisher-distance pruning paper — enter the Evolve catalog, both rated high for reproducibility.

Papers in the Evolve catalog

• 2026-09-17first author

  Do Quantum Models Scale Like LLMs?

• 2026-09-14first author

  Optimal Pruning for Neural Architectures using Fisher Information Distances

Frequent co-authors