Michigan ML researcher: optimization for LLM pretraining (weight averaging) and LLM-based automated theorem proving
2
papers here
2
reviewed by Evolve
2
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-08) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.
Research profile
Output—
Publication-volume data (total and per-year output) is not yet available for this profile; this axis would measure cumulative papers and productivity over time. Only the two 2026 arXiv preprints in the Evolve catalog can currently be confirmed.
Impact—
Citation and influence metrics are not yet available; this axis would measure citations, h-index-style aggregates, and downstream uptake of the work. No claim about the reception of the two 2026 preprints can be made from current data.
Focus0/100
The focus score is 0 because the cataloged field distribution splits exactly evenly — 0.5 LLM pretraining optimization and 0.5 LLM-based automated theorem proving — rather than concentrating in one field. The two lines are tightly linked, though, by a shared theme of optimization and test-time compute for LLMs, so the low score reflects even two-way breadth within a coherent program rather than scattered work.
Collaboration22/100
The collaboration score is 22 on a 0–100 scale, matching a compact footprint of just 2 distinct coauthors across the two cataloged 2026 papers. Yixin Wang and Ambuj Tewari are the recurring collaborators, and Ousherovitch is first author on both papers (non-corresponding), a pattern consistent with focused, small-team work rather than large multi-institution collaborations.
Activity—
Recent-activity data (publication cadence, venue throughput over recent years) is not yet available; this axis would measure how consistently the researcher publishes over time. With only two 2026 preprints confirmed, no trend in research pace can be established.
Reproducibility55/100
Evolve's reproducibility assessment averages 55 across 2 scored papers, and both are rated medium readiness. This sits at a mid-range level: each paper's central claims are backed by concrete experiments — NanoChat pretraining runs for the TSA study and Lean ATP benchmarks for the search study — but the reported effects are small and the experimental scale is described as preliminary, so replication effort would be nontrivial.
Research summary
Adam Ousherovitch is a machine learning researcher at the University of Michigan whose work in the Evolve catalog spans the two halves of one problem: optimization for large language models. The catalog holds two 2026 arXiv papers, both first-authored, and the field distribution splits exactly evenly — machine learning for LLM pretraining optimization at 0.5 and LLM-based automated theorem proving at 0.5. One paper proposes Terminal Shrinkage Averaging (TSA), a shrinkage blend of the final iterate and a checkpoint average, with NanoChat experiments exposing an interaction between learning-rate schedules — including cooldown — and weight-averaging estimators such as LAWA. The other extends compute-aligned training to tree search in Lean automated theorem proving via weighted cross-entropy losses, pairing careful theory with consistent but modest gains over plain cross-entropy fine-tuning. The collaboration footprint is compact: 2 distinct coauthors, with Yixin Wang and Ambuj Tewari recurring, and a radar collaboration score of 22. Evolve's reproducibility assessment averages 55 across 2 papers, each rated medium readiness. With output, impact, and activity metrics still missing, this is best read as an early-stage profile of a focused, theory-driven optimizer.
Collaboration pattern
Within the Evolve catalog, Adam Ousherovitch's collaboration footprint is compact: 2 distinct coauthors across the two 2026 first-authored arXiv papers, with Yixin Wang and Ambuj Tewari as the recurring collaborators. Both papers list him as first author and not corresponding author, a pattern consistent with contributions made inside a small research group at the University of Michigan; the radar collaboration score is 22 on a 0–100 scale, placing the network at the smaller end of the catalog.
Research keywords
Highlights
Terminal Shrinkage Averaging links learning-rate schedules to weight-averaging estimators
Direct generator optimization brings compute-aligned training to Lean tree search
Timeline
- 2026First cataloged first-authored arXiv paper: 'Terminal Shrinkage Averaging Reveals a Schedule-Estimator Interaction in LLM Pretraining', proposing TSA with NanoChat experiments.
- 2026Second cataloged first-authored arXiv paper: 'Direct Optimization of Generators for Search in Automated Theorem Proving', applying weighted cross-entropy compute-aligned training to Lean tree search.
- 2026Both papers reviewed by Evolve, each rated medium reproducibility readiness; the reproducibility mean is 55 across the 2-paper catalog.
Papers in the Evolve catalog
Frequent co-authors
About this profile
This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.
- Bibliometric inputs and radar scores are computed from public bibliometric data (OpenAlex-style aggregation); coverage of non-indexed venues, code releases, or datasets may be incomplete.
- The profile rests on only 2 papers (both 2026 arXiv preprints); output, impact, and activity axes have insufficient data and are left unscored.
- No ORCID is linked, so identity is name-resolved from public bibliographic data and cross-source confirmation is limited.
- Both papers are preprints rated medium reproducibility readiness, with effects described by Evolve review as small and scale preliminary.