Google researcher advancing self-improving LLM agent harnesses and autonomous AI scientists for scientific discovery
3
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-17) 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—
Not yet scored. This axis measures publication volume and productivity over time — papers per year and venue spread. Per-year output counts are not available in the current bibliometric snapshot, so no estimate is offered.
Impact—
Not yet scored. This axis measures citation-based influence — citations, h-index, and referencing literature. No citation metrics exist in this snapshot, so the axis is left blank rather than estimated.
Focus0/100
Focus scores 0 because reviewed activity splits evenly — 50/50 — across two labeled fields: machine learning (LLM agents and automated harness/agent-system optimization) and artificial intelligence — autonomous research agents (AI for science). The split is thematically coherent rather than scattered: RRSI optimizes the agent harness itself, while ScientistTwo applies agent systems to automated scientific discovery.
Collaboration68/100
Collaboration scores 68, drawn from 29 distinct coauthors in the underlying data. Twelve recur across papers, including Xueqing Wu, Ashwin Balasubramanian, Dawei Zhu, Kai-Wei Chang, Nanyun Peng, Rujun Han, and Zifeng Wang. Authorship positions — middle on RRSI and PaperBanana-Interact, last on ScientistTwo — fit a pattern of contributing within larger Google-affiliated teams while leading at least one project.
Activity—
Not yet scored. This axis measures recent research tempo — publication recency and cadence across years. Year-level volume data is missing from this snapshot, so no activity estimate can be made.
Reproducibility38/100
Reproducibility scores 38 (reproMean 38), computed from only 2 reviewed papers. Evolve rates RRSI low: out-of-distribution gains across 8 benchmarks come without significance tests. ScientistTwo rates medium: 80.4% success on 107 papers, but headline evidence leans on AI reviewers inside its own loop. With so few papers, the score is indicative, not stable.
Research summary
Tomas Pfister is a researcher at Google (United States) whose reviewed work centers on automating machine learning research itself. The Evolve catalog places his 2026 output in two equally weighted strands — machine learning (LLM agents and automated harness/agent-system optimization) and artificial intelligence, autonomous research agents for AI-driven science — each carrying half of his labeled field activity. He contributes as a middle author to RRSI, which adds proposal and selection regularizers to recursive self-improvement of agent harnesses, yielding out-of-distribution transfer gains across 8 benchmarks — though no significance tests are reported and reproducibility readiness is rated low. He is last author on ScientistTwo, Google's autonomous research agent reported to improve 107 top-venue papers with an 80.4% success rate and a +25.2% average gain; Evolve cautions that the headline evidence leans on AI reviewers embedded in the system's own evaluation loop, and rates reproducibility readiness medium. A third paper, PaperBanana-Interact on multi-turn human feedback for scientific diagram refinement, is cataloged but not yet reviewed. Bibliometrically, the snapshot records 29 distinct coauthors (collaboration score 68) and a reproducibility score of 38 computed from 2 reviewed papers; output, impact, and activity axes lack sufficient data and are deliberately left unscored. The profile is name-resolved from public bibliographic data; no ORCID is provided.
Collaboration pattern
The snapshot records 29 distinct coauthors, giving a collaboration score of 68. Twelve coauthors recur, including Xueqing Wu, Ashwin Balasubramanian, Bingxuan Li, Dawei Zhu, Kai-Wei Chang, Yale Song, Yiwen Song, Rui Meng, Nanyun Peng, Peng Xia, Rujun Han, and Zifeng Wang. The pattern is large-team, Google-affiliated work: middle-author roles on RRSI and PaperBanana-Interact, with last-author leadership on ScientistTwo.
Research keywords
Highlights
Regularized recursive self-improvement for LLM agent harnesses
Last-author role on Google's ScientistTwo autonomous research agent
Multi-turn human feedback for scientific diagram refinement
PaperBanana-Interact (2026), with Pfister as a middle author, explores refining scientific diagrams through multi-turn human feedback. It is present in the Evolve catalog but awaits review, so no verdict, keywords, or reproducibility rating are available yet.
Timeline
- 2026RRSI, a middle-author paper on regularized recursive self-improvement of agent harnesses, appears on arXiv.
- 2026Last-author paper ScientistTwo reports autonomous AI improving 107 top-venue papers.
- 2026PaperBanana-Interact on multi-turn human feedback for diagram refinement is cataloged, pending Evolve review.
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.
- Radar dimensions are computed from public bibliometric snapshots; citation counts, h-index, and per-year publication volume were not available for this profile.
- Only papers reviewed in the Evolve catalog are reflected here; the author's other works, if any, are out of scope.
- No ORCID is provided; the profile is name-resolved from public bibliographic data and same-name authors may have been conflated.
- PaperBanana-Interact has not yet been reviewed by Evolve, so no verdict, keywords, or reproducibility rating exist for it.