Salesforce researcher spanning LLM post-training, agentic tool use, private MoE training, and efficient inference
4
papers here
4
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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—
No publication-volume metrics are available for this axis yet. It would measure total output — papers per year and venue mix — so Evolve cannot score productivity; the four cataloged 2026 arXiv preprints are the only output evidence here.
Impact—
Citation counts, h-index and venue influence are absent from the current snapshot. This axis would measure how widely the work is cited and adopted; until those data arrive, impact cannot be scored or estimated.
Focus0/100
The focus score is 0, and the underlying field distribution explains why: four Evolve-assigned fields each hold exactly 0.25 weight — differential privacy with sparse Mixture-of-Experts training, LLM post-training/reinforcement learning, LLM agents and post-training, and KV-cache compression. That is a perfectly even spread across four distinct areas rather than concentration, reading as broad applied LLM-systems work rather than a single narrow specialty.
Collaboration82/100
Collaboration scores 82 on a 0-100 scale, backed by 59 distinct coauthors in the snapshot and a recurring core of twelve frequent collaborators including Wenting Zhao, Jielin Qiu and Huan Wang. All four cataloged papers are middle-author team efforts rather than solo or corresponding-author work, consistent with a senior contributor embedded in large, rotating project teams.
Activity—
Recent-activity data — publication cadence by year — are not sufficient in this snapshot. The axis would measure ongoing research tempo; only four 2026 arXiv entries are present, so no yearly trend can be computed.
Reproducibility57/100
Reproducibility scores 57, computed as the mean readiness across all four cataloged papers. The spread is wide: the KV-cache eviction study rates high readiness with a rigorous, convincing null result, the belief-shift branching and DP-MoE papers sit at medium with single-run or untested-gain caveats, and the enterprise agent model is low, pending significance testing.
Research summary
Silvio Savarese is a researcher at Salesforce (United States) whose cataloged work concentrates on the engineering and training of large language models. The Evolve snapshot lists four 2026 arXiv preprints, all middle-author team efforts, spread evenly — 0.25 each — across four fields: differential privacy for sparse Mixture-of-Experts training, LLM post-training with reinforcement learning, LLM agents and post-training for enterprise tool use, and KV-cache compression for efficient inference. The pattern is coherent despite the spread: each paper targets a concrete bottleneck in making LLM systems cheaper, safer, or more capable in production. Notable entries include a belief-shift branching method for tree-structured RLVR with step-level credit assignment, a role-aware DP-SGD scheme for MoE with a parallel-composition guarantee, and a rigorous null result showing random per-head KV-cache eviction matches the strongest baselines — the catalog's highest reproducibility-readiness paper. Collaboration is the strongest measured dimension at 82/100, with 59 distinct coauthors and a recurring core of twelve frequent collaborators. Reproducibility readiness averages 57 across the four reviewed papers, ranging from low for the enterprise agent model to high for the KV-cache study. Output, impact and activity metrics are missing from the current snapshot, so productivity and citation reach cannot be assessed; all scores derive from OpenAlex-derived bibliometrics plus Evolve's own paper reviews.
Collaboration pattern
Collaboration is this profile's strongest measured dimension (82/100): the snapshot records 59 distinct coauthors and a recurring core of twelve frequent collaborators, including Wenting Zhao, Jielin Qiu, Huan Wang, Khai Le-Duc and Florent Draye. All four cataloged 2026 papers are middle-authored, non-corresponding team outputs, suggesting a senior contributor who joins large, rotating project teams rather than running single-author efforts.
Research keywords
Highlights
A convincing null result on KV-cache eviction
Role-aware differential privacy for sparse MoE
Fork placement for tree-structured RLVR
Enterprise agentic model with spec-driven GRPO
Timeline
- 2026Publishes Random Attention, a rigorous null result showing that random per-head KV-cache eviction with prompt protection matches the strongest baselines across 4 models and 6 tasks.
- 2026Publishes RAPTOR, applying role-aware differential-private training to sparse Mixture-of-Experts with a clean parallel-composition guarantee.
- 2026Publishes belief-shift branching for tree-structured RLVR, with fork placement validated pre-RL and across three RL setups.
- 2026Contributes to Salesforce Koa, an enterprise language model for agentic tool use, where spec-driven GRPO lifts BFCL multi-turn from 54.1 to 59.5.
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 dimensions are computed from OpenAlex-derived data only; output, impact and activity metrics were absent from this snapshot.
- No ORCID or country value was provided, so this profile is name-resolved from public bibliographic data and tied to Salesforce by the source record.
- All cataloged papers are 2026 arXiv preprints reviewed by Evolve; peer-reviewed versions and earlier publication history are out of scope.
- Author-position metadata comes from Evolve's catalog alone — all four entries are middle-author and non-corresponding — and may not reflect the complete authorship record.