Systems researcher at UW–Madison spanning efficient FP8 LLM training and agentic control of federated HPC resources
2
papers here
2
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 data is available in this snapshot, so output is not scored. This axis would measure papers per year, venue mix and authorship roles over time; it will appear once underlying catalog records are complete.
Impact—
No citation-based metrics were available, so impact is not scored. This axis would measure citation counts, h-index and uptake of the FP8 training-stability and HPC-agent work by other groups, once citation records can be linked.
Focus0/100
The focus score is 0 because the recorded field distribution is split exactly evenly: 0.5 to machine learning systems (efficient LLM training) and 0.5 to high-performance computing (agentic systems / distributed systems). The two reviewed 2026 papers occupy one field each — FP8 reinforcement-learning training stability on one side, the Diamond Agent for federated HPC control on the other. This reads as dual-track breadth across two systems areas rather than narrow specialization.
Collaboration59/100
The collaboration score is 59, computed from 18 distinct co-authors in the available records, 12 of whom recur frequently — among them Fanchao Chen, Ziheng Jiang, Ziyun Wei, Chi Zhang, Haibin Lin and Minu Mathew. The last-author position on the FP8 reinforcement-learning paper points to a senior supervisory role on a large training-systems team, while the middle-author Diamond Agent work adds collaborators from federated HPC resource management.
Activity—
Temporal data is insufficient, so activity is not scored. This axis would measure publication cadence and recent-year output; only two 2026 arXiv preprints are on file here, too few to establish any trend.
Reproducibility55/100
Reproducibility scores 55, the mean readiness rating across the 2 papers with reproducibility assessments, both rated medium. Evolve's review of the FP8 paper notes strong analysis but single-run evidence, which caps confidence in the reported training-stability gains, while Diamond Agent reports live results from 4 sites that are encouraging but small-scale. Releasing multi-seed FP8 runs and larger federated deployments would raise this axis.
Research summary
Shivaram Venkataraman is a researcher at the University of Wisconsin–Madison whose Evolve-reviewed record (ORCID 0000-0001-9575-7935) spans two systems areas in near-equal measure: machine learning systems for efficient LLM training, and high-performance computing built around agentic and distributed systems. In the first track, the 2026 arXiv preprint 'Towards Full Pipeline FP8 Reinforcement Learning for LLMs' — with this researcher in the last-author position — diagnoses how FP8 quantization noise in the PPO importance ratio leads to over-clipping and entropy collapse, and shows that Calibrated Clipping recovers scores close to BF16 baselines; Evolve rates the analysis strong but notes the evidence is single-run. In the second track, 'Diamond Agent: Agentic Control of Federated HPC Resources as a Service' demonstrates one agent steering jobs across multiple supercomputers using typed skills and event-driven continuation on Globus Compute, with encouraging live results from 4 sites at small scale. The radar profile scores collaboration at 59 (18 distinct co-authors, 12 recurring) and reproducibility at 55 (mean over 2 assessed papers, both medium readiness), while focus sits at 0, reflecting an exact 0.5/0.5 split between the two fields. Output, impact and activity metrics are not yet available for this snapshot, so the profile is an early-stage record built from two reviewed preprints plus bibliometric data.
Collaboration pattern
Collaboration is team-scale: 18 distinct co-authors appear in the records, and 12 recur frequently, including Fanchao Chen, Ziheng Jiang, Ziyun Wei, Chi Zhang, Haibin Lin and Minu Mathew. On the FP8 reinforcement-learning paper this researcher is the last author (not corresponding), consistent with a supervisory role on a large training-systems team; on Diamond Agent the position is middle. The records do not map individual co-authors to specific papers, so track assignments remain unverified.
Research keywords
Highlights
Calibrated Clipping stabilizes FP8 reinforcement learning for LLMs
Diamond Agent: one agent steering federated HPC resources as a service
Timeline
- 2026Last-author preprint on arXiv: 'Towards Full Pipeline FP8 Reinforcement Learning for LLMs', diagnosing FP8 importance-ratio noise, over-clipping and entropy collapse in PPO-style training.
- 2026Middle-author preprint on arXiv: 'Diamond Agent: Agentic Control of Federated HPC Resources as a Service', event-driven agent control demonstrated live across 4 supercomputer sites.
- 2026Both Evolve-reviewed papers land in the same year, splitting the field distribution evenly (0.5/0.5) between efficient LLM training and HPC/distributed systems.
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 scores are computed from OpenAlex-derived bibliometric records; output, impact and activity axes lack underlying data and are unscored.
- The reviewed record covers two arXiv preprints, both from 2026; earlier and peer-reviewed publications are not reflected in this profile.
- Reproducibility rests on only 2 assessed papers (reproMean 55), so that axis is coarse-grained.
- Identity is name-resolved from public bibliographic data with ORCID 0000-0001-9575-7935; country, career history and awards are not present in the input data.