MIT researcher on efficient LLM reasoning, post-training distillation, and token-efficient agent harnesses.
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-15) 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 output-volume data is available yet. This axis would measure publication counts, output cadence, and authorship positions over time; until such records are ingested, it is left unscored rather than estimated.
Impact—
Impact metrics are not yet available. This axis would measure citations, uptake by other researchers, and downstream use of this work; the current snapshot carries no such signals, so no score is reported.
Focus0/100
The focus score is 0 of 100, driven by the field distribution in the reviewed sample: 50% machine learning (LLM post-training and efficient reasoning) and 50% AI agents / LLM systems (harness-level token efficiency). With exactly two papers, one in each area, no dominant field emerges within this snapshot. The even split reads either as an early, still-concentrating profile or as a deliberate bridge between post-training methods and agent-systems engineering.
Collaboration56/100
The collaboration score is 56 of 100, based on 16 distinct coauthors in the current snapshot. A recurring core includes Yecheng Wu, Han Cai, Haozhe Liu, Tian Ye, Sensen Gao, Qihang Cao, Yitong Li, Mingchen Zhuge, Duomin Wang, Ruihua Zhang, Ping Luo, and Jiawang Bian. Authorship positions — last author on SoL-Pi and middle author on Lightning Weave — suggest involvement across both team-led and mid-list collaborative work rather than large consortia.
Activity—
Recent-activity data is insufficient. This axis would measure publication frequency and ongoing research engagement across recent years; the snapshot spans only two 2026 records, so no activity score is computed.
Reproducibility55/100
The reproducibility score is 55 of 100, computed as the mean readiness across 2 reviewed papers, each rated medium. Evolve's review of SoL-Pi credits a solid isolation protocol but flags single-run statistics and a capability drop on TB4; Lightning Weave is described as broad and credible yet single-run with no seed variance. Multi-seed reporting and released artifacts would most directly lift this axis.
Research summary
Song Han is a researcher at the Massachusetts Institute of Technology whose cataloged recent work sits at the intersection of efficient large language model reasoning and LLM-agent systems. The current snapshot contains two 2026 arXiv papers. SoL-Pi, with Song Han as last author, recursively scales auto-research loops to produce an agent harness matching a Pi baseline at roughly 45% fewer tokens on EdgeBench; Evolve credits its isolation protocol while flagging single-run statistics and a capability drop on TB4. Lightning Weave, a middle-author paper, proposes capability composition through Tilted-Target DOPD to lift the accuracy-efficiency frontier of reasoning models, described as broad and credible though single-run and without seed variance. Bibliometrically, the reviewed sample splits evenly — 50% machine learning (LLM post-training and efficient reasoning), 50% AI agents / LLM systems (harness-level token efficiency) — yielding a focus score of 0. The collaboration score of 56 rests on 16 distinct coauthors, with a recurring core including Yecheng Wu, Han Cai, and Haozhe Liu. Reproducibility averages 55 across two papers, both rated medium readiness. Output, impact, and activity axes remain unscored because the underlying data is not yet available.
Collaboration pattern
Song Han's reviewed work is collaborative, with 16 distinct coauthors in the snapshot (collaboration score 56). The recurring core spans both papers' areas: Yecheng Wu, Han Cai, Haozhe Liu, Tian Ye, Sensen Gao, Qihang Cao, Yitong Li, Mingchen Zhuge, Duomin Wang, Ruihua Zhang, Ping Luo, and Jiawang Bian. Authorship ranges from last author on SoL-Pi to middle author on Lightning Weave, suggesting both senior-oversight and embedded-contributor roles.
Research keywords
Highlights
SoL-Pi: auto-research loops cut agent-harness token cost
Lightning Weave lifts the accuracy-efficiency frontier of reasoning models
Timeline
- 2026SoL-Pi appears on arXiv: recursively scaled auto-research loops yield an agent harness matching Pi at roughly 45% fewer tokens on EdgeBench (last-author position).
- 2026Lightning Weave appears on arXiv: Tilted-Target DOPD composes capabilities to lift the accuracy-efficiency frontier of reasoning models (middle-author position).
- 2026Both papers enter the Evolve catalog with medium reproducibility readiness, anchoring a 50/50 split between LLM post-training work and agent-systems token efficiency.
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 scores derive from a narrow snapshot of two Evolve-reviewed 2026 papers; the output, impact, and activity axes lack the underlying data.
- Profile is name-resolved from public bibliographic data under ORCID 0000-0002-4186-7618; identity beyond this record is not independently verified.
- Both papers are arXiv preprints (2026); peer-review status, venue acceptance, and citation counts are not captured in this data.
- Reproducibility and collaboration scores rest on only 2 papers and 16 coauthors, and may shift as more of the catalog is ingested.