Fei Huang — GRPO-based RL for LLMs: tool learning and reasoning-driven reranking; last author on 2026 arXiv work
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-02) 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 metric is available in the current snapshot, so this axis is unscored. It would measure total publications, venue tiers, and yearly production. Evolve can only confirm the two 2026 arXiv preprints reviewed so far.
Impact—
Impact cannot be scored: no citation counts, h-index, or uptake data are on file. This axis would measure citations and downstream adoption of the reviewed work. Any impact inference beyond the two preprints would be unsupported.
Focus0/100
The radar focus score is 0, matching the raw field distribution: coverage splits exactly 50/50 between machine learning (LLM tool learning / RL for agents) and information retrieval / NLP. Both catalog papers share one methodological core — GRPO-based RL fine-tuning for LLMs — so the even split reflects two application targets of a single method family rather than unrelated lines. Focus may rise as the catalog grows.
Collaboration48/100
Collaboration scores 48 on the radar, computed from 10 distinct coauthors across the reviewed papers. The recurring roster — Lijun Liu, Zhengzong Chen, Yuanyuan Zhao, Shihao Liu, Hao Yin, and Wenyan Li — points to a stable core team rather than a wide network. Fei Huang occupies the last-author position on both 2026 preprints, consistent with a senior role, although neither paper records them as the corresponding author.
Activity—
No year-by-year publication record is available, so activity is unscored. This axis would measure publishing cadence and recent output. The only dated evidence is two arXiv preprints, both stamped 2026 — too thin to establish a trend.
Reproducibility38/100
The radar reproducibility score is 38, the mean readiness across 2 assessed papers. MATCH rates medium: ablations are strong, but results come from single runs without variance tests. MERIT-Rank rates low: Evolve flags its efficiency claims and reproducibility for scrutiny. Both papers build on GRPO-based RL fine-tuning pipelines, so released code, training details, and seed variance will largely determine whether these reported gains replicate.
Research summary
Fei Huang is a machine-learning researcher whose reviewed work centers on reinforcement learning for large language models, applied along two fronts: tool learning for agents and reasoning-driven text reranking. The Evolve catalog currently holds two 2026 arXiv preprints, on both of which Fei Huang is the last-listed author — a position usually associated with senior or supervisory responsibility, though neither paper records them as the corresponding author. MATCH introduces model-aware tool learning that couples gated hierarchical rewards with a policy-tracking curriculum; Evolve's review notes strong ablations but single-run results without variance tests, and rates its reproducibility readiness medium. MERIT-Rank (Think Thrice Before Reranking) combines multi-perspective evidence and reasoning with staged RL to lift LLM reranking on BRIGHT, TREC, and BEIR; its efficiency claims and reproducibility are flagged for scrutiny, with a low readiness rating. The field distribution splits evenly (50/50) between LLM tool/agent learning and IR/NLP, and the radar focus score of 0 reflects that breadth; both papers share the same GRPO-based methodology, suggesting one method family aimed at two application targets. Collaboration is moderate (radar 48, 10 distinct coauthors), with a recurring team including Lijun Liu, Zhengzong Chen, Yuanyuan Zhao, Shihao Liu, Hao Yin, and Wenyan Li. Output, impact, and activity metrics are not yet computable. Note: 'Fei Huang' is a common name; this profile is name-resolved from public bibliographic data.
Collaboration pattern
Fei Huang works with a compact, stable team: 10 distinct coauthors appear across the reviewed papers, with Lijun Liu, Zhengzong Chen, Yuanyuan Zhao, Shihao Liu, Hao Yin, and Wenyan Li recurring. Holding the last-author slot on both 2026 preprints suggests a senior or supervisory role, though neither paper lists them as corresponding author. The radar collaboration score is 48 — moderate breadth, consistent with a small core group rather than a wide external network.
Research keywords
Highlights
MATCH: gated hierarchical rewards and a policy-tracking curriculum for LLM tool learning
MERIT-Rank: multi-perspective reasoning plus staged RL for text reranking
Timeline
- 2026MATCH preprint on RL-based tool learning with curriculum scheduling and hierarchically gated rewards appears on arXiv; Fei Huang is the last-listed author.
- 2026MERIT-Rank preprint on multi-perspective evidence and reasoning integration for text reranking appears on arXiv; Fei Huang is again the last-listed author.
- 2026Both catalog papers share a GRPO-based reinforcement-learning methodology, establishing a consistent methodological thread across tool learning and reranking.
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.
- Bibliometrics are limited to the 2 papers reviewed in the Evolve catalog (both 2026 arXiv preprints); no citation, venue, or publication-volume data are on file.
- Output, impact, and activity radar axes are unscored because the underlying metrics are missing from the current snapshot.
- No institution, country, or ORCID is available; 'Fei Huang' is a common name, so this profile is name-resolved from public bibliographic data and may not refer to a unique individual.
- The reproducibility score (38) is a mean over only 2 assessed papers, so it is sensitive to any single new review.