Machine-learning researcher focused on efficient context and history compression for long-horizon LLM agents
4
papers here
4
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 metrics are available yet. This axis would measure the volume and cadence of publications — paper counts, venue placements and preprint-to-publication conversion. All four cataloged works are 2026 arXiv preprints, too early to score.
Impact—
Impact cannot be scored: the input contains no citation counts, h-index or venue-level metrics. This axis would measure citations, uptake and reuse of the four 2026 preprints once indexing services begin covering them.
Focus0/100
Evolve's radar currently records focus at 0, but the raw field distribution is unambiguous: all four cataloged papers carry near-identical field labels, each weighted 0.25, covering LLM agents, context compression and agent memory. Every paper targets the same problem — deciding when and what to compress in long-horizon agent histories — so the topical footprint is one narrow seam rather than a spread across areas.
Collaboration53/100
The collaboration dimension scores 44, built on 8 distinct co-authors across the four cataloged papers. The same names recur — Mingxuan Wang, Bo Wang, Guorun Yao, Yinglong Guo, Chao Ning, Hongyue Chen, Yanbiao Ma and Jungong Han — indicating one stable, small team rather than shifting partnerships. On every paper Luo sits in a middle authorship position without corresponding status, consistent with a contributing specialist inside a continuing group.
Activity—
Activity is unscored because no year-by-year publication counts or recency signals beyond the 2026 preprint cluster were provided. It would measure publishing tempo, recent output and sustained engagement over a career timeline.
Reproducibility46/100
Reproducibility scores 46 as a mean over four papers. Evolve rates readiness medium for three works (GEM, StateComp and the PaMER study) and low for DRSR. The weaknesses recur: single-run evaluations without statistical rigor, results reported without error bars, an unpaired 13-baseline comparison with untested reward parity, and labels generated by an LLM rather than humans. Reported token savings are concrete but await independent replication.
Research summary
Fei Luo is a machine-learning researcher whose cataloged work concentrates on one sharply defined problem: deciding when and what to compress in the interaction history of long-horizon LLM agents. Four 2026 arXiv preprints in the Evolve catalog attack it from complementary angles. GEM exploits representation geometry (effective rank) to keep agent memory stable, cutting agent tokens 21.4% at equal reward. StateComp learns the timing of compression, trimming 52.27% of agent tokens at near-flat reward. A third study finds that hidden states encode memory-control signals before action (AUROC .831/.765), enabling PaMER to cut tokens ~73% at near-par scores. DRSR learns set-level deletion risk through counterfactual supervision, reporting +0.103 reward at -20.82% tokens. Across all four papers Luo authors in a middle position without corresponding status, within a stable group of eight co-authors including Jungong Han and Mingxuan Wang. Evolve's radar scores collaboration at 44 and reproducibility at 46 (mean over four papers), reflecting flagged weaknesses: single-run evaluations, missing error bars, an unpaired 13-baseline comparison and LLM-generated labels. Output, impact and activity lack underlying bibliometric data and are left unscored.
Collaboration pattern
Luo works within a single recurring team: all 8 distinct co-authors appear across the four 2026 papers, including Mingxuan Wang, Bo Wang, Guorun Yao, Yinglong Guo, Chao Ning, Hongyue Chen, Yanbiao Ma and Jungong Han. Authorship is consistently middle-position and non-corresponding, suggesting a hands-on contributor rather than the group's lead. With a collaboration score of 44, the pattern points to a small, closed collaboration circle rather than wide external networking.
Research keywords
Highlights
GEM: 21.4% agent-token reduction at equal reward
StateComp learns when to compress agent history
Memory-control signals emerge before action
DRSR learns set-level deletion risk
Timeline
- 2026Publishes GEM, using representation geometry to cut long-horizon agent tokens 21.4% at equal reward.
- 2026Releases StateComp, learning state-based compression timing for 52.27% agent-token savings at near-flat reward.
- 2026Shows memory-control signals emerge in hidden states before action (AUROC .831/.765), enabling ~73% token cuts via PaMER.
- 2026Introduces DRSR, learning set-level deletion risk via counterfactual supervision (+0.103 reward, -20.82% tokens).
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 by Evolve from OpenAlex-derived public bibliographic data; no citation, h-index or venue metrics were available for output, impact or activity.
- All four cataloged papers are 2026 arXiv preprints; peer-review status and final venues are unknown.
- 'Fei Luo' is a common name; this profile is name-resolved from public bibliographic data, with no ORCID or affiliation provided to confirm identity.
- Institution and country are missing; collaboration findings cover only the co-authors on the four cataloged papers.