Machine learning researcher on LLM agents — context compression and memory control for long-horizon 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 data is available yet: the record contains only the four reviewed 2026 preprints, with no full publication count. This axis would measure total research output and its distribution across years and venues.
Impact—
Impact cannot be scored yet: no citation counts or influence metrics exist in the available data. This axis would quantify citations to the four agent-compression papers and their uptake by other research groups.
Focus0/100
Focus scores 0, the lowest consistency band. The raw field distribution, however, lists four near-identical labels — machine learning with LLM agents and context management — each covering 25% of the record, so the zero appears driven by inconsistent field labeling across the four papers rather than genuinely scattered topics; substantively, all four papers target long-horizon agent history compression.
Collaboration53/100
Collaboration scores 44/100. The record shows 8 distinct coauthors across four 2026 papers, with a stable recurring group — Mingxuan Wang, Fei Luo, Bo Wang, Guorun Yao, Chao Ning, Hongyue Chen, Yanbiao Ma, and Jungong Han. Guo is a middle author on every paper and never corresponding author, consistent with a contributing member of a small team; the moderate score reflects limited evidence of independent leadership rather than absent collaboration.
Activity—
Activity data is insufficient: only four 2026 preprints are on record, all within a single year. This axis would measure publishing cadence and sustained research output across multiple years, which cannot yet be assessed.
Reproducibility46/100
Reproducibility scores 46/100, the mean across four scored papers: GEM, StateComp, and PaMER reach medium readiness, while DRSR is rated low. Evolve's verdicts consistently cite statistical thinness — single-run evaluations without error bars (StateComp, DRSR), an unpaired 13-baseline comparison with untested reward parity (GEM), and labels generated by an LLM rather than human annotation (PaMER) — even as the reported token savings are large.
Research summary
Yinglong Guo is a machine learning researcher whose reviewed work concentrates on one well-defined problem: making long-horizon LLM agents cheaper by deciding which parts of an agent's interaction history to keep, compress, or delete. All four papers in the Evolve catalog are 2026 arXiv preprints on this theme, approached from four angles: representation geometry (GEM, a 21.4% agent-token reduction at equal reward), learned compression timing (StateComp, 52.27% reduction at near-flat reward), pre-action memory-control signals decoded from hidden states (PaMER, roughly 73% reduction at near-par scores, with .831/.765 AUROC for predicting memory needs), and set-level deletion risk trained by counterfactual supervision (DRSR, +0.103 reward at -20.82% tokens). This is an early-stage profile: Guo appears as a middle author on all four papers and never as corresponding author, alongside 8 distinct coauthors including Jungong Han and Mingxuan Wang. Evolve's reproducibility scoring averages 46 across the four papers — three medium, one low — reflecting thin statistical reporting: single-run evaluations, missing error bars, an unpaired baseline comparison, and LLM-generated labels. Citation impact, output volume, and multi-year activity are not yet measurable, and no institutional affiliation is recorded.
Collaboration pattern
Guo works within a compact, recurring team: 8 distinct coauthors appear across the four 2026 papers — Mingxuan Wang, Fei Luo, Bo Wang, Guorun Yao, Chao Ning, Hongyue Chen, Yanbiao Ma, and Jungong Han. Guo holds middle authorship on every paper and is never the corresponding author, suggesting a steady contributor role in a small group rather than an independent lead. No cross-institution or industry collaboration data is recorded.
Research keywords
Highlights
StateComp: 52.27% agent-token reduction at near-flat reward
PaMER: memory-control signals emerge before action (AUROC .831/.765)
GEM: 21.4% token cut at equal reward via stable geometry
DRSR: learned set-level deletion risk via counterfactual supervision
Timeline
- 2026Publishes GEM (Stable Geometry with Divergent Task Evidence), reporting a 21.4% agent-token reduction at equal reward.
- 2026Publishes StateComp, learning when to compress history, with a 52.27% token reduction at near-flat reward.
- 2026Publishes the PaMER study showing memory-control signals emerge before action (AUROC .831/.765), cutting tokens by roughly 73%.
- 2026Publishes DRSR, learning set-level deletion risk via counterfactual supervision (+0.103 reward at -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.
- Profile is name-resolved from public bibliographic data only; no ORCID, institutional affiliation, or country is recorded, so identity confidence is limited.
- All bibliometrics derive from the Evolve-reviewed catalog: 4 arXiv preprints, all dated 2026; no citation impact, output-volume, or multi-year activity data is available.
- Reproducibility score (mean 46) comes from Evolve's review of 4 papers (3 medium, 1 low); it reflects reported statistical rigor, not attempted replications.
- Authorship data is limited to the four reviewed papers (all middle authorship, none corresponding); broader authorship history is unknown.