Machine learning researcher focused on context compression and memory control 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—
Insufficient data: no verified publication count, venue record, or career output total is available yet. This axis would measure cumulative papers and preprints over time and the tiers of venues they appear in.
Impact—
No citation counts, h-index, or adoption signals are available for this profile yet. This axis would measure citations, citing works, and uptake of the agent-compression methods by other research teams.
Focus0/100
The radar's focus axis is recorded at 0, and the raw field distribution supports full topical concentration: all four reviewed papers fall under machine learning for LLM agents and context management, split 0.25 across four near-identical label variants. This axis measures spread across topics; here every data point sits in one problem area — compressing and controlling agent history — rather than across diversified fields.
Collaboration53/100
Radar scores collaboration at 44, based on 8 distinct coauthors across the four reviewed papers: Mingxuan Wang, Fei Luo, Bo Wang, Guorun Yao, Yinglong Guo, Chao Ning, Yanbiao Ma, and Jungong Han. On every paper Chen holds a middle author position and is never the corresponding author, suggesting a contributing specialist within a small, stable team rather than an independent group lead.
Activity—
Insufficient data on publication cadence and recent momentum. This axis would measure papers per year and continuity of output; the four 2026 preprints are the only signal and are too thin to score.
Reproducibility46/100
Reproducibility averages 46 across 4 scored papers (reproMean 46, reproPapers 4). Evolve rates three papers medium readiness and one, DRSR, low. Verdicts cite recurring weaknesses: GEM's 13-baseline comparison is unpaired with untested reward parity, StateComp relies on single-run evaluations, PaMER's labels are LLM-made with thin statistics, and DRSR reports +0.103 reward at -20.82% tokens without error bars.
Research summary
Hongyue Chen is a machine-learning researcher whose reviewed work concentrates on a single, tightly defined problem: making long-horizon LLM agents cheaper by deciding what history to keep, compress, or delete. The Evolve catalog holds four 2026 arXiv preprints, all in this area and each attacking it from a complementary angle — geometry-aware state analysis (GEM, cutting agent tokens 21.4% at equal reward), learned compression timing (StateComp, 52.27% reduction at near-flat reward), pre-action hidden-state probes of memory needs (AUROC .831/.765, with PaMER cutting tokens ~73%), and set-level deletion-risk learning (DRSR, +0.103 reward at -20.82% tokens). On every paper Chen is a middle author and never the corresponding author, working alongside 8 distinct coauthors. Radar signals are mid-range: collaboration 44, reproducibility 46 averaged over 4 papers, and a focus reading of 0 against a field distribution entirely inside LLM-agent context management. Output, impact, and activity axes lack sufficient data. Evolve's verdicts repeatedly cite single-run evaluations, missing error bars, and LLM-generated labels, so reported gains are promising but statistically thin. This profile is name-resolved from public bibliographic data.
Collaboration pattern
Chen works within a compact, recurring team: 8 distinct coauthors (collaboration score 44), including Mingxuan Wang, Fei Luo, Bo Wang, Guorun Yao, Yinglong Guo, Chao Ning, Yanbiao Ma, and Jungong Han. Across all four papers Chen holds a middle author position and is never the corresponding author, consistent with a technical contributor role. The coauthor composition is essentially constant across the catalog, pointing to one stable collaboration cluster rather than many ad-hoc partnerships.
Research keywords
Highlights
GEM cuts agent tokens 21.4% at equal reward via geometry-aware compression
StateComp learns when to compress, halving agent tokens
Memory-control signals emerge before action, enabling ~73% token cuts
DRSR learns set-level deletion risk for history pruning
Timeline
- 2026arXiv preprint on GEM: geometry-aware compression cutting agent tokens 21.4% at equal reward for long-horizon agents.
- 2026arXiv preprint StateComp: learning state-based compression timing, reducing agent tokens 52.27% at near-flat reward.
- 2026arXiv preprint showing memory-control signals in hidden states before action (AUROC .831/.765), with PaMER cutting tokens ~73%.
- 2026arXiv preprint DRSR: set-level deletion risk via counterfactual supervision, reporting +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.
- Radar scores (collaboration 44, focus 0, reproducibility 46) are computed by Evolve from public bibliometric data; the output, impact, and activity axes have insufficient data and are not scored.
- All four reviewed papers are 2026 arXiv preprints with middle authorship; no peer-reviewed venue, institution, country, or ORCID is on record.
- Hongyue Chen is a common name; this profile is name-resolved from public bibliographic data and may reflect homonym aggregation.
- Evolve verdicts report single-run evaluations, unpaired baselines, missing error bars, and LLM-generated labels, so the quoted performance numbers should be read as preliminary.