Machine-learning researcher on efficient 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—
Data insufficient: no publication-volume metrics are available. This axis would measure total output — number of papers, annual productivity, and venue tiers — none of which the current bibliometric snapshot provides, so no score is estimated.
Impact—
Data insufficient: no citation or influence metrics are available. This axis would measure scholarly impact — citation counts, h-index, and downstream uptake of the four arXiv preprints — none of which the input provides, so it is left unscored.
Focus0/100
The focus axis scores 0/100. The underlying field distribution splits evenly across four label variants — 'machine learning (LLM agents / context management)', 'machine learning (LLM agents and efficient context management)', 'machine learning — LLM agents and context management', and 'machine learning (NLP, LLM agents)' — each at 0.25. The measured dispersion thus reflects fragmented labeling of what is substantively one agenda: compression and memory control for long-horizon LLM agents.
Collaboration53/100
The collaboration axis scores 44/100, computed over 8 distinct coauthors. A stable core team recurs across all four 2026 papers — Mingxuan Wang, Fei Luo, Bo Wang, Guorun Yao, Yinglong Guo, Chao Ning, Hongyue Chen, and Jungong Han — indicating a persistent working group. All cataloged papers place this researcher in a middle authorship position with no corresponding-author role, which tempers the score.
Activity—
Data insufficient: no longitudinal cadence metrics are available. This axis would measure publication regularity and recent-year momentum over time, which cannot be computed from a snapshot whose cataloged papers all date to 2026.
Reproducibility46/100
The reproducibility axis averages 46/100 across 4 papers. Evolve rates three — GEM, StateComp, and the memory-control-signals study — at medium reproducibility readiness, and DRSR at low. Reviews repeatedly flag thin statistics: an unpaired 13-baseline comparison, single-run evaluations, LLM-generated labels, and missing error bars. Consistent reporting of token savings and reward deltas partially offsets these limitations.
Research summary
Yanbiao Ma is a machine-learning researcher whose reviewed work concentrates on efficient context management for long-horizon large language model (LLM) agents. The Evolve catalog holds four 2026 arXiv papers, all in middle authorship positions, forming a tightly clustered agenda. GEM compresses agent trajectories by 21.4% at equal reward, grounded in representation geometry and effective rank. StateComp learns when to compress history, cutting agent tokens 52.27% at near-flat reward. A probing study finds memory-control signals emerge in hidden states before action (AUROC .831/.765), supporting PaMER's roughly 73% token reduction at near-par scores. DRSR learns set-level deletion risk through counterfactual supervision, reporting +0.103 reward at -20.82% tokens. Bibliometric dimensions score collaboration at 44/100 over 8 distinct coauthors, focus at 0/100, and reproducibility at 46/100 averaged across 4 papers; output, impact, and activity lack sufficient data. The recurring caveat across reviews is statistical rigor: unpaired baseline comparisons, single-run evaluations, LLM-generated labels, and absent error bars. The profile is name-resolved from public bibliographic data; no affiliation or career facts appear in the input.
Collaboration pattern
Collaboration is concentrated in a single recurring team: all four 2026 papers draw on a pool of 8 distinct coauthors, including Mingxuan Wang, Fei Luo, Bo Wang, Guorun Yao, Yinglong Guo, Chao Ning, Hongyue Chen, and Jungong Han. The collaboration dimension scores 44/100. Every cataloged paper places this researcher in a middle authorship position with no corresponding-author role, suggesting a contributor within a stable group rather than an independent lab lead.
Research keywords
Highlights
GEM cuts agent tokens 21.4% at equal reward via representation geometry
StateComp learns when to compress history, saving 52.27% of agent tokens
Memory-control signals detected before action (AUROC .831/.765)
DRSR learns set-level deletion risk via counterfactual supervision
Timeline
- 2026Publishes DRSR, learning set-level deletion risk with counterfactual supervision (+0.103 reward at -20.82% tokens).
- 2026Publishes the memory-control-signals study, showing hidden states encode memory needs before action (AUROC .831/.765) and enabling PaMER's ~73% token cut.
- 2026Publishes StateComp, which learns when to compress history and cuts agent tokens 52.27% at near-flat reward.
- 2026Publishes GEM, cutting agent tokens 21.4% at equal reward through representation-geometry analysis.
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 are computed from OpenAlex-based public bibliometric data only; no citation counts, h-index, or venue metrics were provided.
- Affiliation, country, and ORCID are missing; the profile is name-resolved from public bibliographic data, and 'Yanbiao Ma' may match multiple individuals.
- Output, impact, and activity dimensions are unscored because the underlying data is insufficient.
- All four cataloged papers are 2026 arXiv preprints reviewed by Evolve; peer-review status and post-publication uptake are outside the provided scope.