First-author arXiv researcher on compression and memory control for long-horizon LLM agents
4
papers here
4
reviewed by Evolve
4
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—
Output is not yet scored: publication-volume and venue metadata are insufficient for this name-resolved profile. This axis would measure total paper count, first-author share, and venue spread over time, none of which can currently be verified.
Impact—
Impact is not yet scored: citation counts, h-index, and downstream usage are absent from the available data. This axis would quantify how often the four 2026 arXiv papers are cited and adopted in agent-efficiency research.
Focus0/100
Focus is not yet scored: field and keyword metadata are null for all four catalog entries. This axis would measure how tightly the publication record concentrates on specific topics such as agent history compression.
Collaboration53/100
Evolve scores collaboration at 44 on a 0-100 scale, computed from public bibliometric data for this name-resolved profile. The raw evidence is 8 distinct co-authors — Fei Luo, Bo Wang, Guorun Yao, Yinglong Guo, Chao Ning, Hongyue Chen, Yanbiao Ma, and Jungong Han — recurring across a catalog in which Mingxuan Wang is first author on every paper, indicating a compact stable core team rather than a wide collaboration network.
Activity—
Activity is not yet scored: yearly output and recency signals rest on only four 2026 arXiv entries. This axis would track publication cadence and consistency across recent years once more records are linked.
Reproducibility46/100
Reproducibility is not yet scored: code-release and reproReadiness fields are null for every catalog paper. This axis would assess artifact availability, benchmark openness, and whether reported results can be independently replicated.
Research summary
Mingxuan Wang is an early-stage researcher whose Evolve profile is built from public bibliographic data and centers on efficiency for long-horizon LLM agents. The catalog contains four 2026 arXiv preprints, each first-authored: StateComp on learning when to compress interaction history, DRSR on learning set-level deletion risk, a study reporting that memory control signals emerge before action, and a method combining stable geometry with divergent task evidence for agent compression. Read together, they define a coherent program: deciding what an agent keeps, deletes, and compresses in its own history, framed as a learned control problem rather than a fixed truncation heuristic. Eight co-authors recur — Fei Luo, Bo Wang, Guorun Yao, Yinglong Guo, Chao Ning, Hongyue Chen, Yanbiao Ma, and Jungong Han — giving a collaboration score of 44 on Evolve's 0-100 radar; the other five axes (output, impact, focus, activity, reproducibility) are unscored because the underlying data is unavailable. No institution, country, ORCID, citations, or venue acceptances are attached, so impact and career-stage claims are withheld. Mingxuan Wang is a common name; this profile is name-resolved from public bibliographic records.
Collaboration pattern
Bibliometrics show 8 distinct co-authors in the current snapshot, all recurring: Fei Luo, Bo Wang, Guorun Yao, Yinglong Guo, Chao Ning, Hongyue Chen, Yanbiao Ma, and Jungong Han. Every catalog paper is first-authored by Mingxuan Wang with no corresponding-author flag, a pattern consistent with leading work inside a small stable team rather than broad external collaboration. The collaboration dimension score of 44 reflects this compact network.
Research keywords
Highlights
StateComp: learning when to compress agent history
First-author 2026 arXiv preprint that frames history compression in long-horizon agents as a learned decision problem — the model determines when compression should occur rather than applying a fixed truncation rule.
DRSR: set-level deletion risk for long-horizon agents
First-author 2026 arXiv preprint proposing to learn deletion risk at the set level, targeting efficiency gains for agents that must prune accumulated state over long horizons.
Memory control signals emerge before action
First-author 2026 arXiv study reporting that signals controlling memory manipulation appear ahead of the agent's action, suggesting memory decisions are anticipatory rather than purely reactive.
Stable geometry with divergent task evidence for agent compression
First-author 2026 arXiv preprint combining stable geometric structure with divergent task evidence as a basis for efficient long-horizon agent compression.
Timeline
- 2026First-author arXiv preprint 'Stable Geometry with Divergent Task Evidence for Efficient Long-Horizon Agent Compression' enters the catalog.
- 2026First-author arXiv preprint 'StateComp: Learning When to Compress History in Long Horizon Agents' enters the catalog.
- 2026First-author arXiv preprint 'Memory Control Signals Emerge Before Action in Long Horizon Agents' enters the catalog.
- 2026First-author arXiv preprint 'DRSR: Learning Set-Level Deletion Risk for Efficient Long-Horizon Agents' enters the catalog.
Papers in the Evolve catalog
- first author
Stable Geometry with Divergent Task Evidence for Efficient Long-Horizon Agent Compression
- first author
StateComp: Learning When to Compress History in Long Horizon Agents
- first author
Memory Control Signals Emerge Before Action in Long Horizon Agents
- first author
DRSR: Learning Set-Level Deletion Risk for Efficient Long-Horizon Agents
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.
- Identity is name-resolved from public bibliographic data; 'Mingxuan Wang' is a common name, and no ORCID, institution, or country is attached to this profile.
- Bibliometrics (collaboration score of 44, 8 distinct co-authors) come from the OpenAlex-based public snapshot; the other five radar axes lack sufficient data.
- All four papers are 2026 arXiv preprints with no venue, field, keyword, reproducibility, or Evolve verdict metadata yet.
- No citation counts, h-index, or venue acceptances are available, so impact and career stage cannot be assessed.