Machine-learning researcher on efficient context and memory management 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 score yet: the underlying publication-volume data — total papers, venues, and annual counts across indexed databases — is unavailable. Only four Evolve-reviewed 2026 arXiv preprints exist, too few to measure productivity.
Impact—
Impact is unscored: no citation counts, h-index, or adoption signals are present in the data. This axis would measure how often the four context-compression papers are cited and built upon by other work.
Focus0/100
The focus score is 0. Its raw input, the field distribution, splits evenly at 0.25 across four near-duplicate labels — all variants of machine learning, LLM agents and context management — so the taxonomy reads fragmentation while the actual portfolio is strikingly narrow: every reviewed paper targets compression or memory control for long-horizon LLM agents.
Collaboration53/100
Collaboration scores 53, computed from a raw count of 13 distinct coauthors. Ning holds a middle author position on all four papers and is corresponding author on none, fitting a contributing role within a stable team: Guorun Yao, Yanbiao Ma, and Yinglong Guo recur most often, alongside Mingxuan Wang, Fei Luo, Bo Wang, Hongyue Chen, and Jungong Han.
Activity—
Activity is unscored: the data cannot yet support a cadence measure. This axis would track publication rhythm and continuity over recent years; all four known papers fall in 2026, leaving no basis for a longer trend.
Reproducibility46/100
Reproducibility averages 46 across the four reviewed papers (reproMean 46, reproPapers 4). Three — GEM, StateComp, and the PaMER probe study — rate medium readiness; DRSR rates low. The recurring obstacle is statistical thinness: single-run evaluations, an unpaired 13-baseline comparison, missing error bars, and labels generated by LLMs rather than human annotation.
Research summary
Chao Ning is a machine-learning researcher whose reviewed work concentrates on a single problem: deciding what a large language model (LLM) agent should keep, compress, or delete from its history during long-horizon tasks. The Evolve catalog holds four 2026 arXiv preprints that sweep the design space of context management. GEM selects what to drop using representation geometry and effective rank, saving 21.4% of agent tokens at equal reward. StateComp learns when to compress, cutting 52.27% of agent tokens at near-flat reward. A probe study shows hidden states encode memory needs before action (AUROC .831/.765), letting PaMER remove roughly 73% of tokens at near-par scores. DRSR learns set-level deletion risk through counterfactual supervision, adding 0.103 reward while removing 20.82% of tokens. Evolve's verdicts are consistent: mechanism-level analysis is strong, but evaluation is statistically thin — single runs, unpaired baselines, no error bars, and LLM-generated labels. Ning appears as a middle author on all four papers, never as corresponding author, within a pool of 13 distinct coauthors, with Guorun Yao, Yanbiao Ma, and Yinglong Guo recurring most often. The computed focus score of 0 reflects fragmented field labels rather than real breadth — every paper targets the same theme.
Collaboration pattern
Ning works inside a compact, stable team: 13 distinct coauthors appear across the catalog, with Guorun Yao, Yanbiao Ma, and Yinglong Guo recurring most frequently, alongside Mingxuan Wang, Fei Luo, Bo Wang, Hongyue Chen, and Jungong Han. Holding a middle author position on all four papers with no corresponding role, Ning reads as a technical contributor within a group that produced multiple context-compression papers within a single year.
Research keywords
Highlights
GEM: geometry-aware compression cuts agent tokens 21.4% at equal reward
StateComp learns when to compress history, halving agent tokens
Hidden states signal memory needs before action (AUROC .831/.765)
DRSR learns set-level deletion risk via counterfactual supervision
Timeline
- 2026Publishes DRSR, learning set-level deletion risk through counterfactual supervision (+0.103 reward at -20.82% tokens).
- 2026Publishes the Memory Control Signals study: pre-action hidden states predict memory need (AUROC .831/.765), and PaMER cuts tokens by about 73%.
- 2026Publishes StateComp on state-based compression timing, cutting agent tokens 52.27% at near-flat reward.
- 2026Publishes GEM, a geometry-aware compression method saving 21.4% of agent tokens at equal reward.
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.
- Evidence base is limited to four Evolve-reviewed 2026 arXiv preprints; no peer-reviewed venue, citation, or publication-count data is included.
- No institution, country, or ORCID is on record, and Chao Ning is a common name; this profile is name-resolved from public bibliographic data and may not refer to a unique person (the coauthor list even contains Chao Ning itself, likely a data artifact).
- Output, impact, and activity dimensions are unscored because the underlying bibliometric data is insufficient.
- The reproducibility score (46) averages only four papers, and the focus score (0) partly reflects fragmented field labels rather than genuinely scattered research topics.