Tsinghua researcher advancing efficient context management and memory compression 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 metrics are available in the current data; this axis would measure publication volume, venue tier, and yearly consistency. The reviewed set covers only four 2026 arXiv preprints.
Impact—
Impact data is insufficient; this axis would measure citations, derivative works, and adoption of the compression methods. No citation counts or usage signals appear in the provided snapshot.
Focus0/100
The computed focus score is 0, but the raw field distribution tells a narrower story: all four cataloged papers sit at 0.25 apiece under four near-identical labels for machine learning, LLM agents, and context management. The zero likely reflects fragmented field tagging in the source data rather than genuinely scattered research; on the reviewed evidence, this profile is tightly concentrated on one problem area — deciding what agents can safely forget.
Collaboration53/100
With a collaboration score of 53 and 13 distinct coauthors across the four papers, this researcher works with a compact, stable group. The frequent-coauthor list repeats Guorun Yao, Yanbiao Ma, Yinglong Guo, and Chao Ning, alongside Mingxuan Wang, Fei Luo, Bo Wang, and Hongyue Chen. All four cataloged papers place this researcher in the last author position, consistent with a senior supervisory role.
Activity—
Activity data is missing; this axis would measure publishing cadence and recent output intensity. Only the year 2026 is represented, so no multi-year activity trend can be computed.
Reproducibility46/100
Reproducibility scores 46 across all four rated papers (reproMean 46, reproPapers 4). Evolve's verdicts flag a consistent pattern: StateComp and the PaMER study rest on single-run evaluations with thin statistics, DRSR reports +0.103 reward at -20.82% tokens without error bars, and GEM's 13-baseline comparison is unpaired with reward parity untested. Three papers rate medium repro-readiness, one rates low.
Research summary
Jungong Han, affiliated with Tsinghua University, is a senior last author in machine learning research on large language model agents, centered on efficient context management for long-horizon tasks. All four papers reviewed by Evolve — GEM, StateComp, the PaMER memory-control study, and DRSR — are 2026 arXiv preprints, each carrying Han in the last author position. The agenda is coherent: deciding what an agent can safely forget. GEM exploits stable representation geometry against divergent task evidence to cut agent tokens 21.4% at equal reward; StateComp learns when to compress history, removing 52.27% of agent tokens at near-flat reward; the PaMER study shows hidden states encode memory needs before action (AUROC 0.831/0.765) and cuts tokens by roughly 73%; DRSR learns set-level deletion risk through counterfactual supervision, adding +0.103 reward while dropping 20.82% of tokens. Evolve rates three papers medium and one low for reproducibility, repeatedly citing single-run evaluations, missing error bars, unpaired baselines, and LLM-generated labels. The collaboration score is 53 across 13 distinct coauthors with a stable core team; output, impact, and activity dimensions lack sufficient data for scoring.
Collaboration pattern
This researcher appears as last author on all four cataloged 2026 papers, a consistent senior position. The collaboration score is 53 with 13 distinct coauthors. A recurring core group — Guorun Yao, Yanbiao Ma, Yinglong Guo, and Chao Ning — appears twice each in the frequent-coauthor list, suggesting the same team across multiple papers, joined by Mingxuan Wang, Fei Luo, Bo Wang, and Hongyue Chen. No external or cross-institution collaborators are visible in the reviewed set.
Research keywords
Highlights
GEM cuts agent tokens 21.4% at equal reward using representation geometry
StateComp halves agent tokens by learning when to compress history
Hidden states encode memory needs before action; PaMER cuts tokens ~73%
DRSR learns set-level deletion risk via counterfactual supervision
Timeline
- 2026Publishes DRSR, learning set-level deletion risk for long-horizon agents via counterfactual supervision and teacher-forced likelihood (+0.103 reward at -20.82% tokens).
- 2026Shows memory-control signals emerge in agent hidden states before action (AUROC 0.831/0.765); PaMER cuts tokens by about 73%.
- 2026Introduces StateComp, learning when to compress history and cutting agent tokens 52.27% at near-flat reward.
- 2026Releases GEM, using stable representation geometry against divergent task evidence to cut agent tokens 21.4% 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.
- Bibliometric dimensions are derived from OpenAlex-based public bibliometric data; citation counts, h-index, and venue-level metrics are not part of this profile.
- All four reviewed papers are 2026 arXiv preprints; no peer-reviewed venue record appears in the reviewed set.
- ORCID and country are not provided; the profile is name-resolved from public bibliographic data, which can be ambiguous for common names.
- Output, impact, and activity dimensions lack sufficient underlying data and are left unscored rather than estimated.