Researcher on persona-stable LLM dialogue, agent memory compression, and vision-language physical-reasoning evaluation
3
papers here
3
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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 metrics are unavailable in this snapshot: there are no publication counts, venue records, or productivity totals beyond the three reviewed preprints. This axis would measure cumulative publication volume, venue tiers, and lead- or corresponding-author share.
Impact—
Impact cannot be scored yet: no citation counts, h-index, or downstream usage data are present in the input. This axis would measure citations, citing works, and uptake of the three papers by the wider research community.
Focus0/100
The focus score is 0, reflecting an unusually even field split: the distribution assigns about 33.3% each to vision-language model evaluation (physical reasoning), NLP on LLM personalization and multi-turn dialogue, and AI on LLM agent memory. With one paper per area and no dominant field, this is a deliberately broad three-front portfolio rather than a concentrated specialization, so a topical focus cannot yet be established.
Collaboration61/100
Collaboration scores 61, computed from 20 distinct coauthors across the three catalogued papers, with 12 recurring collaborators including Sourajit Saha, Shubhashis Roy Dipta, Fangming Liu, Jing Li, and Qiheng Wang. On every paper Li sits in a middle authorship position without corresponding-author status, consistent with contributing within established teams, and the coauthor pool spans all three research lines rather than one single long-term group.
Activity—
Activity cannot be scored: the snapshot has no longitudinal cadence or submission history, and all three catalogued papers fall in 2026, too short a window to gauge research tempo. This axis would track output rate and research continuity over time.
Reproducibility55/100
Reproducibility averages 55 across three assessed papers, each rated medium readiness. The work ships concrete artifacts — paired benchmark designs, routing ablations, and memory-compression numbers — but each carries a caveat: headline gains in the personalization paper rest on a GPT-5.1 judge, the vision-language study covers only six open models in an exploratory setting, and MemForest reports compression and retention figures without significance tests.
Research summary
Siyuan Li is a researcher working across three connected corners of modern large language model systems: personalization in multi-turn dialogue, memory management for LLM agents, and evaluation of vision-language models on physical reasoning. The current Evolve catalog holds three 2026 arXiv preprints, in each of which Li is a middle-listed author rather than corresponding author. The first paper routes persona updates through COMMIT/DEFER/IGNORE decisions to mitigate persona drift, supported by broad baselines and ablations though headline gains rest on a GPT-5.1 judge. The second contributes a paired benchmark testing whether vision-language models change their experiment selection when new evidence arrives, a rigorous design that remains exploratory with six open models. The third, MemForest, reports 50% agent-memory compression with 97-99.7% retention and speedups, though without significance testing. Bibliometrically, Li's work splits evenly across three fields (roughly 33% each), with 20 distinct coauthors and 12 frequent collaborators, yielding a collaboration score of 61 and a reproducibility mean of 55 across three assessed papers. That even field split — focus scored 0 — marks a broad, exploratory three-front portfolio rather than a narrow specialization; output, impact, and activity metrics are not yet available in the snapshot.
Collaboration pattern
Li's three papers draw on 20 distinct coauthors, with 12 frequent collaborators — among them Sourajit Saha, Shubhashis Roy Dipta, Nobin Sarwar, Shaswati Saha, Yuxuan Jiang, Qiheng Wang, Youyuan Zhang, Fangming Liu, Jing Li, Junxi Wang, Te Sun, and Jiayi Zhu. Holding middle authorship on all three papers without corresponding status, Li appears to contribute within established teams, with the collaborator pool spread across the personalization, vision-language, and agent-memory lines rather than concentrated in one group.
Research keywords
Highlights
COMMIT/DEFER/IGNORE routing mitigates persona drift in multi-turn dialogue
Paired benchmark tests whether VLMs change experiment choice with new evidence
MemForest reports 50% agent-memory compression at 97-99.7% retention
Timeline
- 2026Publishes 'MemForest', proposing EventTree partitioning and progressive merging for efficient LLM agent memory management.
- 2026Publishes 'New Evidence, Same Choice', a paired benchmark probing physical experiment selection in vision-language models.
- 2026Publishes 'Toward Robust Personalized Alignment for LLMs', addressing persona drift in multi-turn dialogue via uncertainty-aware COMMIT/DEFER/IGNORE routing.
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 come from an OpenAlex-based snapshot; only the three Evolve-reviewed arXiv preprints are covered, with no venue-acceptance status.
- Output, impact, and activity axes are unscored — no citation counts, publication totals, or multi-year cadence data are available.
- 'Siyuan Li' is a common name; this profile is name-resolved from public bibliographic data, and no ORCID or institutional affiliation was provided.
- All three papers list Li in a middle authorship position; individual contribution levels are inferred only from catalog metadata.