Tianmin Shu

Johns Hopkins University

Unclaimed · AI-structured profileOpenAlex

Cognitive-science-driven AI researcher at Johns Hopkins: LLM cognition benchmarks, personalization, continual learning

papers here

reviewed by Evolve

first-author

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—

Not yet scored: verified publication counts, venue metadata, and authorship-role totals are unavailable in this snapshot. This axis would measure overall scholarly output — number of papers, venue quality, and share of first- or corresponding-author positions over time.

Impact—

Not yet scored: citation counts and h-index-type metrics are absent from the current record. This axis would measure citation-based influence — citations to the reviewed papers, field-normalized impact, and evidence of the work being built upon by others.

Focus0/100

The focus score is 0, reflecting an almost perfectly even three-way split in the field distribution (0.333 each): computational cognitive science / AI evaluation, natural language processing for LLM personalization, and machine learning for continual learning. Across the three reviewed 2026 papers no field dominates; the breadth deliberately connects cognition benchmarking, personalization, and long-horizon memory.

Collaboration84/100

Collaboration scores 84, supported by 68 distinct coauthors in the bibliometric record and twelve frequent collaborators, including Lance Ying, Jinzhou Wu, Yingshan Susan Wang, Katherine M. Collins, Akshay K. Jagadish, and Guangyuan Jiang. Across the three reviewed 2026 papers Shu holds middle and last author positions without a corresponding role, consistent with team-based projects spanning cognitive science, NLP, and machine learning.

Activity—

Not yet scored: a full per-year publication timeline is unavailable; only three 2026 arXiv preprints are verified in this snapshot. This axis would measure publication cadence and sustained research activity across recent years.

Reproducibility85/100

Reproducibility scores 85 as the mean over all 3 reviewed papers. HyperTrace and the continual-learning composition study rate high readiness; CogGym rates medium. Evolve highlights the continual-learning paper's rigorous, honest framing — its gains apply to exact-match memorization only — while HyperTrace's moderate effects rest on LLM evaluators over small samples.

Research summary

Tianmin Shu, based at Johns Hopkins University, works at the intersection of computational cognitive science and modern AI systems. The Evolve-reviewed record comprises three 2026 arXiv papers spread evenly (0.333 each) across three areas: AI evaluation of cognition, LLM personalization, and continual learning for language models. CogGym unifies 258 cognitive experiments into a comparative benchmark, finding that LLM-human behavioral alignment grows with model size yet remains far below human reliability. HyperTrace introduces a training-free, Sequential Monte Carlo-based hypothesis-tracing method for online LLM personalization, with consistent but moderate effects judged mainly by LLM evaluators on small samples. The third paper applies a 2^4 factorial design to show that composed continual-learning mechanisms raise 100-task memorization retention from 1.2% to 34.9%, while openly limiting its claims to exact-match memorization. Bibliometrically, the profile shows strong collaboration (score 84; 68 distinct coauthors, twelve frequent collaborators) and high reproducibility (score 85, the mean across all 3 reviewed papers, two rated high readiness). The even three-way field split yields a focus score of 0, signaling breadth rather than a dominant field so far. Output, impact, and activity dimensions lack sufficient data and are left unscored. This is a full-tier profile built from public bibliographic records and Evolve's paper reviews.

Collaboration pattern

The bibliometric record lists 68 distinct coauthors and twelve frequent collaborators, giving a high collaboration score of 84. Recurring names include Lance Ying, Jinzhou Wu, Yingshan Susan Wang, Shivam Aarya, Luca M. Schulze Buschoff, Katherine M. Collins, Sean Dae Houlihan, Akshay K. Jagadish, and Guangyuan Jiang. Across the three reviewed 2026 papers, Shu appears in middle and last author positions with no corresponding-author role, consistent with team-based work spanning cognitive science, NLP, and machine learning groups.

Research keywords

LLM evaluationhuman-machine behavioral alignmentcognitive science benchmarkcommonsense reasoningscaling lawsLLM personalizationonline preference learningSequential Monte Carlotraining-free adaptationcontinual learningcatastrophic forgettinggenerative replay

Highlights

CogGym: benchmarking human vs. machine cognition at scale
Unifies 258 cognitive experiments into a single comparative evaluation suite; finds LLM-human behavioral alignment scales with model size but remains far below human reliability. Evolve calls it a rigorous, well-validated benchmark, with medium reproducibility readiness.
Training-free preference tracing for online LLM personalization
HyperTrace frames personalization as hypothesis-based tracing solved with Sequential Monte Carlo, requiring no model training. Effects are consistent but moderate, judged mainly by LLM evaluators on small samples; reproducibility readiness is high.
Composed continual-learning mechanisms multiply long-horizon retention
A 2^4 factorial study shows that combining continual-learning mechanisms (LoRA, generative replay, self-distillation) lifts 100-task memorization retention from 1.2% to 34.9%; evaluation covers exact-match memorization only, a limit the authors state openly. Reproducibility readiness is high.

Timeline

2026Co-authored CogGym, a large-scale comparative evaluation of human and machine cognition unifying 258 cognitive experiments.
2026Co-authored HyperTrace, a training-free, hypothesis-based preference-tracing method for online LLM personalization.
2026Last-authored a 2^4 factorial study showing composed continual-learning mechanisms raise 100-task retention from 1.2% to 34.9%.

Papers in the Evolve catalog

Frequent co-authors

Lance Ying×1 Jinzhou Wu×1 Yingshan Susan Wang×1 Shivam Aarya×1 Luca M. Schulze Buschoff×1 Harry Chen×1 Katherine M. Collins×1 Andrea de Varda×1 Shuhao Fu×1 Sean Dae Houlihan×1 Akshay K. Jagadish×1 Guangyuan Jiang×1

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.

Profile is name-resolved from public bibliographic records; country and ORCID are absent from this snapshot, so identity confidence rests on the institutional affiliation.
Collaboration, focus, and reproducibility scores are computed from the provided bibliometric snapshot; no citation counts or venue metadata were available.
Output, impact, and activity dimensions are unscored because citation, venue, and per-year publication data are missing from the input.
All three reviewed papers are 2026 arXiv preprints; peer-reviewed versions and earlier or non-arXiv work fall outside this profile's evidence base.