LLM evaluation researcher; co-authored 2026 arXiv benchmarks on agentic competence and compute-matched MoE scaling.
3
papers here
1
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-17) 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. Aggregate publication volume (total and per-year work counts) is unavailable in the current snapshot, which lists only four 2026 works; a fuller, disambiguated corpus is needed before this axis can be rated.
Impact—
Not yet scored. Citation indicators in the snapshot are effectively empty (h-index and i10-index both 0), so research influence cannot be meaningfully measured; this axis would track citations, citation velocity and downstream uptake.
Focus100/100
Scores 100/100 on focus. The topic profile is tightly concentrated: among topic assignments, Topic Modeling accounts for 4, Natural Language Processing Techniques 2, AI-based Problem Solving 2, AI Applications 2, and Advanced Neural Network Applications 1 — all clustering around large-language-model evaluation and efficient architectures, consistent with the three reviewed 2026 papers.
Collaboration79/100
Scores 79/100. The record lists 52 distinct co-authors against only 4 works in the recent three-year window — a very high collaborator-per-paper ratio — with recurring partners including Wenhao Huang, Shaowen Wang, Ge Zhang, Yuhao Wu and Shen Yan, indicating team-based, consortium-style LLM research.
Activity47/100
Scores 47/100. Visible activity is concentrated in a single year: the snapshot records 4 works in 2026 and none in earlier years, with 4 works across the recent three-year window. This pattern likely reflects a newly minted, name-resolved record rather than inactivity, so the score should be read cautiously.
Reproducibility55/100
Scores 55/100, computed from 1 reproducibility-assessed paper (mean readiness 55/100). Across the three reviewed works, both benchmark studies carry medium reproducibility readiness, while the compute-matched MoE scaling study is rated low: its evidence is single-run and proprietary. Widening assessment coverage beyond one paper would firm this score up.
Research summary
Jiaheng Liu is an early-career researcher working on large language model evaluation and efficient LLM architectures; this profile is name-resolved from public OpenAlex bibliographic data and Evolve's reviews of three 2026 arXiv papers. All three catalog entries place Liu as a middle author without corresponding-author status, consistent with a contributor role on team-produced work. The portfolio spans three complementary threads: xDailyBench, a 248-task benchmark of real-user professional consultation requests finding that implicit needs trail explicit ones by at least 9 points across 11 models; HarnessDev, which tests whether LLMs can build and evolve their own agent harnesses and finds LLM-built systems runnable but behind human systems on code and search; and SMELT, a compute-matched scaling study of MoE looped transformers claiming 6.8–18% compute savings under matched FLOPs, parameters and KV-cache. OpenAlex topics reinforce this concentration — Topic Modeling (4 assignments), Natural Language Processing Techniques, AI-based Problem Solving and AI Applications — producing a maximal focus score of 100. The record shows 52 distinct co-authors across only 4 recent works (collaboration score 79) and visible activity concentrated in 2026. Citation indicators currently read zero, so research impact cannot yet be assessed.
Collaboration pattern
The record lists 52 distinct co-authors across just 4 recent works, an unusually wide collaborative footprint (collaboration score 79). Frequent partners include Wenhao Huang, Shaowen Wang, Ge Zhang, Yuhao Wu, Shen Yan, Qingshui Gu and Jian Li, suggesting stable team clusters behind the three 2026 arXiv papers. All reviewed papers place Liu in a middle-author, non-corresponding role, consistent with specialized contributions within larger LLM research groups.
Research keywords
Highlights
248-task benchmark shows LLMs miss implicit needs in real consultations
LLMs can build runnable agent harnesses, but still trail human systems
Compute-matched scaling laws favor MoE looped transformers
Timeline
- 2026Co-authored xDailyBench, benchmarking 11 LLMs on 248 real-user professional consultation tasks as a middle author.
- 2026Co-authored HarnessDev, testing whether LLMs can create and evolve their own agent harnesses.
- 2026Co-authored SMELT, deriving compute-matched scaling laws for MoE looped transformers with claimed 6.8–18% compute savings.
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 indicators (topics, co-author and activity counts) come from an OpenAlex name-resolved record; 'Jiaheng Liu' is a common name, so the profile may aggregate or split distinct researchers.
- Institution, country and ORCID are absent from the source data, so no affiliation context is included.
- Citation-based measures appear empty in the snapshot (h-index and i10-index both 0), so impact cannot be assessed.
- Only the three 2026 arXiv papers reviewed by Evolve are covered; earlier work and non-arXiv venues fall outside this profile's scope.