Xiao Liang

University of California, Los Angeles

Unclaimed · AI-structured profileOpenAlex

UCLA NLP researcher at the intersection of LLM reasoning efficiency and long-term memory evaluation for LLM agents

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—

No output metrics are available in this snapshot, so this axis is unscored. Output would measure total publications per year and venue spread; the Evolve catalog currently holds only two 2026 arXiv preprints for this researcher.

Impact—

No citation-based signal exists yet. Impact would measure citation counts, h-index, and influence on follow-up work; both catalog papers are 2026 arXiv preprints, too recent for citation data to accumulate.

Focus0/100

The focus score of 0 reflects an exactly even field split in the raw data: half of the cataloged output in natural language processing (LLM reasoning) and half in natural language processing (LLM agents / long-term memory evaluation). The two strands share an evaluation-and-efficiency theme — steering reasoning toward progress and attributing memory failures — but the 50/50 distribution means no single subfield dominates yet.

Collaboration66/100

The collaboration score of 66 rests on 26 distinct coauthors in the underlying records, with a recurring core of twelve frequent collaborators including Hengyuan Zhang, Chenming Shang, Jing Xiong, Dawei Li, Ruobing Xie, Hayden Kwok-Hay So, and Ngai Wong. On both cataloged 2026 papers Xiao Liang sits in a middle author position and is not the corresponding author, suggesting contributions within larger team-driven projects rather than solo-led studies.

Activity—

Activity tracks publishing cadence across years, but the snapshot contains only two records, both from 2026. Without a multi-year publication history, this axis cannot be scored honestly and remains marked as missing.

Reproducibility55/100

The reproducibility score of 55 is the mean over the two cataloged papers, both rated medium reproducibility readiness. For SPS, training-free latent steering favors reproducibility in principle, but Evolve's verdict flags thin statistical rigor despite consistent Pass@4 gains on Qwen3 1.7B-14B across 9 benchmarks. EvalMem is described as well validated with convincing retrieval-defect findings, though its reported gains are limited.

Research summary

Xiao Liang is a researcher at the University of California, Los Angeles whose recent work sits at the intersection of two fast-moving corners of natural language processing: reasoning in large language models and the evaluation of long-term memory in LLM agents. Evolve's catalog includes two 2026 arXiv papers with Xiao Liang in a middle author position, neither corresponding. The first, 'Efficient Reasoning Exploration via State-Conditioned Latent Steering with Progress Guidance', proposes SPS, a training-free latent steering method that guides reasoning toward measurable progress; Evolve's review reports consistent Pass@4 gains on Qwen3 models from 1.7B to 14B parameters across nine benchmarks while flagging thin statistical rigor. The second, 'EvalMem', is an operation-level diagnostic framework that uses three Examiners to attribute long-term-memory failures in agents to encoding, retrieval, or generation; Evolve judges it well validated, with convincing findings on retrieval defects, though measured gains are limited. Bibliometrically, the cataloged output splits evenly (0.5/0.5) between the two subfields, 26 distinct coauthors appear in the underlying records with a collaboration score of 66, and reproducibility averages 55 across two medium-readiness papers. Because 'Xiao Liang' is a common name, this profile is name-resolved from public bibliographic data; output, impact, and activity metrics are not yet available, so conclusions here rest on a small, early-stage evidence base.

Collaboration pattern

Xiao Liang works within sizeable teams: the underlying records list 26 distinct coauthors, and a stable core of twelve frequent collaborators recurs, including Hengyuan Zhang, Chenming Shang, Jing Xiong, Dawei Li, Kailai Yang, Ruobing Xie, Hayden Kwok-Hay So, and Ngai Wong. On both cataloged papers the position is middle author with no corresponding role, consistent with contributing specialist work inside a lab-scale collaboration rather than leading authorship.

Research keywords

LLM reasoninglatent steeringreasoning efficiencyPass@4 evaluationtraining-free methodslong-term memoryLLM agentsmemory failure attributionevaluation benchmarkretrieval-augmented generationagentic RAGdiagnostic frameworks

Highlights

Training-free latent steering makes LLM reasoning more efficient
The SPS method conditions latent steering on the reasoning state with progress guidance, delivering consistent Pass@4 gains on Qwen3 models from 1.7B to 14B across 9 benchmarks; Evolve notes the statistical rigor is thin.
EvalMem traces agent memory failures to specific operations
EvalMem uses three Examiners to attribute long-term-memory failures to encoding, retrieval, or generation; Evolve finds the validation thorough and the retrieval-defect findings convincing, though the gains are limited.
A dual-track profile linking reasoning efficiency and memory diagnostics
The two 2026 papers split evenly across LLM reasoning and LLM agent memory evaluation (field distribution 0.5/0.5), suggesting a research agenda that treats efficient reasoning and reliable agent memory as complementary evaluation problems.

Timeline

2026First cataloged arXiv paper on reasoning efficiency: SPS, a training-free state-conditioned latent steering method with progress guidance, reporting consistent Pass@4 gains on Qwen3 1.7B-14B across 9 benchmarks.
2026Second cataloged arXiv paper: EvalMem, an operation-level diagnostic framework attributing long-term-memory failures in LLM agents to encoding, retrieval, or generation.
2026Both cataloged papers appear in the same year, making this an early-stage profile whose recorded output to date concentrates entirely in 2026.

Papers in the Evolve catalog

Frequent co-authors

Hengyuan Zhang×1 Chenming Shang×1 Zunhai Su×1 Hui Shen×1 Jing Xiong×1 Dawei Li×1 Shiping Yang×1 Kailai Yang×1 Wei Zhang×1 Ruobing Xie×1 Hayden Kwok-Hay So×1 Ngai Wong×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.

Dimension scores are computed from public bibliometric data (OpenAlex-sourced); output, impact, and activity axes lack sufficient data and are unscored.
'Xiao Liang' is a common name; this profile is name-resolved from public bibliographic records and identity should be verified before contact.
Coverage is limited to two 2026 arXiv preprints in the Evolve catalog; peer-reviewed venue acceptance is not reflected.
The reproducibility score covers only the two reviewed papers, on neither of which this researcher is the corresponding author in the catalog.