Shao-Qun Zhang

Nanjing University of Science and Technology · China

Unclaimed · AI-structured profileOpenAlex

Graph-based uncertainty quantification for LLM reasoning & ReAct agents — Nanjing Univ. of Science and Technology

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: this axis would aggregate total works and publication venues over time. The OpenAlex-linked record is too new and sparse for a meaningful count, with an h-index of 0 so far.

Impact—

Not yet scored: this axis would track citation counts, h-index and i10-index across the record. Both indexed indicators currently stand at 0, reflecting an early-stage profile rather than a scored result.

Focus0/100

Scores 0/100 because the raw field distribution splits evenly — 'machine learning / NLP — uncertainty quantification for LLM reasoning' at 0.5 and 'machine learning (LLM agents, uncertainty quantification)' at 0.5. In substance the two labels describe one tightly concentrated agenda: graph-based uncertainty quantification for LLM reasoning and ReAct agents. Topic counts corroborate this — Topic Modeling (2) and Adversarial Robustness in Machine Learning (2) lead, with NLP techniques and AI planning at 1 each.

Collaboration32/100

Scores 32/100 on a thin record: 4 distinct coauthors appear across the OpenAlex-linked works, with Xin-Yu Hu, Shuang Liang and Xiang-Jun Ou as the recurring collaborators. Both cataloged papers list Shao-Qun Zhang in the last-author position without corresponding-author status, consistent with a junior member of a small, stable team rather than an independent or internationally networked group; the axis should be revisited as the coauthor set grows.

Activity32/100

Scores 32/100, consistent with a very recent start. OpenAlex attributes 2 works to this profile, both in 2026, and recent3yWorks is likewise 2; no earlier years appear in countsByYear. The two cataloged arXiv preprints in a single year indicate concentrated output since entering the record, but a one-year window cannot establish a sustained cadence, so the activity score should be read as 'new entrant' rather than 'low performer'.

Reproducibility55/100

Scores 55/100, the mean over 2 papers assessed by Evolve, both rated 'medium' reproducibility readiness. Evolve's verdicts flag concrete gaps: GRUET is called promising but heuristic and under-ablated, while GUT shows strong gains across 45 baselines yet is Qwen3-only with possible test-set tuning. The consistent graph-complexity methodology across both works aids replication, but single-model dependence and thin ablations keep readiness at the midpoint rather than higher.

Research summary

Shao-Qun Zhang is an early-career researcher at Nanjing University of Science and Technology in China; this profile is name-resolved from OpenAlex public bibliographic data. The record centers on one clearly defined problem: quantifying the uncertainty of large language model reasoning, for both single-chain reasoning and agentic reasoning-and-acting (ReAct) processes. Two 2026 arXiv preprints in the Evolve catalog define the agenda. GUT introduces graph-complexity uncertainty quantification, merging reasoning-branch DAGs with NLI-based semantic merging and optimizing the measure through reinforcement learning (GRPO), with strong reported gains across 45 baselines. GRUET extends the same graph-based view to agentic trajectories, reporting +17 AUROC over the best of ten uncertainty-quantification baselines. Evolve rates both papers 'medium' reproducibility readiness, flagging GRUET as promising but heuristic and under-ablated, and GUT as Qwen3-only with possible test-set tuning. Bibliometrically the profile is very new: OpenAlex attributes 2 works, both in 2026, with 4 distinct coauthors — Xin-Yu Hu, Shuang Liang and Xiang-Jun Ou recurring most often — and h-index and i10-index currently at 0. The field distribution splits evenly between LLM-agent uncertainty quantification and NLP-side uncertainty quantification, but in substance this is a single concentrated research line. Zhang appears as last author on both papers without corresponding-author status, consistent with a junior position in a small team. The output and impact axes are not yet scored.

Collaboration pattern

The coauthor footprint is small and stable: 4 distinct coauthors in the OpenAlex record, with Xin-Yu Hu, Shuang Liang and Xiang-Jun Ou recurring across works. On both cataloged papers Shao-Qun Zhang sits in the last-author slot without corresponding-author status, suggesting a junior collaborator within one compact team rather than a hub of an external network; the collaboration score of 32 reflects record thinness as much as working style.

Research keywords

uncertainty quantificationLLM reasoningLLM agentsReActgraph complexitydirected acyclic graphsemantic entropyselective generationreinforcement learningGRPO

Highlights

GRUET: DAG-based uncertainty estimates for agentic ReAct trajectories
Builds DAG graphs over reasoning branches in ReAct processes and reports +17 AUROC against the best of ten uncertainty-quantification baselines; Evolve calls the result promising but heuristic and under-ablated, with medium reproducibility readiness.
GUT: graph-complexity uncertainty with reinforcement-learning optimization
Merges reasoning-branch DAGs with NLI-based semantic merging and optimizes the uncertainty measure via GRPO, reporting strong gains across 45 baselines; Evolve flags Qwen3-only models and possible test-set tuning for scrutiny.
One graph-complexity toolkit spanning chain reasoning and agent trajectories
Both cataloged papers instantiate the same idea — represent reasoning as a DAG of branches, then read uncertainty off graph complexity — moving from single-chain LLM reasoning (GUT) to full ReAct reasoning-and-acting processes (GRUET).

Timeline

2026GUT preprint posted to arXiv, introducing graph-complexity uncertainty quantification for LLM reasoning with reinforcement-learning (GRPO) optimization.
2026GRUET preprint posted to arXiv, extending the graph-based uncertainty approach to agentic ReAct reasoning-and-acting trajectories.
2026Evolve reviewed both papers, rating each 'medium' reproducibility readiness; OpenAlex records 2 works for this profile, all in 2026.

Papers in the Evolve catalog

Frequent co-authors

Xin-Yu Hu×2 Shuang Liang×1 Xiang-Jun Ou×1 Shuang Liang×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.

Bibliometrics are derived from OpenAlex only; no ORCID is linked, so this profile is name-resolved from public bibliographic data and may merge or split records for a common Chinese name.
Both cataloged papers are 2026 arXiv preprints; no peer-reviewed venue record is available yet.
Output and impact dimensions are unscored because indexed citation aggregates (h-index and i10-index currently 0) are not yet meaningful for this new record.
All scores rest on a thin base of 2 reviewed papers and 4 distinct coauthors; treat the collaboration, activity and reproducibility axes as provisional.