Xin-Yu Hu

Nanjing University of Science and Technology · China

Unclaimed · AI-structured profileOpenAlex

Early-career NJUST researcher quantifying uncertainty in LLM reasoning and ReAct agents via graph complexity

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 scored yet. The snapshot shows only 3 works recorded in 2026 with no longer publication history, so the output axis — which measures total publication volume and multi-year productivity trend — cannot be computed reliably; any estimate would be premature.

Impact—

Not scored yet. Citation-based indicators currently stand at h-index 0 and i10-index 0, consistent with an early-stage record, so the impact axis — which would measure citation uptake and influence on follow-up work — lacks sufficient data to score.

Focus0/100

Scores 0/100 for concentration. The OpenAlex field distribution splits evenly (0.5/0.5) between two overlapping labels — uncertainty quantification for LLM reasoning, and LLM agents with uncertainty quantification — and topic assignments spread across at least four clusters: Topic Modeling (3), Adversarial Robustness (3), NLP Techniques (2), and AI-based Problem Solving and Planning (1). Despite this dispersed topical footprint, both reviewed papers share one concrete theme: graph-based uncertainty quantification for LLM reasoning.

Collaboration32/100

Scores 32/100, indicating a small, concentrated network. OpenAlex records 4 distinct co-authors, with Shao-Qun Zhang, Shuang Liang, and Xiang-Jun Ou recurring most often. In both catalog papers Hu holds a middle author position without corresponding-author status, consistent with contributing inside a compact, stable team rather than leading large multi-group collaborations. No international or cross-institution co-authorship patterns are visible in the provided data.

Activity40/100

Scores 40/100. The counts-by-year record shows 3 works in 2026 and zero in all earlier years covered by the snapshot, with recent3yWorks also equal to 3 — meaning all visible activity is concentrated in a single recent year. Both catalog papers are 2026 arXiv preprints. This matches a researcher whose documented output has just begun; the moderate score reflects thin multi-year history rather than a slowdown.

Reproducibility55/100

Scores 55/100, the mean across 2 reviewed papers, both rated medium reproducibility readiness. Evolve's review flags concrete concerns: GRUET is described as promising but heuristic and under-ablated, while GUT delivers strong gains in a 45-baseline comparison yet evaluates only Qwen3 models, with possible test-set tuning that needs scrutiny. The mid-range score reflects real methodological detail alongside unverified generality across models and settings.

Research summary

Xin-Yu Hu is a researcher at Nanjing University of Science and Technology (China) working on uncertainty quantification for large language model reasoning and agentic systems. The Evolve catalog includes two 2026 arXiv preprints, both middle-authored, that share one technical signature: they cast LLM reasoning as a directed acyclic graph and quantify uncertainty from its structure. GUT merges DAG structure with NLI-based semantic comparison into a graph-complexity uncertainty signal and couples it with GRPO reinforcement-learning optimization, reporting strong gains across a 45-baseline comparison; Evolve's review flags Qwen3-only evaluation and possible test-set tuning. GRUET applies the same idea to ReAct-style agent trajectories, reporting +17 AUROC over the best of ten UQ baselines, with the caveat that the method is heuristic and under-ablated. OpenAlex (name-resolved, no ORCID linked) records 3 works in 2026, h-index and i10-index of 0, and 4 distinct co-authors — Shao-Qun Zhang, Shuang Liang, and Xiang-Jun Ou recur — an early-stage, recently started documented record. Topic data spreads across Topic Modeling (3), Adversarial Robustness (3), NLP Techniques (2), and AI-based Problem Solving and Planning (1), yielding a focus score of 0, while reproducibility averages 55 across the two reviewed papers.

Collaboration pattern

The collaboration footprint is small and concentrated: OpenAlex records 4 distinct co-authors, with Shao-Qun Zhang, Shuang Liang, and Xiang-Jun Ou appearing repeatedly. On both reviewed 2026 papers Hu is a middle-position author without corresponding status, suggesting work inside a compact, stable team at Nanjing University of Science and Technology rather than distributed or international collaborations. No senior-collaborator or cross-institution patterns are visible in the provided data.

Research keywords

uncertainty quantificationLLM reasoningLLM agentsReActgraph complexitydirected acyclic graphsemantic entropyselective generationreinforcement learningGRPO

Highlights

GRUET: uncertainty quantification for ReAct-style agent trajectories
Represents each agentic reasoning-and-acting (ReAct) process as a DAG of reasoning branches and quantifies uncertainty over that structure to support selective generation; Evolve's review reports a +17 AUROC gain over the best of ten UQ baselines, while flagging the method as heuristic and under-ablated.
GUT: graph-complexity uncertainty measures coupled with RL optimization
Merges DAG structure with NLI-based semantic comparison into a graph-complexity uncertainty signal for LLM reasoning, then optimizes reasoning with GRPO reinforcement learning; Evolve notes strong gains across a 45-baseline comparison, with caveats that evaluation covers only Qwen3 models and possible test-set tuning needs scrutiny.
One methodological signature carried from reasoning chains to agent trajectories
Both 2026 catalog papers apply the same core idea — representing LLM reasoning as a directed acyclic graph and reading uncertainty off its structure — extending from single-answer reasoning chains in GUT to complete ReAct-style agentic processes in GRUET.

Timeline

2026First visible catalog entry: the GUT preprint appears on arXiv, introducing graph-complexity uncertainty quantification for LLM reasoning that merges DAG structure with NLI-based comparison and optimizes reasoning with GRPO.
2026The GRUET preprint appears on arXiv the same year, extending uncertainty quantification from single reasoning chains to full ReAct-style agentic processes via DAG reasoning-branch graphs.
2026OpenAlex records 3 works in 2026 (recent3yWorks = 3), all concentrated in this single year, with h-index and i10-index at 0 — a newly started, early-stage documented record.

Papers in the Evolve catalog

Frequent co-authors

Shao-Qun Zhang×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.

Bibliometric scores, topics, and co-author counts are derived from OpenAlex and are name-resolved; no ORCID is linked, so identity disambiguation relies on public bibliographic data only.
Both reviewed papers are 2026 arXiv preprints; h-index and i10-index of 0 reflect an early-stage record, not a negative quality signal.
Output and impact dimensions are unscored: the snapshot contains only 3 works recorded in 2026, too little history for reliable productivity or citation metrics.
The researcher's Chinese name characters and full affiliation history are unconfirmed; only the romanized name is available.