Shuang Liang

Research profile

Research summary

Shuang Liang is affiliated with Nanjing University of Science and Technology and works across two machine-learning tracks that split the reviewed record evenly: agentic uncertainty quantification for LLM agents, and multimodal graph reasoning spanning computer vision and graph learning. The Evolve catalog holds two 2026 arXiv preprints. The first, GRUET (first author), quantifies uncertainty in ReAct-style reasoning-and-acting trajectories by modeling reasoning branches as directed acyclic graphs and reports a +17 AUROC improvement over the best of ten uncertainty-quantification baselines; Evolve's review calls the method promising but heuristic and under-ablated, with medium reproducibility readiness. The second, VGCompiler (last author), trains an 8B vision-language model with reinforcement learning (GRPO) to compile graph images into executable code, raising average accuracy from 62.8 to 91.7, though part of the gain is attributed to task-specific training; its reproducibility readiness is rated low. The reproducibility mean across the two papers is 38/100, the collaboration score 48/100 over 10 distinct coauthors — including frequent partners Rongzheng Wang, Zhe Wang and Ke Qin — and the focus score 0/100 because output divides exactly 50/50 between the two subfields. Output, impact and activity axes are unscored because citation counts, publication totals and annual cadence are absent from the underlying data; the profile should be read as early-stage and name-resolved from public bibliographic records.

Research keywords

Highlights

• GRUET: +17 AUROC for ReAct trajectory uncertainty

  First-author 2026 arXiv preprint that quantifies uncertainty of reasoning-and-acting (ReAct) processes by building DAG reasoning-branch graphs, reporting a +17 AUROC improvement over the best of ten uncertainty-quantification baselines; Evolve flags it as promising but heuristic and under-ablated, with medium reproducibility readiness.

• VGCompiler: RL-trained 8B VLM compiles graph images into code

  Last-author 2026 arXiv preprint training an 8B vision-language model with reinforcement learning (GRPO) to compile graph images into executable code for visual graph reasoning, lifting average accuracy from 62.8 to 91.7; Evolve notes part of the gain comes from task-specific training, with low reproducibility readiness.

• Two-track 2026 output with first- and last-author roles

  Both cataloged papers appeared as arXiv preprints in 2026, one led as first author (GRUET, LLM-agent uncertainty) and one closed as last author (VGCompiler, visual graph reasoning), establishing two evenly split research tracks — 50/50 across the two subfields — with 10 distinct coauthors in the recorded network.

Timeline

1. 2026Posts the first-author arXiv preprint GRUET on quantifying uncertainty of agentic reasoning-and-acting (ReAct) processes, reporting +17 AUROC over the best of ten uncertainty-quantification baselines.
2. 2026Appears as last author on the arXiv preprint VGCompiler, training an 8B vision-language model with GRPO reinforcement learning to compile graph images into executable code (91.7 vs 62.8 average accuracy).
3. 2026Both reviewed works land in the same year, splitting evenly between LLM-agent uncertainty quantification and multimodal graph reasoning, with 10 distinct coauthors in the recorded network.

Papers in the Evolve catalog

• 2026-09-21first author

  GRUET: Quantifying Uncertainty of Agentic Reasoning-and-Acting Processes

• 2026-09-16

  Visual Graph Reasoning via Knowledge Compilation

Frequent co-authors