Southeast University  ·  UC San Diego

Shangqing Shi

I work on keeping language and vision-language models reliable when prompts, instructions and label spaces change.

I am a senior in Information Engineering at Southeast University and an exchange student at UC San Diego, where I work at the San Diego Supercomputer Center with Dr. James Meng.

My recent work covers on-policy distillation that is robust to prompt variation (VD-OPD), label-efficient training of models that follow edited instructions (SIGMA), classifiers whose label sets evolve after deployment (SBLS), and test-time correction of multimodal reasoning (QCV). Before that I worked on reinforcement learning and optimization for wireless systems.

Portrait of Shangqing Shi

News

  1. SBLS and QCV were accepted to AAAI 2027.

  2. Released code, data and tests for all of my papers; the index is at OdurShi/research-code.

  3. Started an exchange at UC San Diego and joined the San Diego Supercomputer Center.

  4. MRIME-CD was published in Cluster Computing.

  5. EDECO was published in Biomimetics.

  6. Our access-point deployment paper was presented at the ISAC workshop of ACM MobiCom '24.

Publications

* Equal contribution. Shaded entries are first-author or co-first-author work. Code for every paper is linked from OdurShi/research-code.

Accepted and published

Enlarge figure: Timely edit recall plotted against safety: SBLS lies above and to the right of four baselines and needs 3.5 times fewer observations than fixed-horizon selection.

Open-World Classification as Bayesian Model Selection over Evolving Label Spaces

S. Shi, M. Zhang, Y. Hu, H. Zhang, S. Hu, Z. Huang, Z. Lin, X. Zhang

AAAI 2027First author

SBLS decides when a deployed text classifier's label set should change (stay, add or split) by sequential Bayesian model selection, testing each proposed edit only on observations that arrive later. Without target-task labels it reaches 79.0 Macro-F1 on the 12-dataset BOLT suite and 80.5 Edit-F1 at a 4.0% false-edit probability on streaming benchmarks.

Manuscripts under review

Training Vision-Language Models with Margin-Based Resolution Selection (H. Yu, S. Shi, et al.) is a further manuscript under review.

Projects

Research and engineering projects beyond the papers above, each with code, tests and reproducible results.

Enlarge figure: Map of the ray-traced scene in one time slot: the airspace that is free at each flight level, 24 deployed access points that transmit, receive or are off, 15 users, four UAVs with their paths and the lines of sight used for sensing.

Multi-Agent Placement and Operation for Cell-Free ISAC on a Ray-Traced Scene

Multi-agent optimisation · ISAC · Sionna RT2025 – 2026

Takes the deployment paper from a 2-D grid to a ray-traced urban scene: a Sionna RT simulator supplies path gains for 847 candidate positions and lines of sight to aerial targets. Every access point is an agent that first picks one of 345 sites and then switches between transmitting, receiving and sleeping while UAVs fly between the buildings. I prove that both problems are exact potential games. Placement equilibria are within 0.04% of the exact optimum on small instances; in operation, per-AP best response matches tuned central annealing and overtakes it once central decisions are one slot (0.5 s) old. Its advantage is latency, not solution quality.

Enlarge figure: Architecture of the agent: question, planner, tool executor, connectors and databases, with a provenance store and a citation verifier that lead to the verified answer.

An Evidence-Checked Agent for Chemistry and Biology Databases

LLM agents · tool use2026

A tool-calling agent for PubChem, ChEMBL, UniProt, PubMed and LOTUS that archives every raw response and checks each identifier in its answer against that archive. On a 34-question benchmark against the live databases, 232 of 234 cited identifiers traced back to archived responses and every injected fake identifier was caught.

Enlarge figure: Pearson correlation of four simple baselines for four types of held-out perturbations, with the split-half replicate level as a reference.

A Baseline Gate for Virtual-Cell Perturbation Models

AI virtual cell · evaluation2026

From a model-selection study for an agent-driven virtual cell: a test that a perturbation-response model has to pass before it is trusted. On the Norman et al. 2019 Perturb-seq data, the additive baseline reaches r = 0.86 on double perturbations whose genes were both seen, while for unseen genes no baseline beats the training mean (r = 0.51).

Enlarge figure: 4D Gaussian reasoning segmentation: the task compared with earlier settings, the 4DReSplat pipeline, and a radar chart of results on Reason4D.

4DReSplat: 4D Gaussian Reasoning Segmentation

3D vision · reasoning segmentation2026

Segments what an implicit instruction refers to in a dynamic 4D Gaussian scene, in space and in time: the instruction is parsed into structured visual descriptions, a valid time window is estimated, and Gaussian primitives are selected by spatial, depth and appearance constraints. Comes with Reason4D, 415 reasoning queries over 12 scenes.

Background

  1. University of California, San Diego

    Exchange student · research at the San Diego Supercomputer Center

  2. Southeast University, Nanjing

    B.Eng. in Information Engineering (expected) · GPA 3.88 / 4.0

  3. China Mobile Group Jiangsu, Yangzhou Branch

    Software Development Engineer Intern · data-service APIs and backend modules for the internal data platform

  4. SUPCON Technology, Hangzhou

    Communications Engineer Intern · integration work for industrial communication systems