Zitong Qi

Yongjiang Laboratory / Imou

Jan 2026 to Mar 2026 · SafeCommunity AI · R&D Engineer · Intelligent Video Surveillance

Built a weakly supervised anomaly-detection and VLM-verification pipeline; QLoRA tuning of Qwen3.5-VL improved ECVA test accuracy from 65.68% to 78.53%.

TReNDS Center

Mar 2026 — May 2026 · Georgia State University · Georgia Institute of Technology · Emory University, USA · Remote Research Assistant

Conducted ACVF research on multimodal brain MRI, including method design, implementation, cross-backbone experiments and modality ablations.

Tencent Kaiwu · Reinforcement Learning Competition

2026.04 — 2026.05 · National Third Prize

Designed a 17-token Transformer policy and PPO training with three value estimates, GAE, BFS escape features and rewards for survival, collection and exploration.

Eastern Institute of Technology, Ningbo (EIT)

Jun 2026 to Oct 2026 · EAST-Lab · Prof. Wei Zhang’s group · Research Assistant

Worked on ReInsVLN: led early exploration and integration of visual compression, contributed to its implementation, collaboratively built high-level / low-level coordination, deployed models on Go2/G1, and participated in robot experiments.

ReInsVLN · Training-free vision-language navigation

ICRA 2027 · Submitted · Second author · Core contributor

My contribution: Led early exploration and integration of visual compression, contributed to its implementation, and collaboratively refined memory selection. Collaboratively built the high-level / low-level coordination framework, adapted multiple VLN models, and connected model inference to robot execution. Deployed models on Go2 and G1 and contributed to robot experiments.

1. Compress observation history before the language model, reducing temporal redundancy while protecting poorly covered content.
2. Select visual tokens inside the language model using instruction relevance and position, then restore their original order.
3. Add scheduled history-review cues and connect native navigation decisions to robot execution.

Project results: On all 1,839 R2R Val-Unseen episodes, StreamVLN / InternVLA-N1 reach 64.27% / 64.44% SR, gains of 8.48 / 3.37 percentage points. On matched observation workloads, the InternVLA-N1 compression configuration reduces prefill latency by 21.77% and KV storage by 32.25%.

ACVF · Multimodal fusion of brain MRI and clinical language

ICASSP 2027 · Submitted · First author

My contribution: Independently designed and implemented fusion of MRI, AAL regional volumes and BioClinicalBERT text, using subject-level SigLIP alignment with frozen backbones and cross-backbone ablations.

1. Encode MRI, clinical text and AAL regional volumes while keeping the MRI and BioClinicalBERT backbones frozen.
2. Project features into a shared 256-dimensional space and align the three modalities with a subject-level SigLIP objective.
3. Exclude task-specific label-related fields and evaluate across backbones and ablations on 1,077 ADNI subjects.

Project results: For AD classification excluding cognitive-score inputs, average AUC improves by 0.144 across three backbones. BrainIAC accuracy rises from 64.6% to 84.3% and AUC from 0.732 to 0.922. AnatCL brain-age MAE decreases from 4.924 to 4.373 years.

PairSelect · Token compression for weakly supervised video anomaly detection

ICASSP 2027 · Submitted · First author

My contribution: Independently developed local pairing, norm ranking and group quotas after Transformer block 6. Adapted four detector systems without retraining for compression or budget changes.

1. Contextualize all tokens through the first six dense Transformer blocks.
2. Apply local pairing, norm ranking and group quotas, preserving original vectors and temporal coverage.
3. Run the remaining blocks with fewer tokens and reuse detector weights, without retraining for compression or budget changes.

Project results: On UCF-Crime / XD-Violence at batch size 32, 40% token deletion yields 1.22–1.28× encoder speedup. VideoMAEv2 UCF AUC rises from 80.63% to 81.02%; TimeSformer XD PR-AUC rises from 76.69% to 77.37%. At 60% token deletion, the encoder reaches up to 1.47× encoder speedup.

ProSeCIL

Prototype-Calibrated Semantic Fusion for Exemplar-Free Class-Incremental Learning

Submitted to ICASSP 2027 · Fourth author

TRICE

Tri-Branch Calibrated Ensemble for Optimization-Free Incremental Adaptation in Few-Shot Class-Incremental Learning

Submitted to ICASSP 2027 · Third author