WorkExperience

Research Intern

Pisa, Toscana, Italy · On-site

Independent research: memory for multimodal agents (first-author paper under review; arXiv · project page)

• Controlled experiments located the bottleneck at the delivery step. Holding the retrieved set fixed and changing one variable at a time, delivering the original images is worth +13.87 points and perfect retrieval only +2.31 on Qwen3-VL-8B; the delivery gain grows on a stronger backbone (+17.92 on Qwen3-VL-32B).
• Proposed DeliverMem, a training-free delivery method that leaves the memory store unchanged (original modality, readable identifiers and temporal position, plus a retrieval-side adaptation): state of the art on DMV-Bench, ahead of the best prior method at every setting on both backbones by 7.63 to 22.62 points; on MemLens it leads the strongest published memory-augmented agent by at least 8.20 points at all four context lengths, on a tenth to a seventieth of the full-context input. At MemLens 32K with dev-fitted selection it surpasses GPT-5.4 and matches Gemini-3.1-Pro while delivering about 2,400 tokens, a tenth of what they read.

Business project: LLM post-training for in-cabin AI (sole owner)

• SFT: LoRA fine-tuning on Qwen3-4B / 8B and Qwen3.6-35B-A3B (MoE) with ms-swift; implemented a set-prediction loss (minimum-permutation CE), hard-case weighted SFT and rejection-sampling fine-tuning (RFT).
• Preference optimization and RL, with reward design: DPO and triplet-level DPO (preference pairs split by Hungarian alignment), GRPO / DAPO; designed a hierarchical multi-component GRPO reward (format validity, schema, hierarchical triplet F1, direction, relation class and process rewards).
• Evaluation: built a triplet evaluator (Hungarian matching, placeholder equivalence, relation-synonym normalization, per-class and per-step reporting) with batched vLLM inference; audited and repaired the business memory evaluation set, replacing LLM-based pairing with a deterministic memory-level metric.

Research Intern

Trento, Trentino-Alto Adige, Italy · On-site
Supervisors: Yiming Wang, Elisa Ricci

• Proposed Active Instance Verification and built its first benchmark, PInVerify. Navigation benchmarks count arrival as success, yet an agent can stop in front of the wrong instance of the right category. Formalized this stage as a finite-horizon decision process and built 3,000 evaluation episodes over 18 object categories in photorealistic 3D scenes, with trap views and unreachable sectors in the navigation topology.
• Designed the evaluation protocol and failure diagnostics. Positive, same-category and cross-category pairs are balanced 1:1:1 to separate calibration bias from category recognition; beyond accuracy, the protocol reports steps to decision and navigation failure rate, and the training pool is disjoint from the test set. Judging differences by 95% confidence intervals located the main bottleneck in detection (+3.1 pp with ground-truth boxes) and found no reliable difference among three next-best-view strategies.
• Built two baseline tracks, training-free and trained. The training-free track is a modular pipeline (attribute decomposition, multi-view tracking and fusion, three next-best-view strategies), evaluated on Qwen3-VL-4B / 8B, SenseNova-SI-InternVL3-8B, CLIP and SigLIP2; the trained track is an end-to-end Qwen3-VL-4B + LoRA agent that outputs the verification decision and the next action in one forward pass.
• Built a CoT dataset and post-trained the agent. Generated about 22K CoT samples from benchmark ground truth and per-step visibility metadata, then compared DPO, GRPO and GSPO on top of SFT with a reward combining verification correctness, action quality and format. Accuracy rose from 70.6% untuned to 84.8% with SFT and 85.6% with SFT+GSPO (88.9% with ground-truth boxes), and adding either GRPO or GSPO on top of SFT cut navigation failures from 18.2% to about 9%.

Algorithm Intern

Chengdu, Sichuan, China · On-site

• Researched the newest machine learning algorithms and industrial anomaly detection algorithms.
• Applied GAN models to reconstruct images to identify anomalies and evaluate the performance.
• Used object detection neural networks to detect abnormal parts in high-speed rail supports.
• Assisted colleagues in developing a set of character recognition patents for LCD meters.