Hello, I'm

Xianda Du

Computer Engineering & AI Researcher

University of Waterloo · 4.0/4.0 GPA
Researching Search Agents, VLM Evaluation & Generative Models
Autonomy Intern @ Waabi

Get in Touch View Research
Xianda Du

About Me

I am an undergraduate student in Computer Engineering with an Option in Artificial Intelligence at the University of Waterloo (CGPA 90.87/100). My research spans Information Retrieval & Search Agents, Vision-Language Model Evaluation, and Generative Models. I am currently an Autonomy and Algorithm Engineer Intern at Waabi, working on novel view synthesis quality evaluation for self-driving simulation.

I have had the privilege of working with Prof. Wenhu Chen and Max Ku on image evaluation, Prof. Amir-Hossein Karimi and Ahmed Abdelaal on generative priors for inverse problems, Prof. Jian-Yun Nie and Yuxing Tian (Université de Montréal) on search agents, Prof. Jimmy Lin on LLM reranking,Prof. Renée Sieber and Dr. Fengran Mo (McGill University) on retrieval benchmarks, and Prof. En-Hui Yang on adversarial robustness.

I am actively seeking graduate and research positions in retrieval, multimodal AI, and generative modeling.

4.0 GPA (out of 4.0)
6 Papers
4 Co-first Author
Waabi Current Intern

University of Waterloo

Bachelor of Applied Science in Computer Engineering, Option in Artificial Intelligence · CGPA 90.87/100

Sep 2022 – May 2027
Coursework
Machine Learning, Deep Learning, Reinforcement Learning, Computational Intelligence, Co-operative and Adaptive Algorithms, Algorithms and Data Structures, Systems Programming
Awards
President’s Scholarship of Distinction (2023); Term Dean’s Honour List (Fall 2022, Winter 2023, Fall 2023); Term Distinction (Winter 2025, Spring 2026)

Publications

* denotes equal contribution.

Published

SIGIR 2026

WeatherArchive-Bench: Benchmarking Retrieval-Augmented Reasoning for Historical Weather Archives

Yongan Yu*, Xianda Du*, Qingchen Hu*, Jiahao Liang*, Jingwei Ni, Dan Qiang, Kaiyu Huang, Grant McKenzie, Renée Sieber, and Fengran Mo

A retrieval-augmented reasoning benchmark over 1M+ archival news segments and 335 expert-curated queries, pairing passage retrieval with IPCC-aligned climate vulnerability assessment.

NeurIPS 2026 Workshop BeNTo

Mean Flows as Controllable Priors for Solving Inverse Problems

Ahmed Abdelaal, Xianda Du, and Amir-Hossein Karimi

Uses pretrained MeanFlow finite-interval transport maps as controllable priors for training-free inverse problems, and introduces Receding-Horizon State Optimization (RHSO).

ACL 2025 Findings

WXImpactBench: A Disruptive Weather Impact Understanding Benchmark for Evaluating Large Language Models

Yongan Yu*, Qingchen Hu*, Xianda Du*, Jiayin Wang, Fengran Mo, and Renée Sieber

The first benchmark for LLM understanding of disruptive weather impacts: 1,736 samples across 6 impact categories, with multi-label classification and ranking-based QA tasks.

Under Review

ICLR 2027 · Under Review

VIEScore2: Unified Image Evaluation with Spatially Grounded Explanations

Xianda Du*, Max Ku*, Weiming Ren, Zhi Rui Tam, Chunlin Ren, Ping Nie, Min-Hung Chen, and Wenhu Chen

A unified VLM evaluator for image generation and editing that jointly predicts quality scores and defect locations on a text-native grid, trained with SFT and GRPO. Manuscript available upon request.

ICLR 2027 · Under Review

Evidence in the Loop: Training Search Agents from Knowledge Graphs

Yuxing Tian*, Xianda Du*, Weixu Zhang, Jia Ao Sun, Fengran Mo, and Jian-Yun Nie

A self-evolving search-agent framework that grounds question generation and process rewards in verified, retrievable corpus evidence from multi-hop KG paths. Manuscript available upon request.

ICLR 2027 · Under Review

Mean Flows as Controllable Priors for Solving Inverse Problems

Ahmed Abdelaal, Xianda Du, and Amir-Hossein Karimi

Full conference version with capacity-matched pixel- and latent-space comparisons across five ImageNet restoration tasks, four solvers, and four priors. Manuscript available upon request.

In Preparation

In Preparation

What Reward Do the Metrics Imply in Multimodal Sentiment Analysis?

Chunlin Ren, Xianda Du, Yuyang Zhang, and Robert Mercer

Ongoing

Render, Generate, or Both? Choosing Visual Generation Methods by Quality, Physical Accuracy, and Cost

Ongoing

Physically Counterfactual Video Generation with Programmable, Physics-Simulated Scenes

Research Experience

Research Assistant — VIEScore2

University of Waterloo · Advisor: Prof. Wenhu Chen · Mentor: Max Ku

Oct 2025 – Present
  • Co-first authored VIEScore2, a unified VLM evaluator for image generation and editing that jointly predicts quality scores and defect locations on a text-native N×N grid, with a parameter-free parser that turns predictions into spatially grounded explanations.
  • Unified 38K training examples from 5 sources (ImagenWorld, RichHF, PAL4VST, etc.) with score-only, localization-only, and joint supervision; fine-tuned Qwen3-VL-8B via SFT (DeepSpeed ZeRO-3) followed by GRPO with cell-level Dice, score-accuracy, and format rewards.
  • Reached overall-score SRCC 0.601 vs. 0.491 for Gemini-3-Flash, the strongest zero-shot VLM baseline; ranked 1st in grid IoU on 3 of 6 defect-localization benchmarks and top-3 on 5, ahead of GPT-5.6 and Claude.

Research Assistant — Controllable Generative Priors

University of Waterloo · Advisor: Prof. Amir-Hossein Karimi · Mentor: Ahmed Abdelaal

Aug 2026 – Present
  • Formulated pretrained MeanFlow finite-interval transport maps as controllable priors for training-free inverse problems, enabling one-evaluation clean-endpoint prediction and running existing flow solvers (D-Flow, PnP-Flow, MPC-Flow) without retraining.
  • Introduced Receding-Horizon State Optimization (RHSO), a model-predictive controller over the generative state; best in 35 of 60 task–metric cells and 200 of 240 pairwise comparisons across 5 ImageNet restoration tasks, 4 solvers, and 4 priors.
  • Built a unified benchmark over pixel- and latent-space priors (JiT, pMF, SiT-XL, iMF) spanning PyTorch and JAX, with backend-agnostic forward operators, sharded SLURM sweeps on Alliance clusters, and PSNR/SSIM/LPIPS reporting.

Research Assistant — Evidence-Grounded Search Agents

Université de Montréal · Advisor: Prof. Jian-Yun Nie · Mentor: Yuxing Tian

Jul 2026 – Sep 2026
  • Co-first authored Evidence-in-the-Loop, a self-evolving Proposer–Solver framework that trains LLM search agents with RL on questions synthesized online from multi-hop Wikidata KG paths, keeping only paths whose every triplet is backed by verified, E5-retrievable Wikipedia evidence.
  • Designed two process rewards, evidence acquisition and evidence utilization, that credit actual retrieval of verified passages instead of entity mentions, curbing the reward hacking of entity-based rewards (E-GRPO, KG-Path).
  • Achieved the best average EM/F1 in all 5 training settings (Qwen2.5-7B Base/Instruct, LLaMA-3.1-8B, Qwen3-8B, Search-R1-7B) over SSP and KG-Path, winning 18 of 20 multi-hop comparisons; Proposer-generated questions lift downstream multi-hop F1 from 30.4 to 35.3.

Research Assistant — LLM Reranking (RankLLM)

University of Waterloo · Advisor: Prof. Jimmy Lin

Jun 2026 – Present
  • Contributing to RankLLM, Castorini’s open-source toolkit for LLM-based listwise reranking: integrated SPLADE-v3 learned sparse retrieval as a first-stage retriever and upgraded the Pyserini backend to v2.4.0.
  • Improved reranking reproducibility by enforcing deterministic decoding, and wrote a Colab CLI onboarding guide covering installation, sanity checks, and end-to-end RankZephyr reranking.

Research Assistant — Weather Archival Rescue

McGill University · Advisor: Prof. Renée Sieber · with RBC Borealis AI and CEIMIA

Sep 2024 – Present
  • WXImpactBench (ACL 2025 Findings, co-first author): built the first benchmark for LLM understanding of disruptive weather impacts (1,736 samples, 6 impact categories) with multi-label classification and 100-passage ranking QA; evaluated 8+ LLMs (Llama-3, Qwen2.5, Mixtral, GPT) with 4-bit quantization.
  • WeatherArchive-Bench (SIGIR 2026, co-first author): built a retrieval-augmented reasoning benchmark over 1M+ archival segments and 335 expert queries with IPCC-aligned vulnerability assessment; benchmarked 13+ retrievers (BM25, SPLADE, ANCE, dense, cross-encoder reranking) and 17 LLMs, peaking at 95.8% Recall@100.
  • WXChat (ongoing): developing a digitization pipeline for Victorian-era newspapers with Tesseract 5 OCR at 600 DPI, layout-aware column segmentation (OpenCV; skew correction, CLAHE, Sauvola binarization), and async GPT-4o-mini post-OCR correction.

Undergraduate Research Assistant — Noise Injection Analysis

University of Waterloo · Advisor: Prof. En-Hui Yang

Jan 2025 – Apr 2025
  • Studied effects of Gaussian noise (9 levels, σ = 0.1–0.9) and FGSM adversarial perturbations on intermediate feature representations in pretrained ResNet-18 on ImageNet.
  • Extracted and analyzed feature maps across 4 network layers (56×56 to 7×7) using PyTorch forward hooks, generating statistical summaries and 2D cross-layer distribution histograms.

Professional Experience

Autonomy and Algorithm Engineer Intern

Waabi · Toronto, ON

Sep 2026 – Dec 2026
  • Working on novel view synthesis (NVS) quality evaluation for sensor-simulation reconstruction models in self-driving, an exploratory effort to design metrics and pipelines that assess rendering fidelity at novel viewpoints.

Machine Learning Engineer

StackAdapt · Toronto, ON

Jan 2026 – Apr 2026
  • Shipped StackAdapt’s Precision Audience LLM pipeline that converts unstructured advertiser ICPs into executable boolean segment plans across 9 audience sources × 10 geo locations per request, built on GPT-5 with Pydantic-typed structured outputs and an async worker pool backed by Redis job tracking.
  • Eliminated silent data corruption by building a 4-layer hallucination detection system (schema enforcement, reference-grounding against candidate sets, strict enum validation, and post-LLM safeguard rules), ensuring every audience segment in the output corresponds to a validated upstream source.
  • Reduced unsafe ad targeting by designing prompt hardening (injection detection, 10-category sensitive-data taxonomy) and a typed exception hierarchy that maps failures to correct HTTP semantics, enabling per-geo partial-success (207) so multi-country requests degrade gracefully.
  • Scaled audience coverage to ~12.7K segments across multiple countries with hybrid vector retrieval on Qdrant (dense + BM25 + RRF, geo-indexed prefilter) and an end-to-end data pipeline with ISO validation, diff-based incremental ingestion, and snapshot/restore rollback to S3.

Software / Machine Learning Engineer

Storytellers.ai

Jun 2025 – Aug 2025
  • Built data quality management features (dynamic filtering, interactive tables, advanced tooltips) for a large-scale analytics platform, optimizing SQL queries to cut load times by up to 15%.
  • Integrated the Snowflake data warehouse via a secure OAuth flow with automated permission checks and metadata retrieval through REST APIs.
  • Developed an AI-powered documentation system on Cloudflare Workers AI and Durable Objects that generates table descriptions from schema and statistical metadata.
  • Replaced static service-account keys with GCP impersonation in cloud authentication workflows, strengthening platform security.

RAG Project Technical Lead

University of Waterloo, WatAI · Waterloo, ON

Apr 2025 – Dec 2025
  • Designed and implemented a research-oriented RAG system for educational QA, supporting Qwen, OpenAI, and Gemini through a flexible LangChain-based interface with configurable prompts, per-user API keys, and streaming inference.
  • Built an end-to-end retrieval pipeline over course PDFs and slides with document chunking, dense retrieval in Supabase, optional hybrid BM25 + dense search, re-ranking, and citation-grounded generation.
  • Developed evaluation and monitoring workflows using Recall@k, nDCG, and LLM-as-a-judge faithfulness checks, with run tracing (model, prompt version, retriever settings) to tune RAG on real-user chats.
  • Implemented tool-augmented querying (retrieval, file parsing, syllabus/FAQ lookup) and a model router balancing quality, latency, and cost, with timeouts, retries, and safe fallbacks.

Platform Developer

University of Waterloo, ECE Department · Waterloo, ON

Sep 2024 – Dec 2024
  • Fine-tuned Hugging Face sentiment analysis models with TensorFlow/PyTorch for media content classification, achieving 80% accuracy.
  • Built a pipeline that transformed embeddings and metadata from a PostgreSQL database on GKE into a self-designed neural network for scraped-article classification, using pandas for data curation and labeling.
  • Optimized news-feed ranking with Reciprocal Rank Fusion (RRF), and set up a CI/CD pipeline on GCP (Network Endpoint Groups, Load Balancers, GKE workloads) for Dockerized services.

Software Engineer

LTIMindtree

Jan 2024 – Apr 2024
  • Developed, tested, and deployed internal APIs with Spring and OAuth2, handling data transformation and system integration.
  • Automated cloud-based log downloads with Gson, HttpClient, SOAPUI, and MuleSoft, including stress testing.
Jun 2025 – Nov 2025
  • Framed 15-min crypto returns as ranking under Weighted Spearman; trained TabNet, XGBoost, MLP, and LightGBM via Optuna, finalizing an Optuna LightGBM scoring 0.57 Weighted Spearman on the public leaderboard.
  • Implemented leak-safe, rolling-window features (RSI divergence, distance-to-high/low, volatility) with vectorized NumPy/pandas; importance screening kept 80 features, improving generalization and training speed.
  • Built an 8-model LightGBM ensemble (varying depth, regularization, seeds, losses) with weighted blending and adaptive clipping, winsorization, normalization, and smoothing to stabilize ranks.

Skills & Service

Machine Learning & AI

PyTorchJAXTensorFlowKerasTransformers Scikit-learnDeepSpeedXGBoostLightGBMHPO

Research Areas

Information RetrievalSearch AgentsRAG Vision-Language ModelsRL for LLMs (GRPO)Generative Models

Programming Languages

PythonJavaScript/TypeScriptC/C++ JavaSQLBashRMATLABHTML/CSS

Developer Tools

GitDockerKubernetesSLURM Cloud PlatformsLangChainScrapySupabaseQdrant

Data & Analysis

NumPyPandasOpenCVMatplotlib

Languages

Chinese (Native)English (Fluent)

Service & Open Source

Get in Touch

I am actively looking for graduate and research positions in retrieval, multimodal AI, and generative modeling. Feel free to reach out.

a32du@uwaterloo.ca github.com/XiandaDu linkedin.com/in/xianda