Portrait of Tianhu Xiong

Tianhu Xiong

Mathematics & Computer Science at the University of Illinois Urbana-Champaign. I work on large language and vision-language models and LLM agents, from distributed training on TPU and GPU clusters to interpretability tooling, and I write high-performance C++.

Seeking a Summer 2027 internship
GPA 4.0 / 4.0James Scholar Honors · UIUC
CVPR 2026 · EMNLP 2026Co-first author · Main track
Kaggle SilverLMSYS Chatbot Arena · RSNA Lumbar Spine, 2024
USACO PlatinumNOIP 2022 First Prize, China

About

At UIUC I work with Klara Nahrstedt in the MONET group. My research spans interpretability of vision-language models, multi-agent systems for editing slide decks, and test-time training for diffusion language models. Before college, I learned to program through C++ and competitive programming.

I am interested in software engineering on AI-powered products and foundation models: understanding how models behave, the systems that train and serve them, the agents built on top of them, and the products people use them through. I also enjoy high-performance backend work, from networking to distributed systems.

Research

Research Assistant · MONET Group, UIUC

Oct 2025 – present

Advised by Prof. Klara Nahrstedt.

Figure 1 of the paper: an attribution graph tracing how Gemma-3-4B recognizes a sea otter from image and text tokens Figure 13 of the paper: circuit tracing for an image of '1 + 2 =', where digit and chart-range features feed a math subgraph that outputs 3
CVPR 2026 Findings co-first author

Circuit Tracing in Vision–Language Models: Understanding the Internal Mechanisms of Multimodal Thinking

Jingcheng Yang*, Tianhu Xiong*, Shengyi Qian, Klara Nahrstedt, Mingyuan Wu  (*equal contribution)

The first circuit-tracing framework for vision-language models. It shows which internal features Gemma-3-4B uses to reach an answer, and lets you turn those features up or down to test the explanation.

My contributions
  • Trained 34 per-layer TopK transcoders for Gemma-3-4B, one per MLP block. I extended EleutherAI Sparsify to Gemma-3's multimodal architecture and ran 30k steps on 8×H100 (about 60 hours) over a 360k-example text and image corpus.
  • Built the feature-activation pipeline. I added vision-language hook points to TransformerLens at every MLP input and output, and computed activations on demand for only the features in each attribution graph (under 2,000 of 5.6M) instead of precomputing all of them, which cut compute and storage.
  • Wrote the activation-patching and feature-steering modules that verify the circuits are causal and trace hallucinations down to individual features.
Overview of SlideForge: deck state graph, slide-native agent, theme-preserving reconstruction, verification and self-repair
EMNLP 2026 Main

SlideForge: An LLM Agent for Controllable Editing of Slides as Structured Artifacts

Haozhen Zheng*, Fulin Wang*, Tianhu Xiong, Yingjie Yu, Shengyi Qian, Hanchao Yu, Alex Schwing, Klara Nahrstedt, Mingyuan Wu

An LLM agent that edits real PowerPoint decks through their object model instead of redrawing screenshots, so every edited slide stays native and editable.

My contributions
  • Helped build the Deck State Graph pipeline as a multi-agent system that accepts a PPTX deck or a plain slide image. Agents map regions onto a 306-type component taxonomy, plan how segmentation calls are batched so detections are not duplicated, and arbitrate overlapping boxes. Native PPTX objects are attached when the source file is available.
  • Helped build the self-verification loop in decomposition. Accepted regions are whited out of the rendered slide, a VLM reviewer checks what remains for missed content, invalid detections and overlapping candidates, and its findings drive another detection pass until the slide is fully covered.
  • Built the evaluation harness: it rebuilds each slide from its extracted components, renders it headlessly and scores it against the original, and counts editable shapes with python-pptx. On 89 unseen decks, reconstruction reaches 0.950 SSIM against 0.813 for Claude Code, and restyled slides keep 72.8 editable shapes against 32.4 for a GPT-Image-2 + Claude pipeline.
COLM 2026 Workshop NonAR-LM · poster

Interpreting Diffusion Language Models Over Denoising Time

Jingcheng Yang, Shangtong Wu, Yifan Kang, Tianhu Xiong, Mingyuan Wu, Haozhen Zheng, Alex Schwing, Klara Nahrstedt

Brings circuit-style interpretability to diffusion language models. Sparse autoencoders and cross-layer transcoders trained on Dream and LLaDA show how features evolve across denoising steps, and timing-controlled interventions test which of them are causal.

In progress

Diffusion Language Models with Test-Time Training

Jun 2026 – present

A diffusion language model with long-range memory. Text is denoised block by block, left to right, and test-time-training layers keep updating a small set of fast weights from one block to the next, so the model can use information beyond its attention window. Training runs on Cloud TPU v4 through my TRC grant.

  • Built the 1.3B-parameter model in JAX/Equinox: diffusion within each block, and test-time-training layers whose causal fast-weight updates carry state across blocks.
  • Built the distributed training setup with hybrid sharding: FSDP within each node and DDP across nodes. The same code runs on TPU and GPU, from a 32-chip TPU v4 pod to the NCSA DeltaAI cluster.
JAXEquinoxXLATPU v4FSDPdistributed training

Projects

In progress

Financial Agent Safety Bench · adversarial testing for LLM agents

Oct 2026 – present

Tests whether a tool-using financial research agent stays accurate when its tools return tampered filings, false search results or hidden instructions, and whether it leaks confidential client details.

  • Built an attack environment that intercepts the agent's tool calls (web search, SEC EDGAR, page reader) and injects attacks: 374 test cases covering 16 attack types across 5 classes, with a dataset generator and schema validation.
  • Built the evaluation pipeline: deterministic leak checks on outbound tool arguments, a rubric-based LLM judge with calibration, and confidence intervals bootstrapped over companies.
PythonLLM agentsfault injectionevaluation
LiveTranslate subtitle overlay in light and dark themes, showing an English sentence and its Chinese translation side by side

LiveTranslate · offline simultaneous interpreter

May – Aug 2025

Live bilingual subtitles for anything playing on a Windows PC, with speech recognition and translation running fully on-device.

  • Built a C++/Qt desktop app that captures system audio through WASAPI loopback and resamples it with Media Foundation. Capture and rendering run on separate worker threads so the UI never blocks.
  • Integrated quantized streaming ASR (Zipformer) and offline translation through ONNX Runtime and CTranslate2, reaching under 0.5 s end-to-end latency on a laptop with no cloud calls.
C++QtWASAPIONNX RuntimeCTranslate2multithreading
P2P client main window: logged-in device, transfer-rate chart at 85 MB/s, and a list of client and server tunnels

P2P transport & NAT traversal

Jan – Apr 2025

Direct transfers between devices behind NATs. A signaling server helps the peers find each other, then data moves over a custom reliable UDP protocol.

  • Built a reliable UDP transport in C++17 and standalone Asio with adaptive retransmission timeouts and end-to-end encryption: about 90 MB/s on a clean link and 30+ MB/s with 20% simulated packet loss.
  • Built an async signaling backend on Django REST and Channels WebSockets that coordinates STUN hole punching and session setup, authenticated with device-bound JWTs.
C++17AsioUDPnetworkingDjango Channels

Education

University of Illinois Urbana-Champaign

Expected May 2028

B.S. in Mathematics & Computer Science · James Scholar Honors · GPA 4.0/4.0

Coursework: Data Structures · Algorithms & Models of Computation · Computer Architecture · System Programming · Programming Languages & Compilers · Probability & Statistics · Discrete Structures · Abstract Linear Algebra

Honors

Programming
USACO Platinum Division, 2023 US Open · NOIP First Prize, 2022 (China National Olympiad in Informatics in Provinces)
Research
CVPR 2026 Broadening Participation Scholarship, IEEE/CVF · Google TPU Research Cloud grant, 2026
ML contests
Kaggle Silver Medal × 2, 2024: LMSYS Chatbot Arena Human Preference Prediction; RSNA Lumbar Spine Degenerative Classification
Math & physics
AIME 2024, USAMO qualifier · AMC 12 Distinction, 2023 · Physics Bowl National Gold, 2024

Skills

Languages
C++, Python · also Java, JavaScript, SQL, Bash
ML systems
PyTorch, JAX / Equinox / XLA, distributed training on TPU and GPU (FSDP, DDP), TransformerLens, sparse autoencoders and transcoders
LLM agents
Multi-agent pipelines, tool use, LLM-as-judge evaluation, adversarial testing
Systems
Qt, Asio, ONNX Runtime · Django, WebSockets, Redis, MySQL · Linux, Docker, Slurm, Git