Han Muchen (Miranda)

Research

Co-first author of a paper submitted to NeurIPS 2026. Title, anonymous repository identifier, and distinctive method / numerical details are withheld below to preserve double-blind review integrity; full information is available on request after the review period.

Co-first author. Deployment-failure audit suite for Vision-Language-Action (VLA) models — mutually-exclusive failure-mode taxonomy with calibrated confidence intervals and a CI-style ship-gate; demonstrates shipping-blocking failure modes that look acceptable on aggregate-success metrics alone.

Experience

Founding Researcher, Φ(fight) Research May 2026 – present

Independent research collective for physical-world AI fighting: adversarial multi-agent VLA, world models under contested dynamics, mechanistic interpretability of self-play policies.
Three papers in progress for ICLR 2027 (co-authored with Liu Yuchen): one benchmark (Φ-Arena, co-led), one mechanistic interpretability paper (collaboration), one energy-bounded adversarial games paper (collaboration).

Undergraduate Research Opportunities Program (UROP), HKUST 2025 – Present

Explainable AI from Cognitive Science Perspectives. Advisor: Prof. Janet H. Hsiao, Attention Brain and Cognition (ABC) Lab.

AI-side of a project that uses human attention data (eye-tracking, EEG) to evaluate and improve saliency-based XAI methods for deep vision models, building on the lab's line of work on human-attention-guided XAI for object detection.
Comparing model saliency maps (GradCAM, attention rollout) against human attention as ground truth; using EEG-decoded neural representation quality as an additional faithfulness signal for XAI explanations.

Education

Hong Kong University of Science and Technology Sep 2025 – 2029 (expected)

BEng in Computer Science, School of Engineering.

Technical Skills

Languages. Python, Rust, TypeScript.

Research & ML.

Statistical evaluation for ML deployment — failure-mode taxonomies, Wilson and bootstrap confidence intervals, CI-style ship gates for closed-loop VLA rollouts.
Computer vision & XAI — saliency methods (GradCAM, attention rollout), human-attention-guided faithfulness evaluation.
EEG & signal processing — multi-channel EEG, decoding-based readouts of neural representation quality.
Stack — PyTorch with HuggingFace Transformers / datasets; NumPy / SciPy / pandas.

Rust. Async services and CLIs with tokio / clap / reqwest / serde; self-hosted VPS proxy manager (sing-box, VLESS+Reality, Hysteria2).

Web. Astro / TypeScript; built and deployed a static research-collective site.

Platforms. Native development across Windows, macOS, and Linux (Ubuntu, Arch Linux).