Systems, experiments, and applied research.
Explore Lifan Hu’s work in graph reasoning, model evaluation, agent behavior, memory systems, multimodal AI, and creative code.
- MisGraph — A reproducible graph-reasoning benchmark with paired node-renaming tests, inspectable model responses, and a browser-based evaluation runner.
- shuorenhua Bench — A context- and population-conditioned benchmark for naturalness, direct usability, revision burden, and semantic preservation in model-generated Chinese text.
- Pink Elephant — A provider-agnostic experiment harness for measuring how corrected mistakes can persist and spread through later model outputs and project artifacts.
- AGEA — Adaptive Graph Evidence Acquisition: a sequential framework that selects useful graph evidence before LLM-based fraud reasoning under token and latency constraints.
- C-OS — A persistent cognitive memory system with temporal graph memory, contradiction-aware fact versioning, hybrid retrieval, and an evaluation harness.
- CIRT — A PyTorch implementation of Competitive Information Routing Transformers, replacing purely additive residual mixing with learned competitive routing.
- Linguine — A natural-language programming language with formal semantics and a clean compiler pipeline, designed to make readable intent executable without giving up structure.
- GAiVE — A story-aware pipeline that connects events across gameplay footage, selects a narrative arc, and produces a machine-readable short-video edit plan.
- Pose-Based Risk Detection — A privacy-first, real-time multi-person pose pipeline for fall, instability, inactivity, and bed-exit risk detection without storing raw frames.
- Lie Group Trajectory Encoder — Learned generators from trajectories in SE(2), SE(3), SO(3), and SL(2,R), with manifold visualizations and robustness studies.
- Recursive Web — A browser-based, multi-level ARG with terminal commands, hidden clues, binary hints, and an ASCII-art finale.