Tianqi Liu

An open benchmark and evaluation framework for measuring whether coding agents complete evolving product requirements in existing enterprise repositories. It reconstructs PRDs, synthesizes and validates atomic Code Points, localizes code evidence, and attributes unmet requirements to implementation and process symptoms. The project reports 43 enterprise tasks, 750 audited Code Points, 11 model–scaffold settings, and 95.2% expert-consensus agreement.

A device-cloud collaborative framework for personalized mixed-precision quantization in sequential recommendation. It uses frozen weights, channel-wise quantization, user/device-specific sensitivity analysis, and one-forward-pass customization to balance recommendation quality, compression, latency, and heterogeneous device resources.

A multimodal GUI-reasoning paradigm that inserts UI elements between screen and action in a cyclic Screen–UI Elements–Action loop. It introduces UI-element-driven reinforcement fine-tuning and a 26K-sample benchmark covering localization, semantic understanding, and practical usage.

A benchmark for desktop GUI agents on long-horizon professional creative and engineering workflows with human-in-the-loop collaboration. It covers 538 executable tasks, multiple difficulty levels, and mid-turn and post-turn interaction protocols across design, video, audio, and 3D software.