Hi, This is Jingming Chen

Benchmark on Latent and Shifting User-Intent in Long-Horizon Agents

• Designed a benchmark that represents user intents as graphs, enabling precise manipulation and measurements of human-agent intent alignment and changes in long-horizon agentic tasks via graph projection and edits.
• Built an automated pipeline that generates text and multimodal long-horizon tasks covering coding and deep research domains, and publicly released 256 difficulty-calibrated tasks with verifiable web-evidence records.
• Evaluated 11 frontier closed- and open-source models across various configurations using a suite of process-level and outcome-based metrics.

Agentic Framework and Data-Generation Pipeline for Deep Information-Seeking

• Designed a novel training-free agentic framework for deep research with performance that surpasses strong baselines on challenging benchmarks such as BrowseComp, GAIA, and Xbench.
• Implemented innovative components for the agentic framework that combine hierarchical planning, self-correction, confidence-weighted reflection, folding and checkpoint-based context management, critical-error diagnosis, and trajectory rollback error-correction.
• Utilized the framework as a training-data generation pipeline and performed SFT on two open-source LLMs, resulting in models that also outperformed baselines across multiple benchmarks.

• Researched and designed a multi-agent pathfinding (MAPF) algorithm for the competition.
• Built a multi-agent LLM system that solves the MAPF problem through hierarchical roles.
• Constructed a traditional multi-agent pickup and delivery (MAPD) solver with windowed priority-based search, rolling-horizon replanning, orientation-aware A* pathfinding, and FIFO task constraints.

CSG Smart Science & Technology

• Researched, designed, and built machine-learning models that use weather data and sky imaging to generate real-time forecasts of photovoltaic power output for large-scale solar farms so that power grids can anticipate sudden changes in solar output to optimize grid efficiency.
• Extracted motion vectors from ground-based sky imaging to create ultra-short-term forecasts of cloud movement and cloud coverage to calculate power output.
• Integrated forecasting models with the company's energy management system to create a user-friendly GUI for visualizing power forecasts.