Literature Secretary
Ada
Collects papers, parses PDFs, keeps Zotero-oriented metadata tidy, and prepares structured source material for the rest of the lab.
PDF ingest
Zotero
Search
One Person Lab / AI Research Studio
HLF Lab
A personal research studio led by one human leader and supported by a small AI crew, built to shape a complete, rational research system with elegance, discipline, and genuine curiosity.
- Leader
- Lingfeng He
- Base
- NJUST
- Focus
- Reasoning / Generation / Agents
AI Crew
Five working roles that carry the lab from literature intake to durable research output.
Reading Director
Mira
Reads papers in chunks, synthesizes reports, performs method deep dives, and turns claims, methods, results, and limits into usable judgment.
Deep read
Methods
Limits
Proposal Director
Noah
Searches for gaps, turns observations into candidate proposals, and sketches experiment plans that stay grounded in falsifiable questions.
Gap finding
Ideas
Experiments
Knowledge Secretary
Iris
Converts reading into literature cards, journals, and durable memory, keeping the lab's local knowledge base organized and reusable.
Cards
Journal
Memory
Communications Director
Theo
Shapes drafts, presentation plans, and research-group briefings from the lab's accumulated reading, proposals, and experiment direction.
Writing
Slides
Notion
Research Lens
Current Research Directions
Multimodal LLM Reasoning
Understanding and improving how multimodal models reason across language, images, context, and implicit intent.
Image Editing and Generation
Designing generative systems that follow high-level goals, preserve semantic consistency, and produce controllable visual changes.
AI Agents
Exploring agents that can plan, call tools, inspect feedback, and complete long-horizon tasks with dependable reasoning.
Lab Output
Selected research, publications, and systems.
Papers
arXiv preprint
Reasoning-Guided Generative Image Editing
This work explores image editing tasks that require world knowledge, abstract intent, and implicit user needs, studying how multimodal language understanding can be connected with visual generation through a more robust reasoning bridge.
IGARSS 2023 · EI indexed
Weakly-supervised ROI Extraction Method Based on Contrastive Learning for Remote Sensing Images
This work uses Grad-CAM to generate pseudo labels and applies contrastive learning to mitigate label noise in weakly supervised ROI extraction for remote sensing images.
Projects
YOLO + DeepSeek
Animal Health Monitoring System
A farm animal monitoring system that uses YOLO-based behavior detection, tracking, video streams, and DeepSeek-powered health analysis to help inspect activity patterns and potential risks.
Multi-agent research assistant
LabCrew Agent Design
A local-first research assistant built around a coordinator and focused subagents for paper reading, literature cards, proposal generation, experiment planning, writing, and presentation work.
WorldQuant BRAIN workflow
brain_agent Design
A local orchestration system for WorldQuant BRAIN research that manages idea generation, alpha simulation, diagnostics, variant search, memory, submit-readiness checks, and run reports.
Research Notes
Multimodal reasoning
Short notes on model behavior, failure cases, benchmarks, and reasoning traces.
Agent Logs
LabCrew and generation
Implementation records for agent workflows, image editing systems, and experiments that deserve to be remembered.
Quiet Journal
Literature, art, and private weather
A restrained space for reading, images, fragments, and the slower questions that sit beside technical work.
Contact