Haifeng Li

Research profile

Research summary

Haifeng Li is a researcher at Central South University (China) whose recent work sits at the intersection of multimodal large language models and physically grounded AI evaluation. The Evolve-reviewed portfolio centers on two 2026 arXiv efforts: FireWorldBench, a 520-world, 9,074-question benchmark for coupled-field fire dynamics that probes whether multimodal LLMs can reason about complex physical systems, and RS-Claw-Evolution, an environment-feedback-driven evolution framework that trains lightweight remote-sensing agents for long-horizon tasks, lifting Qwen3-4B to 65.9% on Earth-Bench AP. OpenAlex records attribute 7 works over the recent three-year window, spread across five topic clusters led by multimodal machine learning applications (4) and advanced neural network applications (3), with an even 50/50 split between LLM agents for remote sensing and multimodal physical-reasoning benchmarks. Collaboration data show 20 distinct coauthors with 12 recurring partners, yielding mid-range collaboration (61) and activity (61) scores; reproducibility is comparatively strong (78), driven by a high-readiness benchmark release and a medium-readiness agent pipeline. Citation-based output and impact cannot yet be assessed, so this profile should be read as an early-stage, preprint-centered view of an emerging applied-AI researcher.

Research keywords

Highlights

• FireWorldBench: 520 coupled-field fire worlds probing MLLM physical reasoning

  Co-authored benchmark effort (middle authorship, 2026, arXiv) built on FDS simulation with 520 fire worlds and 9,074 QA pairs, exposing limits of multimodal LLM physical-world reasoning. Evolve notes the caveat that real-world gold answers are simulation-derived rather than measured; reproducibility readiness is rated high.

• RS-Claw-Evolution: lightweight remote-sensing agents beating 32B baselines

  Last-author 2026 arXiv work introducing environment-feedback-driven evolution for long-horizon remote-sensing tasks: a three-stage training pipeline lifts Qwen3-4B to 65.9% on Earth-Bench AP, surpassing 32B baselines. Evolve flags single-run, benchmark-coupled evidence and medium reproducibility readiness.

• Strong reproducibility signal across scored outputs (78/100)

  Across the two Evolve-reviewed papers the reproducibility mean is 78: the benchmark release rates high readiness while the agent-training pipeline rates medium. This artifact-oriented pattern — reusable evaluation resources alongside training code — is the most distinctive quality signal currently on record for this researcher.

Timeline

1. 2026FireWorldBench released on arXiv: a coupled-field fire dynamics benchmark with 520 fire worlds and 9,074 QA pairs, rated high reproducibility readiness.
2. 2026RS-Claw-Evolution published on arXiv: environment-feedback-driven evolution lifts lightweight agent Qwen3-4B to 65.9% on Earth-Bench AP, above 32B baselines.
3. 2026OpenAlex countsByYear and recent3yWorks both record 7 works, with the entire visible recent output concentrated in 2026.
4. 2026Collaboration network spans 20 distinct coauthors in the recent window, with 12 recurring frequent partners in the Evolve record.

Papers in the Evolve catalog

• 2026-09-19

  FireWorldBench: Benchmarking Complex Physical World Intelligence through Coupled-Field Fire Dynamics

• 2026-09-06

  RS-Claw-Evolution: Environment-Feedback-Driven Evolution for Lightweight Remote Sensing Agents in Long-Horizon Tasks

Frequent co-authors