Fengrui Lab

AI scientists for the physical world.

We build full-stack agent systems that connect scientific reasoning with the physical world.

Group

Fengrui Lab is a research group at Harbin Institute of Technology (HIT), led by Professor Stamoulis Dimitrios. We are based in the Faculty of Computing and the State Key Laboratory of Smart Farm Technologies and Systems, directed by Professor Jie Liu (IEEE Fellow).

We are hiring. We welcome prospective students interested in AI agents, scientific discovery, and real-world systems.

Vision

Our goal is to build AI scientists capable of tackling complex, long-horizon tasks whose evidence and outcomes change across space and time. Scientific work requires more than a single prediction: it involves forming hypotheses, planning experiments, interpreting results, and revising decisions as new observations arrive.

We study the full stack that makes this possible: event-driven world models for cyber-physical systems, edge deployment, multi-agent controllers and orchestration, scientific hypothesis generation, multimodal reasoning, and reusable knowledge libraries. Our evaluation suites follow entire scientific workflows, examining the quality of decisions, their consequences, and the resources they consume.

We connect these methods to smart agriculture, smart buildings, smart cities and low-altitude aerial systems, and bioengineering, including crop breeding. These applications ground our research in real data and physical constraints. We aim to make expert-level investigation more accessible, reproducible, and useful in practice.

Publications

Recent papers

  1. EMNLP 2026(Accepted) CCF-B

    CoFiE: Coarse-to-Fine Evidence Selection for Efficient Streaming Video Understanding

    J. Jiang, Y. Ling, R. Li, D. Stamoulis, J. Liu

    A two-stage framework for efficient streaming video understanding: novelty-based filtering removes redundant frames before the vision encoder, while query-aware refinement selects relevant evidence during LLM prefill. CoFiE reaches 78.86% accuracy on StreamingBench and 68.72% on OvO-Bench, filtering up to 80% of evidence frames and accelerating end-to-end inference by up to 2.54×.

    arxiv.org/abs/2609.03675

  2. PRICAI 2026 (Accepted) CCF-C

    Position: We Should Evaluate Agentic Memory as Inference, not as a Tape Recorder

    Y. Hadjiyianni, P. Michelakis, E. Neofotistos, J. Jiang, D. Stamoulis, J. Liu

    A position paper holding that memory is fundamentally an inference process, using past experience to form beliefs under uncertainty, and must be evaluated as one. It contributes an architecture-agnostic taxonomy of memory failure modes (schema dominance, trace fragmentation, provenance collapse) and intervention-based criteria that perturb episodic evidence while holding cue and architecture fixed. Models statistically indistinguishable on task accuracy diverge sharply under intervention, most of all in provenance.

    URL pending

  3. ACM SIGSPATIAL 2026 (Accepted) CCF-C

    Deploying and Evaluating a Smart-Agriculture Agentic Engine for Full-Season Soybean Farm Operations

    A. Qu, P. Michelakis, L. Han, Y. Hadjiyianni, K. Ouyang, K. Siskos, F. Li, R. Meng, J. Jiang, D. Stamoulis, J. Liu

    FAIRY, a full-stack agentic engine deployed on an operating soybean research farm, built around an “everything is an event” execution paradigm that treats sensing, crop-growth transitions, machinery actions, and management interventions as state-changing events in a shared process engine. Nine agent controllers are evaluated across one hundred full-season scenarios on agentic success, full-path spatiotemporal correctness, token cost, and edge-device runtime.

    arxiv.org/abs/2609.00106

  4. LAW 2025NeurIPS Workshop

    CORE: Full-Path Evaluation of LLM Agents Beyond Final State

    P. Michelakis, Y. Hadjiyianni, D. Stamoulis

    A framework built on finite automata, with five metrics that score an agent's entire execution path, not just whether the final answer happens to be correct.

    arxiv.org/abs/2509.20998