Linfeng Fan

Public record / 2026

Privacy reviewed

Curriculum Vitae

Linfeng Fan

Verifiable world models for embodied agents

Beijing, China2023200424@ruc.edu.cnlinfengfan.com
Back to profileDownload CV PDF

01 / Profile

Research profile

I study how agents can ground decisions in the right evidence, detect when that evidence is insufficient, and make targeted corrections before acting.

Current role
Undergraduate researcher
Affiliation
Gaoling School of Artificial Intelligence, Renmin University of China
Research question
How can an agent recognize when its evidence is no longer sufficient for action, then correct the right part of its state instead of continuing blindly?
Long-term goal
Self-correcting dexterous manipulation through latent predictive world models.

02 / Research

Research areas

  1. 01

    Predictive world models for embodied agents

    Build compact latent states grounded in vision, proprioception, and contact, with the long-term goal of self-correcting dexterous manipulation.

  2. 02

    Evidence-grounded agent systems

    Trace which evidence may control an action, and detect when environmental change breaks an operational obligation rather than merely changing an incidental value.

  3. 03

    Reliable multimodal reasoning

    Locate spatiotemporal evidence at the decoding layers where video grounding can still be repaired, then test inconsistent temporal reasoning.

03 / In progress

Ongoing research

Oct. 2026-Present

First-author research project

Targeting ICML 2027

What Should Mass Knowledge Editing Preserve? Key Magnitude and Factual Confidence

Testing whether activation magnitude identifies factual knowledge worth preserving during mass knowledge editing. The initial pilot has not established an advantage for confidence weighting.

  • CompletedAnalyzed the implicit squared key-norm weighting in MEMIT's old-memory regularizer and reproduced MEMIT and Forward Replay on GPT-J 6B.
  • CompletedBuilt five trace- and ridge-matched regularizer variants over 50,000 background token positions across layers 3-8, testing norm removal, token likelihood, their combination, and shuffled controls on 100 jointly edited facts.
  • CompletedIn the 100-edit pilot, norm removal preserved efficacy, paraphrase, and neighborhood preference scores at 100.0%, 96.5%, and 81.5% for both variants; token-likelihood weighting did not outperform shuffled controls.
  • In progressDesigning a pre-edit known-fact benchmark and stress tests with at least 5,000 edits to evaluate fact-level confidence weighting.

04 / Publications

Preprints

  1. 2026First author

    The Granularity Mismatch in Agent Security: Argument-Level Provenance Solves Enforcement and Isolates the LLM Reasoning Bottleneck

    Linfeng Fan*, Ziwei Li*, Yuan Tian, Yichen Wang, Rongsheng Li, Xiong Wang†

    arXiv preprint · 2605.11039

    arXiv record
  2. 2026First author

    Skill Drift Is Contract Violation: Proactive Maintenance for LLM Agent Skill Libraries

    Linfeng Fan, Yuan Tian, Ziwei Li, Zhiwu Lu†

    arXiv preprint · 2605.10990

    Planned submission to CVPR 2027

    arXiv record
  3. 2026First author

    STEAR: Layer-Aware Spatiotemporal Evidence Intervention for Hallucination Mitigation in Video Large Language Models

    Linfeng Fan, Yuan Tian, Ziwei Li, Zhiwu Lu†

    arXiv preprint · 2604.03045

    Planned submission to CVPR 2027

    arXiv record

* Equal contribution · † Corresponding author

05 / Collaborations

Additional collaborative research

  1. Collaborator

    Known-Subspace Rate Ladder for Streaming Low-Rank Regression with Fast Coefficient Drift

    Under review at ICLR 2027

  2. Collaborator

    Auditing Applied Dose in Multi-Layer Subspace Ablations

    Under review at ICLR 2027

  3. Collaborator

    Residual Scaling for Language Models: An Exact Layer Score and an Error Budget for Composing It

    Under review at ICLR 2027

06 / Experience

Research & projects

Sep. 2026-Present

Beijing, China

LatentVerse

Dexterous-Hand Foundation Models and Teleoperation

Embodied AI Algorithm Intern

Building teleoperation infrastructure for the team's dexterous-hand foundation-model effort and human-in-the-loop experiments on physical hardware.

  • CompletedIndependently developed a Meta Quest-based teleoperation system for a Tianji robotic arm and publicly released its code on GitHub.
  • CompletedBuilt glove-based teleoperation pipelines using MANUS and Wuji systems for the company's in-house dexterous hand and Wuji dexterous hands.
  • In progressContributing teleoperation infrastructure to responsive human-in-the-loop experiments on real hardware within the team's foundation-model effort.

Apr. 2026-Oct. 2026

Beijing, China

Beijing Noetix Robotics Technology Group Co., Ltd.

Dexterous-Hand and Humanoid Embodied Intelligence

Project Lead · Advisor: Eric Tian

Real-system work spanning dexterous-hand policy training, cross-platform integration, and ongoing work toward joint G1 whole-body and dexterous-hand training.

  • CompletedTrained a Wuji dexterous hand mounted on a Tianji robotic arm across multiple manipulation tasks using NVIDIA GR00T.
  • CompletedMigrated and integrated the same Wuji hand onto a Unitree G1 humanoid and coupled it with the team's GMT control stack.
  • In progressExtending the system toward joint G1 whole-body and dexterous-hand training, with iterative work on coordination, manipulation, and robustness.
  • In progressDesigning a HumanEgo-inspired wearable-glasses demonstration and human-to-dexterous-hand retargeting pipeline to reduce task-specific real-robot demonstration collection.

Jun.-Aug. 2025

Beijing, China

RUC-HUHA Humanoid Robot Team

Bimanual Teleoperation System and Sim-to-Real Deployment

Deputy Captain

Contributed to integration, real-robot debugging, and coordination for an 11-person humanoid robotics team.

  • CompletedBuilt and debugged the VR-to-ROS teleoperation pipeline, including relative coordinate mapping and proportional workspace scaling for dual-arm control.
  • CompletedContributed to MoveIt2-based bimanual motion planning and sim-to-real deployment, including execution deviation, workspace, collision-avoidance, and safety debugging.

07 / Education

Education

Sep. 2023-Jun. 2027 (expected)

Renmin University of China

Bachelor of Science in Artificial Intelligence · Elite Talent Training Program

Gaoling School of Artificial Intelligence · Beijing, China

GPA 3.7/4.0

08 / Honors

Selected honors

2025

Renmin University of China

Merit Student Award

2025

World Humanoid Robot Games

RUC-HUHA Humanoid Robot Team

2nd Place (Team)

2024

Renmin University of China

Merit Student Award

2024

National Undergraduate Mathematical Modeling Contest

First Prize, Beijing Division

2024

National Undergraduate Innovation and Entrepreneurship Training Program

Beijing Outstanding Project Grant