01research agenda
Predict
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.
Undergraduate researcher / Renmin University of ChinaBeijing, China
LinfengFan
B.S. candidate in Artificial IntelligenceGaoling School of Artificial Intelligence, Renmin University of China
Verifiable world models for embodied agents
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?
Research position
A reliable action needs a traceable state.
03 connected questions
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?
Reliable embodied agents need predictive states whose claims remain traceable to vision, proprioception, contact, and the origin of information.
I study how agents can ground decisions in the right evidence, detect when that evidence is insufficient, and make targeted corrections before acting.
Long-term goal
Self-correcting dexterous manipulation through latent predictive world models.
02published evidence
Verify
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.
03published evidence
Correct
Reliable multimodal reasoning
Locate spatiotemporal evidence at the decoding layers where video grounding can still be repaired, then test inconsistent temporal reasoning.
Decision boundary / research agenda
Where should an action stop?
Vision, proprioception, contact, and provenance converge at an evidence boundary. The question is whether to act, correct the relevant state, or stop when evidence is insufficient.
Evidence boundarySufficient?
OutputAct / correctIf evidence fails, repair the relevant state or stop.
Personal archive / 02 frames
Outside
the margins.
Two photographs, presented without a brief.
Selected work
Systems and papers,
claims and limits.
02 focused records
- Type
- system
- Status
- ongoing
- Role
- Project Lead
- Period
- Apr. 2026-Present
Embodied Dexterity
Dexterous-Hand and Humanoid Embodied Intelligence
Claim
Real-system work spanning dexterous-hand policy training, cross-platform integration, and ongoing work toward joint G1 whole-body and dexterous-hand training.
Evidence
Trained a Wuji dexterous hand mounted on a Tianji robotic arm across multiple manipulation tasks using NVIDIA GR00T. Migrated and integrated the same Wuji hand onto a Unitree G1 humanoid and coupled it with the team's GMT control stack.
Industrial research / Apr. 2026-present
Four tracks, one focus.
Two completed foundations, one active training track, and one planned data track converge on the current G1 whole-body and dexterous-hand program.
Current focusG1 + hand
Long-term goalSelf-correctingDexterous manipulation remains the long-term goal.
- Type
- paper
- Status
- arXiv preprint
- Role
- First author
- Period
- 2026
STEAR
STEAR
STEAR: Layer-Aware Spatiotemporal Evidence Intervention for Hallucination Mitigation in Video Large Language Models
Claim
At high-entropy decoding steps, one token-conditioned middle-layer evidence hypothesis can support grounding repair and a temporal counterfactual check.
Evidence
STEAR selects token-conditioned visual evidence, reinjects it at grounding-sensitive middle layers, and reuses it to construct temporal counterfactuals within a single video-encoding pass.
- Evaluated across three backbones and four benchmarks
- LLaVA-Video-7B, InternVL2.5-8B, and Qwen2.5-VL-7B across EventHallusion, VidHalluc, NExT-QA, and MVBench.
- 43.13% to 73.53% accuracy
- On LLaVA-Video-7B, EventHallusion Misleading split.
Publications
Every claim keeps its scope attached.
03 public preprints
Name highlighted = Linfeng Fan* Equal contributionAll records link to arXiv
PACT
The Granularity Mismatch in Agent Security: Argument-Level Provenance Solves Enforcement and Isolates the LLM Reasoning Bottleneck
Claim, evidence & results
Claim
Untrusted information becomes dangerous when its provenance is allowed to bind an authority-bearing tool argument.
Evidence
PACT assigns semantic roles to tool arguments, tracks value provenance across replanning, and checks each argument against a role-specific trust contract before execution.
- 100% utility and 100% security
- Oracle-provenance mixed-trust diagnostic suites.
- 100% security with 38.1-46.4% utility
- The three strongest of five evaluated models on AgentDojo; utility was 8-16 percentage points above CaMeL at the same security level.
SkillGuard
Skill Drift Is Contract Violation: Proactive Maintenance for LLM Agent Skill Libraries
Claim, evidence & results
Claim
A skill drifts when environmental change violates a role-bearing execution contract, not whenever a referenced value changes.
Evidence
SkillGuard separates operational obligations from incidental mentions, validates role-bearing environment contracts, and uses violations to localize repair.
- Zero false alarms across 599 negative controls
- No-drift and hard-negative cases; Wilson 95% CI [0%, 0.6%].
- 100% precision and 76% recall
- Known-drift verification with the strongest evaluated backbone.
- One-round repair improved from 10% to 78%
- Repair with contract-violation localization compared with no localization.
STEAR
STEAR: Layer-Aware Spatiotemporal Evidence Intervention for Hallucination Mitigation in Video Large Language Models
Claim, evidence & results
Claim
At high-entropy decoding steps, one token-conditioned middle-layer evidence hypothesis can support grounding repair and a temporal counterfactual check.
Evidence
STEAR selects token-conditioned visual evidence, reinjects it at grounding-sensitive middle layers, and reuses it to construct temporal counterfactuals within a single video-encoding pass.
- Evaluated across three backbones and four benchmarks
- LLaVA-Video-7B, InternVL2.5-8B, and Qwen2.5-VL-7B across EventHallusion, VidHalluc, NExT-QA, and MVBench.
- 43.13% to 73.53% accuracy
- On LLaVA-Video-7B, EventHallusion Misleading split.
Research trajectory / 2026
Ground.Predict.Correct.
The long-term agenda is embodied: compact predictive states, grounded in the evidence that matters, able to detect and repair their own failures before action.
Record
The research sits inside a real trajectory.
Education / experience / recognition
Education / currentSep. 2023-Jun. 2027 (expected)
Renmin University of China
Gaoling School of Artificial Intelligence
- Degree
- Bachelor of Science in Artificial Intelligence
- Program
- Elite Talent Training Program
- GPA
- 3.7/4.0
Experience
Build, integrate, verify.
Apr. 2026-Present / ongoing
Dexterous-Hand and Humanoid Embodied Intelligence
Project Lead · Beijing Noetix Robotics Technology Group Co., Ltd.
Real-system work spanning dexterous-hand policy training, cross-platform integration, and ongoing work toward joint G1 whole-body and dexterous-hand training.
- Completed
Trained a Wuji dexterous hand mounted on a Tianji robotic arm across multiple manipulation tasks using NVIDIA GR00T.
- Completed
Migrated and integrated the same Wuji hand onto a Unitree G1 humanoid and coupled it with the team's GMT control stack.
- In progress
Extending the system toward joint G1 whole-body and dexterous-hand training, with iterative work on coordination, manipulation, and robustness.
- Planned
Planning a HumanEgo-inspired wearable-glasses demonstration and human-to-dexterous-hand retargeting pipeline to reduce task-specific real-robot demonstration collection.
Jun.-Aug. 2025 / completed
Bimanual Teleoperation System and Sim-to-Real Deployment
Deputy Captain · RUC-HUHA Humanoid Robot Team
Contributed to integration, real-robot debugging, and coordination for an 11-person humanoid robotics team.
- Completed
Built and debugged the VR-to-ROS teleoperation pipeline, including relative coordinate mapping and proportional workspace scaling for dual-arm control.
- Completed
Contributed to MoveIt2-based bimanual motion planning and sim-to-real deployment, including execution deviation, workspace, collision-avoidance, and safety debugging.
Selected recognition
Evidence of range.
World Humanoid Robot Games
2nd Place (Team)
RUC-HUHA Humanoid Robot Team
National Undergraduate Mathematical Modeling Contest
First Prize, Beijing Division
National Undergraduate Innovation and Entrepreneurship Training Program
Beijing Outstanding Project Grant