Jin Bai

ECCV 20262 papers ICLR 20272 submissions

I am pursuing an M.Eng. in Computer Technology at the University of Chinese Academy of Sciences (UCAS), School of Space Exploration, and conduct my graduate research at the Technology and Engineering Center for Space Utilization (CSU), Chinese Academy of Sciences. I received the 2026 National Scholarship.

Reliable Visual Intelligence & Scientific AI Agents

Visual Understanding Geometry-aware segmentation, foundation models, thin structures

Scientific AI Agents Reinforcement learning, reusable skills, verified execution

Selected work. At ECCV 2026, FGOS-Net addresses geometry mismatch in thin-structure state-space models, while IP-SAM enables prompt-absent deployment. My current agent projects study verified policy-skill co-evolution and rubric-based evaluation for long-horizon scientific tasks.

Seeking a full-time research internship in Beijing (3-6 months), focusing on AI agents, reinforcement learning, and multimodal models. Also open to Fall 2028 Ph.D. opportunities. Contact: baijin25@mails.ucas.ac.cn.

Full profile

Research perspective. My work sits at the intersection of computer vision, Earth observation, and learning agents. I study how AI systems can remain structurally faithful to visual evidence and operationally verifiable when deployed on scientific tasks.

Computer vision. As first author of FGOS-Net (ECCV 2026), I investigate geometry-aware serialization for thin structures. I also contribute to IP-SAM (ECCV 2026), which enables prompt-conditioned foundation models to operate when deployment prompts are absent.

Agents and evaluation. TerraWeave connects verified Earth-observation execution with reinforcement and reusable skill learning, while EarthArena evaluates long-horizon scientific agents through task-specific rubrics over answers, artifacts, and trajectories.

Open research. I am a core contributor to RSMeM, which received Second Prize in the Open Source GeoAI Practice Challenge at IEEE APGRSS 2026. I welcome collaborations and research internship opportunities around reliable visual intelligence and scientific agents.

News

10/2026 RSMeM won Second Prize in the Open Source GeoAI Practice Challenge at IEEE APGRSS 2026.
10/2026 TerraWeave and EarthArena were submitted to ICLR 2027 and are under review.
06/2026 FGOS-Net (Bridging the Geometry Mismatch) and IP-SAM were accepted to the ECCV 2026 main conference.
01/2026 Fluxamba was released as an arXiv preprint and is under review.
11/2025 M²F-SAM was presented at the 4th China Space Science Assembly in Beijing.
09/2025 Started postgraduate study at UCAS and CSU.

Research Experience

Apr. 2026
- Present

THUNLP & Qiyuan Lab

Research Intern — AI agents and remote sensing foundation models; contributed to research papers on verified execution and reliable evaluation.
Sep. 2025
- Present

Technology and Engineering Center for Space Utilization, Chinese Academy of Sciences (CSU)

Foundation models for lunar surface understanding and Earth observation.

Publications

ECCV

Thin structuresState-space modelsRemote sensing

Bridging the Geometry Mismatch: Frequency-Aware Anisotropic Serialization for Thin-Structure SSMs

J. Bai, H. Zhang, Q. Wen, N. Li, S. Li, A. ur Rahman, X. Tian

ECCV 2026Main Conference · LNCS 17005, pp. 515–532 · First author

Thin structures are easily fragmented by conventional state-space scans. FGOS-Net introduces geometry-aligned serialization to preserve structural connectivity with efficient inference.

arXivFGOS-Net PDF Publisher ECCV Code
ECCV

Foundation modelsPrompt-free segmentationVisual understanding

IP-SAM: Rethinking Prompt-Conditioned Segmentation for Prompt-Absent Deployment

H. Zhang, J. Bai, R. Guo, J. Tan, H. Wang, Y. Li

ECCV 2026Main Conference · LNCS 17020, pp. 20–37 · Second author

Prompt-conditioned segmenters fail when no user prompt is available; IP-SAM generates intrinsic prompts for automatic deployment across camouflage and polyp segmentation.

arXivIP-SAM PDF Publisher ECCV Code
ICLR

EO agentsReinforcement learningSkill evolution

TerraWeave: Grounding Policy-Skill Co-Evolution through Verified Execution for Earth Observation Agents

J. Bai et al.

ICLR 2027Under review · First author

Successful tool calls can still leave later steps without the evidence or resources they need. TerraWeave learns policies and reusable skills from verified execution state, making credit assignment more reliable across long-horizon Earth-observation workflows.

Figure ICLR 2027
ICLR

Agent evaluationRubric evaluationScientific benchmark

EarthArena: Benchmarking Agents on Real-World, Long-Horizon Scientific Tasks in Earth Observation

J. Bai et al.

ICLR 2027Under review · Co-first author

Answer-only scores miss scientific validity; EarthArena evaluates 30 paper-grounded investigations and 101 linked tasks using reproduced references and rubrics over answers, artifacts, and execution.

Figure ICLR 2027

Additional work: Fluxamba · M²F-SAM

Fluxamba: Topology-Aware Anisotropic State Space Models for Geological Lineament Segmentation in Multi-Source Remote Sensing

J. Bai, H. Zhang, Q. Wen, S. Li, X. Tian, A. ur Rahman

arXiv 2026Under review · First author

Axis-aligned scans break discontinuous geological lineaments; Fluxamba follows their intrinsic topology across multi-source remote sensing.

arXivFluxamba PDF
M²F-SAM: Enhancing Foundation Models via Macro-Micro Fusion for High-Fidelity Lunar Lineament Segmentation

J. Bai, Q. Wen, S. Li, X. Tian, A. ur Rahman, Z. Ergashovna, J. Gerts, A. Krivtsun, N. Ruzi

CSSA 20254th China Space Science Assembly · Beijing · Nov. 21-24 · p. 284

Lunar lineaments demand both global context and boundary detail; M²F-SAM fuses macro and micro features for high-fidelity segmentation.

Research Projects

RSMeM

Core Contributor

Knowledge-enhanced remote sensing agents

Second Prize · Open Source GeoAI Practice Challenge · IEEE APGRSS 2026

RSMeM plans and executes multi-step geospatial analyses with hierarchical domain knowledge, then turns failed attempts into reusable experience. The public demo exposes tools, traces, and outputs for reproducible remote sensing agents.

Associated research: ACL 2026 Main Conference

Award Live Demo ACL Paper Code

Education

Sep. 2025
- Present

University of Chinese Academy of Sciences (UCAS)

M.Eng. student in Computer Technology, School of Space Exploration. 2026 National Scholarship.
Sep. 2021
- Jul. 2025

Minzu University of China (MUC)

B.Sc., Information and Computational Science, College of Science.

Acknowledgements

I have been fortunate to meet fellow researchers whose talent is matched by their enthusiasm: Zonghao Guo, Bingxian Wu, Huiyao Zhang, and Zhounan Yu. I hope the road ahead brings many more familiar faces into this shared journey.