Research map / Multi-agent systems and games / LLM agents
EpiAgent: An Agent-Centric System for Ancient Inscription Restoration
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes EpiAgent, an LLM-planned Observe–Conceive–Execute–Reevaluate agent that orchestrates MLLM/OCR analysis, experience priors, diffusion-based restoration tools, and multi-metric self-refinement to restore ancient Chinese inscriptions.
Problem
Full-inscription restoration of ancient stone inscriptions and rubbings under heterogeneous, spatially coupled degradation, where fixed image-to-image pipelines distort glyphs and fail to generalize to real rubbings.
Approach
An LLM central planner (Kimi-K2) runs an iterative observe-conceive-execute-reevaluate loop, choosing per-character sequences of specialized diffusion restoration tools based on degradation analysis, mined experience priors, and text/style-based feedback, rather than a fixed single-pass pipeline.
Key results
- On CIRI Testing Set S, EpiAgent reaches PSNR 22.14, SSIM 0.9684, and 1-NED 0.9069, versus 21.15 PSNR for the best baseline IR3.
- On real split R-II, EpiAgent achieves 1-NED 0.5546 versus 0.5374 for IR3 (second best).
- In a user study with over twenty epigraphers, EpiAgent gets 59.66% Top-1 ranking versus 15.60% for IR3.
Surprising findings
- A standalone MLLM for analysis yields very low 1-NED (0.6481 on Set S), while adding the CLM and RAG raises it to 0.9742.
Limitations
- Baselines were retrained on CIRI with default settings; some tailored and agentic methods could not be reproduced or compared.
- Human feedback is optional and expert review is needed for high-stakes cases.
- Reliance on an external closed LLM planner (Kimi-K2) and MLLM components.
Experimental setup
Datasets: CIRI (Chinese Inscription Rubbing Images) large-scale Chinese corpus (RAG) historical execution logs
Benchmarks: CIRI Testing Set S CIRI Testing Set R-I CIRI Testing Set R-II
Metrics: PSNR SSIM LPIPS CLIP-IQA MUSIQ MANIQA NIMA Top-1 Accuracy Top-5 Accuracy Macro Accuracy
Baselines: CharFormer DocDiff GSDM Restormer MambaIR PromptIR MoCE-IR IR3
Builds on: PromptIR DocDiff RAG Kimi-K2
Closest papers by meaning
- An Intelligent Agentic System for Complex Image Restoration Problems (ICLR 2025 · similarity 0.84)
- OPERA: An Agent for Image Restoration with End-to-End Joint Planning–Execution Optimization (NeurIPS 2026 · similarity 0.84)
- RestoreAgent: Autonomous Image Restoration Agent via Multimodal Large Language Models (NeurIPS 2024 · similarity 0.82)
- OBI-Bench: Can LMMs Aid in Study of Ancient Script on Oracle Bones? (ICLR 2025 · similarity 0.81)
- E-mem: Multi-Agent Based Episodic Context Reconstruction for LLM Agent Memory (ICML 2026 · similarity 0.81)
- Towards Automated Chinese Ancient Character Restoration: A Diffusion-Based Method with a New Dataset (AAAI 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.