Today’s Open Source (2025-10-22): EditScore Released — 7B–72B Parameter Coverage for Accurate Instruction-Guided Image Editing Quality Evaluation
Daily Discovery of Latest LLMs — 2025-10-22
Location: Hong Kong, China
---
📢 Overview
Today’s highlights include:
- EditScore (Reward Model)
- HumanSense (Comprehensive Benchmark)
- CamCloneMaster (Framework)
- AttnRL (Reinforcement Learning Project)
- Reasoning with Sampling (PyTorch Implementation)
- RewardMap (Toolbox)
---
🏆 Foundation Models
1. EditScore
Description:
A series of state-of-the-art open-source reward models (7B–72B) for evaluating and enhancing instruction-driven image editing.
Key Features:
- Introduces EditReward-Bench — first public benchmark for image editing reward models.
- Covers 13 sub-tasks and 11 cutting-edge editing models.
- Simple API for accurate quality scoring in just a few lines of code.
- Serves as a high-fidelity reward signal for reinforcement learning fine-tuning.
Bookmark:
https://sota.jiqizhixin.com/project/editscore
---
2. HumanSense
Description:
A benchmark for human-centric perception and interaction in multimodal LLMs.
Key Features:
- Focus on deep understanding and contextual responses in extended multimodal settings.
- Employs multi-stage, modality-progressive reinforcement learning.
- Tailored prompt design boosts non-reasoning models without additional training.
Bookmark:
https://sota.jiqizhixin.com/project/humansense2
---
🛠️ Frameworks & Essential Tools
1. CamCloneMaster
Description:
Framework for replicating camera motion from reference videos without camera parameters or test-time fine-tuning.
Key Features:
- Supports reference camera control for I2V (image-to-video) and V2V (video-to-video).
- Includes CameraClone dataset rendered with Unreal Engine 5.
Bookmark:
https://sota.jiqizhixin.com/project/camclonemaster
---
2. AttnRL
Description:
Reinforcement learning framework targeting process supervision in reasoning models.
Key Features:
- Integrates attention mechanisms as exploration guides.
- Optimized for mathematical reasoning tasks.
- Provides complete codebase and training dataset.
- Supports scalable, multi-GPU training and evaluation.
Bookmark:
https://sota.jiqizhixin.com/project/attnrl
---
3. Reasoning with Sampling
Description:
PyTorch-based implementation to boost base model reasoning via sampling.
Key Features:
- Includes diverse evaluations: MATH500, HumanEval, GPQA, AlpacaEval 2.0.
- Helps assess reasoning performance across varied tasks.
Bookmark:
https://sota.jiqizhixin.com/project/reasoningwithsampling
---
💡 Monetization & Publishing Tip
If you’re exploring advanced AI models like EditScore or HumanSense and want to publish findings or tutorials, consider using AiToEarn官网:
- Cross-platform publishing to Douyin, Kwai, WeChat, Bilibili, Xiaohongshu, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, X/Twitter.
- Integrated tools for generation, analytics, content ranking.
- Efficient sharing & monetization of AI benchmarking insights.
---
🤖 Robotics Development
1. Dexbotic
Description:
Open-source Vision-Language-Action (VLA) toolkit for embodied intelligence professionals.
Key Features:
- Supports multiple mainstream VLA strategies in one environment.
- Includes pre-trained models for reproducing state-of-the-art methods.
- Continuous updates to include new foundational and industry-leading VLA models.
Bookmark:
https://sota.jiqizhixin.com/project/dexbotic
---
📄 Related Links
---
Pro Tip for Robotics Professionals:
Integrating Dexbotic with AI-powered content workflows enables:
- Automated generation of demos/tutorials
- Global multi-platform publishing
- Access to performance analytics and AI model rankings (AI模型排名)
This strategy expands your reach and efficiently shares your technical achievements with a worldwide audience.
