Research
Fine-tuning 350M Model for Structured Outputs
Hugging Face Bloghigh confidence
Why it matters
- →Demonstrates that smaller models can be fine-tuned to match larger models in specific tasks.
- →Provides a practical approach for developers to enhance model performance using accessible resources.
- →Highlights the importance of structured output capabilities for real-world AI applications.
Hugging Face has outlined a method to fine-tune a 350M model to improve its structured output capabilities, a key requirement for many applications. The process involves using a LoRA adapter and specific reward functions to enhance the model's performance on the IFStruct benchmark, achieving a 22.6% pass rate. This demonstrates that smaller models can be effectively optimized for specific tasks, potentially matching the performance of larger models. The fine-tuning can be conducted on a free-tier GPU, making it accessible for developers seeking to improve model performance efficiently.
Read originalMore from Hugging Face Blog
© Hugging Face BlogModels & Labsmodels
NeoMME: Efficient Multimodal and Multilingual Encoder Released
Hugging Face has unveiled NeoMME, a new family of multilingual multimodal encoders that integrate text and image processing into a single bidirectional Transformer. Unlike traditional models that rely on separate pretrained vision towers, NeoMME processes both text tokens and image patches simultaneously, enhancing efficiency and reducing computational overhead. The model's innovative architecture allows for dynamic image resolution and long bidirectional context, making it highly adaptable for various applications. With its release, NeoMME offers a streamlined approach to visual document retrieval, outperforming existing models in efficiency and storage requirements.
Hugging Face Blog
© Hugging Face BlogVideo & Creative AIimage
AI Model Trained to Paint Watercolours
A language model has been trained to paint watercolours using JavaScript and the p5.brush library, capturing the attention of over 1.5 million viewers. The project uses reinforcement learning to teach the model aesthetic preferences, allowing it to create unique, handmade-style paintings. This approach contrasts with the typical perfection of AI-generated images, offering a fresh perspective on AI's creative potential. The open-source nature of the project invites further exploration and replication by others interested in the intersection of AI and art. The model's ability to generate code that paints images reflects a deeper level of interaction between AI and artistic expression. By limiting the model to ten methods from the library, the project emphasizes creativity within constraints. This initiative not only highlights AI's evolving role in art but also encourages a broader dialogue about the future of AI-driven creativity.
Hugging Face Blog
© Hugging Face BlogCoding Toolscoding
Hugging Face Introduces Funes for Agent Memory
Hugging Face has launched Funes, a tool that provides a durable memory layer for coding agents like Claude Code and Codex. This innovation allows agents to index, retrieve, and rank session logs, enabling them to recall past decisions and rationales during their workflow. Funes operates locally, ensuring privacy and control, and can also sync with a private Hugging Face dataset for broader accessibility. This development enhances the functionality of coding agents by allowing them to seamlessly integrate past knowledge into current tasks, making them more efficient and context-aware.
Hugging Face Blog
More in Research
© Google Research BlogResearchresearch
Google Maps Complete Male Fruit Fly Brain
Google Research, in collaboration with HHMI Janelia, has achieved a significant milestone in connectomics by mapping the complete brain and central nervous system of the male fruit fly. This project, published in Cell, represents the largest brain map to date with over 166,000 neurons and 125 million synaptic connections. The detailed connectome provides a crucial resource for studying neural mechanisms and behaviors, offering insights into how brains function across species. This advancement not only enhances our understanding of fruit fly neuroscience but also sets the stage for future research in more complex organisms.
Google Research Blog
© TechCrunch AIResearchresearch
OpenAI's Astra Model Raises AI Safety Concerns
OpenAI's Astra model introduces a new reasoning technique known as 'recurrent depth,' which has sparked significant concern among AI safety experts. This approach, also referred to as 'opaque recurrence,' allows the model to process queries in a loop, making its reasoning process less transparent and more challenging to monitor. Despite OpenAI's assurances that Astra's use of this technique is limited and that they remain committed to chain-of-thought monitoring, experts worry about the potential for diminished transparency in AI reasoning. The situation underscores the ongoing tension between advancing AI capabilities and ensuring safety and accountability in AI systems.
TechCrunch AI
© WIRED AIResearchresearch
Russian Startup Enables AI Models to Communicate Silently
Mostik, a Russian startup, has developed a novel approach allowing AI models to communicate without generating text output, akin to machine telepathy. This technique leverages the mathematical values in model weights to enable smaller models to benefit from the capabilities of larger ones, enhancing efficiency and performance. By creating a bridge between models like GLM-5.2 and Qwen-3.5, Mostik has demonstrated a cost-effective hybrid system that performs impressively. This innovation could significantly boost the value of open-weight models, challenging the dominance of proprietary models from major labs like OpenAI.
WIRED AI