Gate News message, April 22 — Hugging Face has open-sourced ml-intern, an ML research agent capable of autonomously completing the full workflow of reading papers, organizing datasets, launching GPU training, evaluating results, and iterating improvements. The project is built on Hugging Face’s smolagents framework and provides both CLI and web-based interfaces, with code available on GitHub.
The ml-intern toolchain is designed around the Hugging Face ecosystem. It retrieves papers from arXiv and HF Papers while tracing citation chains for deeper reading; browses datasets on HF Hub, validates quality, and reformats data for training; and when local GPU resources are unavailable, invokes HF Jobs to launch cloud-based training tasks. After training completes, the agent automatically reads evaluation outputs, diagnoses failure causes, and reruns experiments. By default, it uses Claude Sonnet 4.5 to drive the decision loop, with a maximum of 300 iterations per run and automatic context compression when exceeding 170k tokens.
Hugging Face demonstrated three use cases. In a scientific reasoning task, the agent identified OpenScience and NemoTron-CrossThink datasets from citation chains, filtered seven variants from ARC, SciQ, and MMLU by difficulty level, and ran 12 rounds of supervised fine-tuning on Qwen3-1.7B, improving GPQA scores from 10% to 32% in under 10 hours. For a medical application, the agent determined existing datasets were insufficient, wrote scripts to generate 1,100 synthetic data samples, and scaled them 50-fold for training, exceeding Codex performance by 60% on HealthBench. In a competitive mathematics scenario, the agent authored a GRPO training script and launched training on A100 GPUs via HF Spaces, then conducted ablation studies after observing reward collapse.
Disclaimer: The information on this page may come from third parties and does not represent the views or opinions of Gate. The content displayed on this page is for reference only and does not constitute any financial, investment, or legal advice. Gate does not guarantee the accuracy or completeness of the information and shall not be liable for any losses arising from the use of this information. Virtual asset investments carry high risks and are subject to significant price volatility. You may lose all of your invested principal. Please fully understand the relevant risks and make prudent decisions based on your own financial situation and risk tolerance. For details, please refer to
Disclaimer.
Related Articles
Microsoft Unveils AI Agent Commerce Infrastructure: Publisher Marketplace, Merchant Protocols, and Ad Tools
Gate News message, April 22 — Microsoft's AI monetization vice president Tim Frank announced a suite of commercial infrastructure updates designed for the "agentic web" era, enabling publishers, merchants, and advertisers to remain discoverable and tradable as AI agents make purchasing decisions on
GateNews42m ago
NeoCognition Raises $40M in Seed Funding for On-the-Job Learning AI Agents
Gate News message, April 22 — AI research lab NeoCognition announced the completion of a $40 million seed round, emerging from stealth mode. Founded by Ohio State University Associate Professor Yu Su, along with Xiang Deng and Yu Gu, the company is headquartered in Palo Alto, California. The round w
GateNews58m ago
PicWe Launches AI Agent Wallet with On-Device Key Management
PicWe announces public beta of PicWe Wallet, an AI-agent-enabled, on-device key wallet with no recovery phrases. It supports multi-chain assets, swaps, AI-accessible automation, and aims to unify RWA infrastructure.
PicWe has launched the public beta of PicWe Wallet, an AI Agent-enabled wallet that stores keys on-device, eliminates recovery phrases, and keeps critical operations local. The beta supports multi-chain asset management, swaps, and stablecoin-based fees while enabling programmable AI interactions. Broader PicWe initiatives position the platform as unified infrastructure for real-world assets, enabling issuance, circulation, settlement, cross-border payments, tokenization, and supply-chain coordination for enterprise use cases.
GateNews1h ago
Google Research Releases ReasoningBank: AI Agents Learn Reasoning Strategies from Success and Failure
Gate News message, April 22 — Google Research released ReasoningBank, an agent memory framework that enables large language model-driven agents to continuously learn after deployment. The framework extracts universal reasoning strategies from both successful and failed task experiences, storing
GateNews2h ago
Tsinghua Professor Dai Jifeng Launches Naive.ai, Raises ~$300M at $800M Valuation
Gate News message, April 22 — Dai Jifeng, an associate professor at Tsinghua University's Department of Electronic Engineering, has founded Naive.ai, a company focused on open-source model post-training and AI agents. The startup has raised approximately $300 million at an estimated valuation of $80
GateNews3h ago
AWS Expands Multi-Agent AI Workflows, Supports Claude Opus 4.7 on Bedrock
Gate News message, April 22 — Amazon Web Services announced expansion of its agentic AI initiatives through multi-agent workflows, supporting Anthropic's Claude Opus 4.7 on Amazon Bedrock to help customers move beyond generative AI pilots. The company is expanding partner relationships as customers
GateNews3h ago
