Xiaomi unveils CodeMidas coding RL framework

// 1h agoRESEARCH PAPER

Xiaomi unveils CodeMidas coding RL framework

Developed by Xiaomi's MiMo team and Peking University researchers, CodeMidas is an autonomous pipeline that constructs executable reinforcement learning environments directly from raw open-source codebases without relying on external artifacts like GitHub issues. The system synthesized 5,545 verified RL tasks across 3,185 repositories and 23 programming languages, driving significant benchmark performance gains for models like MiMo-V2.5.

// ANALYSIS

Self-synthesized RL environments directly from raw code dismantle the biggest bottleneck holding back frontier coding agents: the shortage of clean, human-curated training data.

• Bypasses the limitations of GitHub issue scraping by deriving specifications and verifiable test oracles directly from repository source code.

• Automated execution filtering and iterative rollout validation eliminate reward hacking and flawed verifiers before training begins.

• Broad linguistic diversity across 23 languages and 15 technical domains exposes agents to realistic, multi-ecosystem software engineering challenges.

• Marks a significant step toward closed-loop agent self-improvement where models autonomously construct and grade their own learning curricula.

// TAGS

codemidasxiaomiagentreinforcement-learningcode-generationsynthetic-datasoftware-engineering

DISCOVERED

1h ago

2026-09-21

PUBLISHED

1h ago

2026-09-21

RELEVANCE

8/ 10

AUTHOR

mark_k

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

OPEN SOURCE14m ago

AutoClip is an open-source, full-stack video processing platform that uses AI to automatically extract highlights and generate short clips from long-form content.

AutoClip is an open-source, self-hosted AI video clipping and highlight generation platform designed to repurpose long-form media—such as podcasts, livestreams, and lectures from YouTube, Bilibili, or local files—into engaging short-form clips. Built on a FastAPI backend, Celery task queue, and React frontend, the application automates video ingestion via yt-dlp, speech transcription via Whisper, and content evaluation using LLMs like Qwen, Gemini, OpenAI, or local models via Ollama and LM Studio. It scores segments for viral potential, cuts video via FFmpeg, manages collections in a modern web interface, and provides developer-oriented integrations including a command-line interface and an MCP (Model Context Protocol) server for agents in Claude and Cursor.

OPEN SOURCE20m ago

PenguinHarness 0.2.13 advances agentic recursive self-improvement

PenguinHarness 0.2.13 is the latest update to the open-source agent development platform created by the team behind LlamaFactory for recursive self-improvement. Supporting over 1,000 models, the platform provides an AI-native workspace where agents autonomously scaffold, evaluate, and evolve other agents with closed-loop evaluations to prevent reward hacking.

MODEL1h ago

Qwen3.8-Omni-Flash drops voice API costs 98%

Alibaba's Qwen team has launched Qwen3.8-Omni-Flash, a native multimodal model built to aggressively lower the cost barrier of omni-modal AI. The model features a 1-million-token context window with native support for text, image, audio, and video inputs, while slashing voice API costs by more than 98% to make real-time multimodal intelligence widely accessible for developers.