SuperAPI AI Directory
Partner with us中文AI newsFavorites

AI developer platforms

3,662 results · Page 37 of 102

R1-Multimodal-Journey

AI developer platforms

R1-Multimodal-Journey (also known as MM-Eureka V0) is an open-source project exploring reinforcement learning (RL) for multimodal large models, aiming to apply R1-style reinforcement learning similar to DeepSeek-R1 to vision-language models (VLMs) and verify its effectiveness on more challenging problems such as geometry. The project is built on Open-R1-Multimodal, integrates open-source achievements such as vLLM, Open-R1, and trl, and adopts the technologies and base models of DeepSeek-R1 and Qwen2.5-VL. Its core improvements include: support for vLLM-accelerated training, which is 5-6 times faster than OpenR1-MM on 8 H100s; fixing some bugs in Open-R1-Multimodal (such as the get_per_token_logps function); and upgrading transformers to 4.49.0.dev0 to support Qwen2.5_VL. The experiment used 20,000 samples from geo170k and found that even a small 3B model under the instruct setting can exhibit an "aha moment" within the first 10 training steps. The project also found that VLMs struggle to show a length growth curve like LLMs, but performance still improves; compared with SFT, R1-style RL is based only on answer correctness rewards, yet exhibits higher data efficiency. The project's latest progress has been migrated to MM-Eureka, and more capabilities, models, and data are also open-sourced in that repository.

Read more

RAGEN

AI developer platforms

RAGEN is an open-source reinforcement learning framework for training reasoning LLM agents, based on the StarPO algorithm and supporting multi-turn trajectory-level reinforcement learning. It provides 10 built-in environments (such as Sokoban, FrozenLake, WebShop, DeepCoder, SearchQA, etc.) and a Gym-compatible interface, making it easy to add custom environments. The RAGEN-2 version further introduces SNR-Adaptive Filtering, a lightweight trajectory filtering method based on reward variance, as well as reasoning collapse diagnostics (detecting template collapse through conditional entropy and mutual information proxy metrics), helping to understand and fix hidden failure modes in agent reinforcement learning (such as Echo Trap and reasoning collapse). The framework supports two optimization methods, PPO and GRPO, and includes tools such as a diagnostics dashboard and training metrics guide.

Read more

RAGEN

AI developer platforms

RAGEN is a reinforcement learning framework for training reasoning LLM agents, supporting multi-turn reinforcement learning based on the StarPO algorithm. It provides trajectory-level and turn-level training, 10 built-in environments, and a Gym-compatible interface, making it easy to add custom environments. RAGEN-2 further introduces SNR adaptive filtering based on reward variance and reasoning collapse diagnostic metrics to mitigate issues such as template collapse and improve training stability.

Read more

RD-Agent

AI developer platforms

RD-Agent is a R&D automation agent launched by Microsoft, aiming to automate high-value general R&D processes through AI-driven data-driven AI. Its core framework consists of two parts: "Research" (R, proposing new ideas) and "Development" (D, implementing ideas). Typical applications include automatic quantitative factory, data mining agent, research copilot (automatically reading papers or financial reports and implementing models/datasets), Kaggle agent, FFT agent, and data science agent. RD-Agent currently only supports Linux systems, depends on Docker and Conda environments, can be installed via PyPI, and supports LiteLLM integration with various large models (including DeepSeek).

Read more

reactive-agents-ts

AI developer platforms

Reactive Agents is an open-source TypeScript agent execution framework, designed for controllable execution, end-to-end observability, and reliable operation, where the same code can run across local 4B models and frontier APIs. It provides a transparent 12-stage execution engine, per-stage hooks, tool call repair, output validation, crash recovery, and a per-run signed evidence receipt. The framework supports 8 reasoning strategies, 4 layers of memory, 8 LLM providers, model-adaptive context, guardrails and validation, Ed25519 identity, and React/Vue/Svelte frontend integration, built on Effect-TS.

Read more