Open Project & Benchmarks

Agent-R1 is a unified, modular training framework for Agentic RL. It models each round of agent interaction as a step-level RL transition and decouples trajectory representation, context construction, environment interfaces, and optimization algorithms. This allows GRPO, PPO, and other algorithms to be compared and extended on a shared foundation for multi-turn agent training.

Claw-R1 provides the Data Foundation for Agentic RL, enabling the systematic collection, evaluation, and curation of high-quality training data from diverse agent interactions. It introduces a middleware layer (Gateway + DataPool) between the Agent Side and the Training Side, focusing on data infrastructure rather than training algorithms themselves.

CastClaw（星辰） is an agent framework for human–AI collaboration in time series forecasting research. It uses three specialized agents, Planner, Forecaster, and Critic, to orchestrate the complete forecasting workflow and incorporates human confirmation at key stages. It packages data analysis, feature engineering, and classical time series model capabilities into an extensible runtime toolbox.

CastMind（星思） is a research initiative on context-driven reasoning for time series projection in complex systems. It explores how historical time series, external context, and domain knowledge can support an extrapolate → understand → project → revise workflow for forecast revision and mechanism explanation. Public code is forthcoming.

CastFactory（炼星坊） is an open-source framework for LLM-driven time series forecasting model training, designed to make it extremely easy for users to build and adapt TSF models with modern large-model pipelines. It provides a unified and practical workflow for continued pre-training (CPT), supervised fine-tuning (SFT), and reinforcement learning (RL), helping researchers and practitioners develop, optimize, and evaluate LLM-based forecasting systems with much lower engineering overhead.

PaperScout is an autonomous scientific literature retrieval agent for complex research queries. It treats paper discovery as multi-turn decision-making, using accumulated retrieval context to choose between Search for new candidates and Expand for following paper references. With process-aware, sequence-level reinforcement learning (PSPO), it learns to refine its retrieval strategy from intermediate feedback and identify papers that match the research topic and query constraints.

Academic Search is a scientific literature retrieval and metadata skill for Codex, Claude Code, and other AI agents. It brings together sources including arXiv, Semantic Scholar, OpenAlex, Google Scholar, and CNKI, supporting query expansion, citation tracing, record validation, cross-source deduplication, and BibTeX export, as well as open-access full-text retrieval. It adapts search strategies to the discipline and records source provenance and access status to support literature reviews and research surveys.