Polaris: An Open‑Source AI Research Agent that Partners with You

Research workflows are long and fragmented: reading countless papers, unclear directions, repeated experiments, and late‑night writing. Existing large‑model tools handle isolated steps, but a full‑cycle AI‑driven research process remains unsolved.

Polaris (named after the North Star) addresses this gap by chaining six stages—literature survey, idea generation, idea review, experiment, paper writing, and paper review—into a single pipeline where the output of each stage naturally feeds the next. The AI drives the process, while humans intervene only at key decision points.

Each morning the platform automatically fetches the latest arXiv submissions for the user’s subscribed topics. An AI first reads each paper, producing a one‑sentence conclusion, a one‑page Chinese summary, and organized metadata (authors, concepts, tags). Researchers can browse, filter, and add selected papers to a shared knowledge base with a single click.

Instead of merely storing papers, Polaris compiles each into a deep‑read page that explains the work’s motivation, method, and reusable aspects, extracts figures with captions, and links concepts across the corpus. The knowledge base supports semantic search, relevance scoring, custom tags, and one‑click export to tools like Obsidian. Important papers can be turned into shareable PPTs.

Idea generation leverages signals such as concept gaps, paper limitations, and research trends. Polaris scores candidate ideas on novelty and feasibility, then routes them to a panel of AI reviewers with differing viewpoints. The reviewers debate each idea, producing an Elo ranking; the highest‑ranked ideas are presented to the researcher for promotion to a full research plan or direct experimentation.

When an idea is approved, Polaris connects to the lab’s GPU servers. An experiment‑intelligence loop plans the experiment, breaks it into a checklist with acceptance criteria, writes code, deploys the environment, and launches training. After each step the system validates exit codes, artifacts, and metrics before proceeding. If results stagnate, the AI revises the plan, adds new steps, and iterates. At decision points (e.g., ambiguous data differences) the AI pauses and asks the researcher, whose answer is incorporated into the next iteration.

For paper writing, Polaris offers an online LaTeX workspace with multi‑file projects, live compilation, and PDF preview. The AI drafts each section, but enforces a hard rule: every numeric claim must trace back to a real experiment record, and every citation must point to an existing reference. After drafting, AI‑driven peer review checks citation authenticity, cross‑verifies numbers with experiment logs, and aggregates reviewer comments. Any fabricated citation causes an immediate rejection.

Polaris also includes a global assistant, PolarisBuddy, accessible via Command J. It greets the user, offers context‑aware guidance, accepts dragged‑in papers or text for queries, and can autonomously plan and execute tasks until a result is produced.

The platform is built for multi‑user labs: a shared literature library is parsed once for the whole team, usage metrics (AI token consumption, GPU budget) are visible on a lab dashboard, and all members can see each other’s progress.

Polaris’s capabilities are exposed via MCP and a skill system. Built‑in skills cover literature summarization, experiment design standards, and automated rebuttal drafting, while users can add custom skills or install community‑shared ones. External tools such as Claude Code, Codex, and DeepSeek Harness can call Polaris resources directly, and desktop clients are available for macOS, Windows, and Linux.

The project is open‑source on GitHub (https://github.com/ZJU-REAL/Polari) and remains under rapid development. Researchers are invited to try the platform and contribute to its evolution.