Auto Research

简体中文 · English

Auto Research is a local-first platform for inspecting paper implementations and running reproducible, iterative experiments. Start by checking what a paper establishes and which parts are implemented locally. Then use compatible components in a controlled research loop.

Browse the paper catalog Install and run a demo

Two workflows

Paper implementation and evaluation

Start with the paper catalog. Entries link to original sources and identify the available write-up language. A write-up records the method, authors' results, local commands or evaluation artifacts, and where the implementation differs from the paper.

Domain	What to look for
Search, advertising, recommendation, and LLM applications	Retrieval, ranking, generative recommendation, and production A/B evidence; local public-data experiments are reported separately.
Foundation models	Model architecture, pretraining, long context, and inference efficiency, with the local model size and compute budget made explicit.
Structured decisions / System One	Typed questions, dynamic candidate sets, calibration, selective prediction, and structured scoring.
Multimodal models	Image, video, audio, and language experiments with traceable checkpoints and evaluation subsets.
LLM post-training	Preference learning, reinforcement learning, on-policy distillation, and training stability; candidate scoring is not the same as free generation.
Agents	Memory, planning, tools, and self-improvement. A local mechanism or simulator is not evidence of general agent capability.

Automated research and evolution

Start from a research topic, or supply a supported system, dataset, and research direction. The controller retrieves paper evidence, proposes testable hypotheses, runs registered candidates under a shared budget, and uses validation performance to guide the next round. Test data stays isolated from selection.

Supported examples include RankMixer and HyFormer, a local micro-LLM, post-training objectives, and combinations of agent memory, planner, tool policy, and critic. Newly discovered methods with no executable implementation remain evidence only. The evolution overview covers adapter and workflow details.

flowchart LR
    A[System or topic] --> B[Research direction]
    B --> C[Paper evidence]
    C --> D[Candidate experiments]
    D --> E[Execute in parallel]
    E --> F[Select on validation]
    F --> G{Continue?}
    G -- Yes --> D
    G -- No --> H[Isolated test]
    H --> I[JSON, Markdown, and HTML]

Read the evidence

The public experiment dashboard indexes committed, traceable results. Compare like with like: an original paper's offline benchmark, its online A/B test, and a local public-data run answer different questions. A synthetic mechanism check or a small smoke test does not establish the full paper's capability or production gains. High-cost experiments that were not run are explicitly labeled not reproduced.

For the source repository, see the English README. Use the learning cards and topic deep dives to study mechanisms and compare papers. The translation coverage report covers the whole site.