Unifying Questions¶

Use this template when reading a new method:

1. Prediction target¶

Does it predict tokens, patches, latents, video frames, future states, actions, or experiment outcomes?

2. Conditioning state¶

Is the input text, an image, a video, a history of trajectories, tool output, or an interactive environment?

3. Feedback source¶

Is error defined by next-token loss, human preference, a rule-based verifier, a simulator, or feedback from the real world?

4. Where improvement happens¶

Does capability growth come from parameters, data, prompts, test-time search, or external tools and environments?

5. Evidence level¶

Record “reported in the paper,” “reproduced with public code,” “independently reproduced,” and “observed in deployment” separately.