Research

See the shape of a research field on one map

Map a field's themes and sub-fields, ask about the key papers and methods under each, and find the areas with the least work.

Explore the full survey

A survey of LLM alignment research, zoomed out to show the whole field. Zoom in on any theme to read the synthesis.

Open the full map

How it works

Enter a topic and the model maps its main themes. Each theme gets its own branch, and every question under it keeps that sub-field in context.

  1. Start with the topic. Type your research question. Ask for the main themes, sub-fields and methods.

  2. Branch into each sub-field. Under a theme, ask for its key papers, the methods they use, and where they disagree. Each answer can take its own follow-ups.

  3. Let auto-expand fill in a branch. Switch a node to an expand node and set a depth. The model writes its own follow-up questions under it and answers them, level by level.

  4. Find the gaps. Branches with few papers or many open questions are candidates for your own work.

The model answers from what it learned in training, or from the web when it searches. Check every paper it names against a library database before you cite it.

Why a map instead of a chat thread

In one chat thread, asking about a second sub-field pushes the first out of view. On a map, each sub-field keeps its own branch, and you can return to any of them later.

Zoom out and each node shortens to a phrase, then a keyword, so the whole field fits on one screen. Zoom in on one theme for the full answers.

Example

Say you are surveying AI in drug discovery. The map splits into molecular generation, property prediction and synthesis planning. Under molecular generation you branch into diffusion models, VAEs and autoregressive approaches. Synthesis planning comes back with far fewer papers than generation, which marks it as a gap to look into.

Now try it yourself

Survey recent papers on large language model alignment
The field clusters into three areas: RLHF and preference learning, constitutional AI and rule-based approaches, and mechanistic interpretability as a path to alignment.
What are the open problems in RLHF?
Reward hacking, where the model exploits the reward signal rather than learning the intended behavior. Also distribution shift — the model encounters inputs at deployment that differ from training.

Ready to try literature surveys?