Skip to content

docs(program): add hypothesis column to results.tsv format#336

Open
morozow wants to merge 1 commit intokarpathy:masterfrom
morozow:exp/program/mar19-hypothesis
Open

docs(program): add hypothesis column to results.tsv format#336
morozow wants to merge 1 commit intokarpathy:masterfrom
morozow:exp/program/mar19-hypothesis

Conversation

@morozow
Copy link

@morozow morozow commented Mar 19, 2026

Add hypothesis column to results.tsv for learning

This PR introduces a lightweight mechanism for the agent to accumulate non-obvious research directions during experimentation.

Changes:

  • Added hypothesis column to results.tsv format
  • Updated program.md: column spec and examples for agent
  • Added "gather hypotheses from results.tsv and look for unexplored combinations" to the "out of ideas" recovery flow

What is a hypothesis here:

Not a restatement of what was tried (that's description), but a non-obvious combination, condition, or connection worth exploring later. Example: experiment fails with GeLU → hypothesis is "GeLU might unlock with RMSNorm instead of LayerNorm", not "GeLU didn't work".

Why this matters:

When the agent exhausts obvious directions after 50+ experiments, it can revisit its own accumulated hypotheses — ideas it generated in context but never tested. This creates a passive memory that survives context window limits.


Open question – needs A/B testing:

The hypothesis gathering instruction is currently placed at the beginning of the "out of ideas" list. Unclear if it should be first (prioritize own insights) or last (use as final fallback). Requires empirical testing.

- Expand results.tsv from 5 to 6 columns, adding new "hypothesis" column
- Update TSV header and column descriptions to include hypothesis field
- Define hypothesis as non-obvious ideas to explore based on results
- Add guidance to use "-" for baseline runs without hypotheses
- Update example table with sample hypothesis values for each experiment
- Clarify autonomous researcher guidance to gather hypotheses from results.tsv when generating new experiment ideas
- Helps structure iterative research by capturing insights and unexplored combinations for future testing
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant