touch-grass

· [wip] · ai · fine-tuning · decision-models · python
github.com/heath0xFF/touch-grass →

agents can get stuck in loops. they repeat failed tool calls, retry without new evidence, or turn a small task into a much larger one.

touch-grass is an experiment in teaching a small model to recognize those situations. it’s a fine-tune of Liquid AI’s d1-3B decision model to score five actions:

  • CONTINUE: useful progress is happening.
  • RETRY: repeating the action is justified.
  • CHANGE_STRATEGY: the task is still solvable, but the approach needs to change.
  • ESCALATE: a person, permission, or missing fact is needed.
  • TOUCH_GRASS: more autonomous work is likely to add cost, risk, or unnecessary scope.

the model judges the task and the agent’s trajectory. it does not execute the chosen action.

building and measuring it

i trained the first version with LoRA on my DGX Spark. the repository includes synthetic training data, evaluation code, measured results, and reproduction instructions.

the tuned model reached 90% accuracy on a 150-case synthetic test set and a separate 30-case authored set. It also incorrectly stopped useful work in one of those authored cases. These small tests show a proof of concept, not production safety.

the released weights are available alongside GGUF and MLX exports. they use native decision scoring, not an ordinary chat interface.

what I am exploring next

Recent work adds challenge cases that distinguish productive repetition from a genuine loop. i am also exploring whether targeted weight edits could correct specific decision errors without damaging other behavior. that is an open experiment, not a demonstrated result.

Liquid AI’s Antidoom inspired the project, but touch-grass is not an implementation of Antidoom or its training method. It works at the agent-trajectory level rather than the token level, and remains advisory.

i’d really also like to explore weight-space programming with this. is there a scenario where LoRA + weight space programming can make a meaningful impact over “plain” or “default” SFT?