DART-SD Stops Distillation From Flattening Agent Diversity
When you distill a multi-turn tool-calling agent, the standard recipe is full-trajectory supervision: take a successful run, train the student to reproduce it end to end. DART-SD's authors point out what that quietly destroys: most real tool-use tasks are diamonds, not lines. Many valid paths diverge after a shared start and reconverge at the goal, and training on one canonical trajectory teaches the model that every other valid path is wrong. Paper: arXiv 2608.18524, currently at 60 upvotes on Hugging Face daily papers.
Their fix models execution as an Interaction-State Transition Graph, finds the exact states where the student's exploration actually fails, and applies supervision only to those recovery steps, leaving the earlier reasoning untouched by gradients. Surgical correction instead of full-body cast. On multi-turn tool-calling benchmarks it substantially beats full-trajectory self-distillation.
The lesson generalizes past distillation: agent training keeps importing habits from single-shot instruction tuning, and "there is one right answer" is the most toxic import. An agent's strength is having five ways around a failed API call. Training methods that preserve path diversity instead of collapsing it are how small models get real agentic ability rather than a memorized script.
← Back to all articles
Their fix models execution as an Interaction-State Transition Graph, finds the exact states where the student's exploration actually fails, and applies supervision only to those recovery steps, leaving the earlier reasoning untouched by gradients. Surgical correction instead of full-body cast. On multi-turn tool-calling benchmarks it substantially beats full-trajectory self-distillation.
The lesson generalizes past distillation: agent training keeps importing habits from single-shot instruction tuning, and "there is one right answer" is the most toxic import. An agent's strength is having five ways around a failed API call. Training methods that preserve path diversity instead of collapsing it are how small models get real agentic ability rather than a memorized script.
Comments