Snorkel at AI Council

Everyone knows bigger models reason better. What’s less obvious is that they often behave worse, especially when tools are involved. In this talk showing a practical application of research around task fidelity, we’ll show how a 4B model was fine-tuned to beat a 235B model on real financial analysis tasks, not by adding more reasoning, but by enforcing tool discipline. Using reinforcement learning with the open-source rLLM framework, the model learned to explore schemas, validate outputs, and retry failures instead of hallucinating confident nonsense.

The key surprise: training on simple tool interactions transferred cleanly to much harder, multi-step problems. If you’re building LLM systems that touch things like databases, APIs, or internal tools, this talk focuses on the behaviors that actually matter and how to teach them without frontier-scale compute.