Stop Making Models Bigger, Make Them Behave — Kobie Crawdord, Snorkel

A conference talk from Snorkel's developer advocate presenting research showing a 4-billion parameter model can outperform a 235-billion parameter model on financial analysis tool-use tasks. Using GRPO-based reinforcement learning with a high-quality expert-curated dataset, the team achieved roughly double the pass@1 performance at under $500 per training run. The key insight: the larger model failed not due to reasoning deficits but due to poor tool-use discipline — it queried non-existent tables and hallucinated answers. The fine-tuned smaller model learned to first discover available tables, inspect schemas, and self-correct errors. Surprisingly, training on single-table examples only yielded the best results, and the improvement generalized to harder multi-table benchmarks as well. The takeaway is that identifying the specific behavioral failure mode and targeting it with the right data is often more effective than scaling up model size.