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

Snorkel and UC Berkeley’s RLLM team show a 4B-parameter model, trained with inexpensive RL on a curated tool-use environment (FinQA), outperforming a 235B model on financial analysis tasks. The win wasn’t more "reasoning"—it was disciplined tool use: listing tables, inspecting schemas, and self-correcting errors before answering.

Why this matters: production agent workflows often fail from poor tool discipline, not lack of intelligence. Smaller, self-hosted models with behavior training can hit reliability targets while lowering cost, latency, and compliance risk. Their RL loop (GRPO) ran ~21 hours for under $500 and doubled pass@1 accuracy on single-table tasks; surprisingly, training only on single-table prompts also lifted multi-table performance.

What’s strong: a concrete, operationally relevant result—cheap, targeted RL can fix the real failure mode (tool misuse). The frameworked environment (OpenEnv/PrimeIntellect) is reproducible, and the evaluation rubrics approach helps isolate behavioral failures worth training.

Caveats: single domain and environment, small datasets (≈290 single-table, 79 multi-table), and an unspecified 235B baseline/practices reduce generality. Potential overfitting risk to the sandbox. Claims shouldn’t be generalized to open-ended reasoning or broader domains without more evidence.

What to do: instrument tool telemetry, adopt rubric-based evals to diagnose failure modes, and budget for small RL runs before defaulting to larger models. Keep architecture modular: default to compact, on-prem models; route edge cases to bigger models if needed.