Program Distillation Offers Scalable, Transparent LLM Evaluation

New research proposes PAJAMA, a system that offers a scalable and transparent alternative to the common "LLM-as-a-judge" paradigm for automated evaluation. While LLM judges are standard, they suffer from high costs, significant latency, and opaque decision-making, hindering their scalability and reliability. PAJAMA addresses these issues through a technique called program distillation. Instead of relying on an LLM for real-time evaluation, PAJAMA distills the LLM's decision logic into a committee of executable programs. These programmatic judges are inherently transparent, allowing for easy inspection and editing, and eliminate per-sample API costs. The system aggregates decisions from these programs and includes a fallback mechanism to selectively escalate low-confidence cases to a human or a larger LLM. Across five datasets and four model families, programmatic judges demonstrated performance comparable to a 13B-sized LLM judge. By using program outputs as routing signals, PAJAMA improved both accuracy and throughput, pushing the Pareto frontier for evaluation. Furthermore, these programmatic judges generate cheap and effective reward signals, outperforming reward models trained on proprietary LLM labels at two orders of magnitude lower API cost, making them valuable for both evaluation and reinforcement learning.