AI

Compiling Expert Procedures into Verified Programs: When a Small Model Is Enough

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 16/10/2026, 02:33:24 UTC.

Briefings

A new program named PLVR compiles expert reasoning traces into executable code for small 8B parameter models. On the Telecom domain, Granite-4.2-8B improved from 43.0% to 91.2% accuracy, while on Retail and LiveMCPBench the program matched best baselines. The spread between base models narrowed from 7 to 2 points, showing the program effectively bridges capability gaps in procedural tasks.

Key points

• The program PLVR takes teacher reasoning traces, clusters recurring operations, and generates a program with one typed primitive per operation using a frozen small model.
• On Telecom, Granite-4.2-8B improved from 43.0% to 91.2% accuracy, while on Retail and LiveMCPBench the program matched best baselines.

Context

This research addresses the gap between base model capabilities and expert procedures in tool-use domains. The PLVR program demonstrates that procedural execution can significantly boost performance on procedural tasks like tool search and multi-turn matching.

Limitations

Results are limited to three Nemotron base models on official tau2-Bench and LiveMCPBench test sets. Performance varies by domain, with Granite-4.2-8B showing strong gains on Telecom but varying results on Retail and LiveMCPBench.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 25/10/2026, 17:36:30 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeidoer3cdibb27kuui4e7iarnu36525rzkv5sx4vf25s2gm3zfqtom