Researchers have released a technique named PLVR that turns a teacher's step‑by‑step traces into a reusable program. The program runs alongside a frozen small language model – even an 8 billion‑parameter model – and guides its actions. On the telecom test set, the 8B Granite‑4.2 model jumps from 43% to 91.2% accuracy when used with PLVR.
How PLVR works
First, the teacher model Kimi‑K3 records its reasoning traces while solving tool‑use problems. PLVR clusters these traces to find repeating operations. For each operation, a “frontier” model writes a short typed primitive that a small frozen model can execute. The system then stitches the primitives together using A* search, optimizing for the correct final action. During inference the program either supplies a chain‑of‑thought for the student model to follow or checks and repairs each proposed tool action before it is sent out.
Results across domains
The same PLVR program was tested with three Nemotron base models and with Granite‑4.2‑8B on three benchmarks: telecom, retail, and LiveMCPBench (a procedural tool‑use suite). On telecom, every Nemotron model reached 93.3‑94.4% accuracy, while the 8B model with PLVR hit 91.2% – just a few points shy. On retail and LiveMCPBench the PLVR program stayed competitive with the best published baselines, though it lagged by 3‑13 points on the latter. The method also reduced the performance spread between base models from about seven points to about two points on average.
Why it matters
For developers building tool‑use assistants, PLVR shows that a relatively cheap 8B model can perform almost as well as much larger models when guided by an explicit program. This could lower compute costs and make such assistants more accessible, though the approach still relies on high‑quality teacher traces and may vary by domain.
