21% accuracy boost for on-device LMs using lightweight Process Supervision.
I paired a Qwen3-0.6B generator with an 8B PRM verifier (LoRA-finetuned on PRM800K) and hit 40.4% on GSM8K vs 33.3% self-consistency.
#LLM#OnDeviceAI#Finetuning
@chantastic I think he's talking about some kind of workflow that finds issues and creates optimized prompts for /goal . Not sure how the verifier process is tho