Distilling Few-Shot Examples into Structured Criteria Improves B2B Conversation Classification
June 13, 2026
The paper introduces knowledge extraction methods that compress verbose few-shot ICL examples into compact structured classification criteria for multi-party B2B sales conversations, paired with a new Call Playbook dataset covering five classification tasks. The approach targets the context-length bottleneck that degrades standard ICL in long conversation settings.
HOW THIS AFFECTS YOU
●
builderIf the distillation approach generalizes, it offers a practical way to reduce prompt length and improve ICL accuracy for domain-specific conversation classification without fine-tuning.