NTT tuning lifts tsuzumi 2 to 83.1% from few manuals

NTT frames the problem as one of scale: conventional tuning to instill company-specific knowledge is said to demand tens of gigabytes of data, while an individual firm's working documents amount to about 1,000 PDFs of up to roughly 100 pages — far too little for standard training. Simply repeating that material, NTT says, overfits the model to specific wording and article numbers, so it cannot answer questions phrased differently, and pulling in similar material from the open web tends to introduce noise.

The technique borrows from how people study: instead of rereading one text, the model acts as a material writer, producing FAQ, table and dialogue-format data that anticipate how the knowledge would be used on the job. NTT says tsuzumi itself drives this generation engine, expanding a small seed set into billions to tens of billions of tokens of training data. In its benchmark, prior paraphrase-based methods such as FinePhrase reached at most 73.3%, against 83.1% for the new technique, and NTT reports no loss of general-domain capability from injecting the specialized knowledge.

NTT positions the method alongside retrieval-augmented generation rather than as a replacement: RAG is recommended where absolute accuracy comes first, while fixing knowledge in the model suits high-speed processing such as real-time analysis of call-center audio, where external retrieval adds latency and GPU cost. NTT says it plans to fold the technique into group solutions and is considering a managed service that runs tuning in its data centers and returns a customized model. At NTT R&D FORUM 2026 it will also present a "self-evolving tsuzumi" concept in which the model autonomously builds and updates a knowledge graph from daily documents and retrains itself — a step toward recursive self-improvement adapted to Japanese corporate data environments, in NTT's framing.