Can models learn multi-token concepts during fine-tuning?

Next-token prediction fragments multi-token concepts into arbitrary subword units. "Ribonucleic acid" becomes "rib" → "on" → "ucle" → "ic" → "acid" — five separate prediction targets with no unified semantic representation. Concept-Aware Fine-Tuning (CAFT) introduces multi-token prediction into post-training, enabling models to learn sequences that span multiple tokens as coherent concepts.

Prior multi-token prediction methods worked only during pretraining — prohibitively expensive and dominated by general language modeling rather than domain-specific concept formation. Attempts to apply multi-token prediction to fine-tuning previously failed because multi-token prediction represents a dramatic distribution shift that short post-training phases cannot absorb. CAFT solves this through self-distilled auxiliary heads: first train auxiliary heads (predicting positions beyond the next token) using an instruction-tuning mixture with self-distilled ground truth, then fine-tune with multi-token loss on top of standard LoRA or full fine-tuning.

The results: CAFT consistently outperforms next-token fine-tuning across text summarization and de novo protein design. CAFT LoRA often outperforms next-token full fine-tuning — suggesting models learn more effectively in a multi-token setting even with fewer trainable parameters. In settings where multi-token prediction is highly advantageous (protein design, where amino acid sequences have multi-residue semantic units), multi-fold performance increases are observed.

This connects to the format-shapes-reasoning finding: since Does training data format shape reasoning strategy more than domain?, the prediction unit (single token vs. multi-token) is a format variable that shapes what the model learns. Multi-token prediction is a higher-level format that encourages conceptual chunking rather than token-by-token prediction.

The democratization aspect matters: pretraining-phase MTP was restricted to well-resourced labs. CAFT brings this to fine-tuning, where any practitioner can apply it. Trained task-agnostic auxiliary heads are provided for popular open-source models.