Zero-shot generalization in Stage 3 capability induction

Demonstrate that the Stage 3 capability induction (instruction tuning) of the NEO framework enables generalization to new zero-shot domain capabilities beyond the tasks used for training.

Background

During Stage 3, the model is instruction-tuned on a variety of discovery tasks, but the paper focuses on introducing a diverse task set rather than demonstrating zero-shot capability extension.

The authors explicitly defer the demonstration of zero-shot generalization in Stage 3 to future work, indicating that the behavior of the instruction-tuned model on unseen domain capabilities remains to be established.

References

In this paper, we focus on simply introducing a large variety of types of tasks to the model including text-based retrieval, recommendation, and text generation across heterogeneous items. We save for future work demonstrating the ability to generalize in this stage to new, zero-shot domain capabilities.

— A Unified Language Model for Large Scale Search, Recommendation, and Reasoning  (2603.17533 - Nadai et al., 18 Mar 2026) in Section 3, Capability Induction via Instruction tuning