Goal: Develop reproducible training methods that steer models from memorization to procedural (algorithmic) solutions under noisy/variable input formats.
Modern neural models frequently achieve high in-distribution accuracy via shortcut learning and memorization, yet fail catastrophically under:
- Compositional shifts
- Paraphrase noise
- Out-of-distribution combinations
This is a core blocker for:
- Reliable agents
- Instruction-following systems
- Tool-using models
Where supervision is inherently noisy (synthetic + human + RLHF) and downstream tasks require procedural competence and invariant behavior.
Goal: Develop a reproducible training recipe and early-warning signals that steer models from memorization to procedural (algorithmic) solutions under noisy/variable input formats.