How much can mitigation techniques like augmentation reduce priming without harming learning?

This explores whether mitigation techniques like augmentation can dial down priming without also dialing down learning — and the honest starting point is that the corpus suggests priming is more predictable than it is freely tunable. The clearest anchor is the finding that keyword priming after training is almost entirely forecastable from a word's probability *before* training: there's a ~10^-3 threshold separating contexts where priming kicks in from those where it doesn't, and just three exposures lock the effect in Can we predict keyword priming before learning happens?. That matters for your question because it reframes the goal: you're not erasing a random side effect, you're trying to keep a model from absorbing a real statistical regularity in its data. So the real question becomes whether you can change *how* a model absorbs information rather than just how much.

That's exactly where the augmentation evidence gets interesting — and double-edged. Thinking-augmented pretraining shows that adding generated reasoning traces to the data improves data efficiency 3x and benchmark performance by 10%+, with harder tokens automatically attracting longer traces Can training data augmentation match test-time compute scaling benefits?. The lesson buried in that result is that augmentation doesn't have to be a blunt 'add more examples' lever; it can reshape *what the model attends to* during learning, surrounding a keyword with reasoning context instead of letting it fire in isolation. Consistency training pushes the same idea more directly at the priming problem: it teaches a model to respond identically to 'clean' and 'wrapped' prompts using the model's own clean answers as targets, building invariance to surface perturbations without the staleness that plagues ordinary fine-tuning Can models learn to ignore irrelevant prompt changes?. That's the closest thing in the corpus to 'reduce priming, preserve capability.'

But the surrounding notes are a warning that mitigation almost always extracts a price somewhere, and the trick is choosing *where* to pay it. Training dense retrievers to be more sensitive to compositional structure reliably degrades their zero-shot generalization by 8–40% — and the authors are explicit that this is a geometric trade-off baked into high-dimensional space, not a tuning bug you can engineer away Does training for compositional sensitivity hurt dense retrieval?. RL post-training shows a parallel collapse: it amplifies one dominant format from pretraining and suppresses the alternatives within a single epoch, meaning 'cleaning up' behavior can quietly narrow the model Does RL training collapse format diversity in pretrained models?. The pattern across both: interventions that suppress one behavior tend to compress diversity or generalization as collateral.

The most hopeful framing in the collection is that this tradeoff is often a *misallocation* problem rather than a hard cost. Fast-Slow Training routes task-specific lessons into optimized prompts (a fast, disposable channel) while keeping weight updates minimal, reaching the same performance 1.4–3x faster with substantially less catastrophic forgetting Can splitting adaptation into two channels reduce forgetting?. The framing there — forgetting isn't inherent, it's bad routing — is exactly the optimistic answer to your question: if priming lives in one channel and the learning you care about lives in another, you can attack one without touching the other. Context-as-playbook approaches make the same bet at inference time, using incremental curation instead of full rewrites to avoid erasing detail Can context playbooks prevent knowledge loss during iteration?.

So the synthesis: there's no published 'X% priming reduction at Y% learning cost' number in this corpus, and you should be suspicious of anyone who quotes one as universal — priming is predictable enough that suppression means fighting a genuine data regularity, and the trade-off literature here (retrieval, RL, fine-tuning) shows costs reliably reappear *somewhere*. The leverage isn't in finding a magic augmentation strength; it's in *separating channels* so the thing you want to suppress and the thing you want to keep aren't sharing the same dial.