Researchers say you can align a language model to your preferences without touching its weights.
The technique, called spec learning, takes a brief user instruction and a small set of examples showing preferred versus rejected outputs. It compiles those into natural-language specifications — essentially structured prompts — that condition the model at inference time. No gradient updates, no retraining, no GPU bill for fine-tuning. The paper, posted to arXiv, reports that responses guided by compiled specs frequently outperform direct preference optimization (DPO) on specialized-domain datasets where preference signal is dense.
The practical gap this closes is real. Hand-writing system prompts to coax a model toward a behavior is fragile and iterative — change one phrase and the model drifts. Fine-tuning is rigorous but expensive and opaque: once weights update, you cannot read what the model "learned." Spec learning sits between those poles, and the specs it produces are human-readable, which means a team can inspect, version, and debate them like any other document.
The comparison to DPO is the boldest claim here. DPO has become the default cheap-and-cheerful alternative to reinforcement learning from human feedback, so outperforming it without any parameter updates would be notable — if it holds outside the dense-signal domains the paper tests on. Sparse preference data, the harder and more common real-world case, is where this approach still needs to prove itself.