A research framework lets developers bolt safety constraints onto a language model after training — no weight updates required.
Called Lagrangian Reward Augmentation, or LARA, the method works by adding a second scoring signal — a cost model — alongside the reward model that normally guides inference-time alignment. Instead of cramming safety requirements into a single blended score (a common workaround that demands manual tuning), LARA uses a mathematical technique called dualization to reduce the problem to a one-dimensional calculation. That produces an augmented reward signal that plugs into existing inference-time methods without redesigning them. The researchers tested it on both sequence-level approaches, like Best-of-N reranking, and token-level decoding methods.
The practical upshot: teams deploying models they cannot or will not retrain — because of cost, contractual limits, or a preference to keep weights frozen — get a principled way to enforce safety budgets rather than hoping a penalty coefficient is tuned correctly. Best-of-N with LARA came closest to matching finetuning-based alignment baselines, which is a meaningful gap to close.
The caveat the paper is honest about: for token-level decoding, LARA yields a "dual-calibrated heuristic" rather than a guaranteed constrained policy — meaning it is a disciplined approximation, not a proof. That distinction matters if vendors start marketing inference-time alignment as equivalent to proper safety training, which they will.