Jailbreak Scaling Laws for Large Language Models: Polynomial-Exponential Crossover

Adversarial attacks can reliably steer safety-aligned large language models toward unsafe behavior. Empirically, we find that adversarial prompt-injection attacks can amplify attack success rate from the slow polynomial growth observed without injection to exponential growth with the number of inference-time samples. To explain this phenomenon, we propose a theoretical generative model of proxy language in terms of a spin-glass system operating in a replica-symmetry-breaking regime, where generations are drawn from the associated Gibbs measure and a subset of low-energy, size-biased clusters is designated unsafe. Within this framework, we analyze prompt injection-based jailbreaking. Short injected prompts correspond to a weak magnetic field aligned towards unsafe cluster centers and yield a power-law scaling of attack success rate with the number of inference-time samples, while long injected prompts, i.e., strong magnetic field, yield exponential scaling. We derive these behaviors analytically and confirm them empirically on large language models. This transition between two regimes is due to the appearance of an ordered phase in the spin chain under a strong magnetic field, which suggests that the injected jailbreak prompt enhances adversarial order in the language model.

Demystifying LLM-as-a-Judge: Analytically Tractable Model for Inference-Time Scaling

Recent developments in large language models have shown advantages in reallocating a notable share of computational resource from training time to inference time. However, the principles behind inference time scaling are not well understood. In this paper, we introduce an analytically tractable model of inference-time scaling: Bayesian linear regression with a reward-weighted sampler, where the reward is determined from a linear model, modeling LLM-as-a-judge scenario. We study this problem in the high-dimensional regime, where the deterministic equivalents dictate a closed-form expression for the posterior predictive mean and variance. We analyze the generalization error when training data are sampled from a teacher model. We draw $k$ inference-time samples and select via softmax at a temperature applied to a quadratic reward. When the reward is not too different from the teacher, the generalization error decreases monotonically with increasing inference time samples $k$. However, the specific reward that optimizes inference-time selection generally differs from the teacher. In contrast, substantial reward misspecification induces a finite optimal $k$ beyond which more sampling can increase the generalization error. For fixed $k$, there exists an optimal sampling temperature. We experimentally verify these facts in large language model inference with an additional large language model as a judge. In the "best-of-$k$" limit with the teacher as reward, we theoretically show that the generalization error decays as $Θ(1/k^2)$ and determine the leading coefficient via extreme value theory. These formulas delineate domains where scaling inference-time computation is provably preferable to collecting more data. Finally, we demonstrate that when task difficulty increases, the previously mentioned advantage of inference-time compute degrades.

From memorization to generalization: a theoretical framework for diffusion-based generative models

Diffusion-based generative models demonstrate a transition from memorizing the training dataset to a non-memorization regime as the size of the training set increases. Here, we begin by introducing a mathematically precise definition of this transition in terms of a relative distance: the model is said to be in the non-memorization/`generalization' regime if the generated distribution is almost surely far from the probability distribution associated with a Gaussian kernel approximation to the training dataset, relative to the sampling distribution. Then, we develop an analytically tractable diffusion model and establish a lower bound on Kullback-Leibler divergence between the generated and sampling distribution. The model also features the transition, according to our definition in terms of the relative distance, when the training data is sampled from an isotropic Gaussian distribution. Further, our study reveals that this transition occurs when the individual distance between the generated and underlying sampling distribution begins to decrease with the addition of more training samples. This is to be contrasted with an alternative scenario, where the model's memorization performance degrades, but generalization performance doesn't improve. We also provide empirical evidence indicating that realistic diffusion models exhibit the same alignment of scales.