Elicitation without Backpropagation: Steering Model Behavior by Optimizing the Latent Posterior

In the \emph{latent posterior model} of transformer behavior, the next-token distribution arises from a posterior over latent predictive models conditioned on the context, mixed to generate continuations. We exploit this model in settings where it is exact, namely Bayes-filtered transformers (BFTs) meta-learned on sequences from a hierarchical prior, to introduce \textbf{Posterior Prefix Tuning (PPT)}, a new method for \emph{eliciting} behavior from a transformer: given a utility function on continuations, find a prompt under which the transformer generates continuations of high expected utility. For a BFT, the elicitation objective factors through the latent posterior, and the gradient of this objective can be estimated from samples of the prior alone. PPT optimizes the parameters of a distribution over hard prompts: it draws prior samples once from the BFT via predictive Monte Carlo (PMC), then estimates the gradient by importance sampling against them. The optimization performs no transformer forward passes and no backpropagation through the transformer, and the prior samples are utility-independent, so a single set of samples drives elicitation against any number of utilities at negligible marginal cost. We validate PPT on Beta--Bernoulli and reinforced urn BFTs across three utility families (reverse cross-entropy, frequency matching, Dyck validity).

Towards Spectroscopy: Susceptibility Clusters in Language Models

Spectroscopy infers the internal structure of physical systems by measuring their response to perturbations. We apply this principle to neural networks: perturbing the data distribution by upweighting a token $y$ in context $x$, we measure the model's response via susceptibilities $χ_{xy}$, which are covariances between component-level observables and the perturbation computed over a localized Gibbs posterior via stochastic gradient Langevin dynamics (SGLD). Theoretically, we show that susceptibilities decompose as a sum over modes of the data distribution, explaining why tokens that follow their contexts "for similar reasons" cluster together in susceptibility space. Empirically, we apply this methodology to Pythia-14M, developing a conductance-based clustering algorithm that identifies 510 interpretable clusters ranging from grammatical patterns to code structure to mathematical notation. Comparing to sparse autoencoders, 50% of our clusters match SAE features, validating that both methods recover similar structure.

Embryology of a Language Model

Understanding how language models develop their internal computational structure is a central problem in the science of deep learning. While susceptibilities, drawn from statistical physics, offer a promising analytical tool, their full potential for visualizing network organization remains untapped. In this work, we introduce an embryological approach, applying UMAP to the susceptibility matrix to visualize the model's structural development over training. Our visualizations reveal the emergence of a clear ``body plan,'' charting the formation of known features like the induction circuit and discovering previously unknown structures, such as a ``spacing fin'' dedicated to counting space tokens. This work demonstrates that susceptibility analysis can move beyond validation to uncover novel mechanisms, providing a powerful, holistic lens for studying the developmental principles of complex neural networks.

Studying Small Language Models with Susceptibilities

We develop a linear response framework for interpretability that treats a neural network as a Bayesian statistical mechanical system. A small, controlled perturbation of the data distribution, for example shifting the Pile toward GitHub or legal text, induces a first-order change in the posterior expectation of an observable localized on a chosen component of the network. The resulting susceptibility can be estimated efficiently with local SGLD samples and factorizes into signed, per-token contributions that serve as attribution scores. Building a set of perturbations (probes) yields a response matrix whose low-rank structure separates functional modules such as multigram and induction heads in a 3M-parameter transformer. Susceptibilities link local learning coefficients from singular learning theory with linear-response theory, and quantify how local loss landscape geometry deforms under shifts in the data distribution.

Generalization Analogies: A Testbed for Generalizing AI Oversight to Hard-To-Measure Domains

As AI systems become more intelligent and their behavior becomes more challenging to assess, they may learn to game the flaws of human feedback instead of genuinely striving to follow instructions; however, this risk can be mitigated by controlling how LLMs generalize human feedback to situations where it is unreliable. To better understand how reward models generalize, we craft 69 distribution shifts spanning 8 categories. We find that reward models do not learn to evaluate `instruction-following' by default and instead favor personas that resemble internet text. Techniques for interpreting reward models' internal representations achieve better generalization than standard fine-tuning, but still frequently fail to distinguish instruction-following from conflated behaviors. We consolidate the 15 most challenging distribution shifts into the GENeralization analogIES (GENIES) benchmark, which we hope will enable progress toward controlling reward model generalization.