While test-time reasoning enables language models (LMs) to tackle complex tasks, searching or planning in natural language can be slow, costly, and error-prone. But even when LMs struggle to emulate the precise reasoning steps needed to solve a problem, they often excel at describing its abstract structure--both how to verify solutions and how to search for them. This paper introduces DisCIPL, a method for "self-steering" LMs where a Planner model generates a task-specific inference program that is executed by a population of Follower models. Our approach equips LMs with the ability to write recursive search procedures that guide LM inference, enabling new forms of verifiable and efficient reasoning. When instantiated with a small Follower (e.g., Llama-3.2-1B or Qwen3-1.7B), DisCIPL matches (and sometimes outperforms) much larger models, including GPT-4o and o1, on challenging constrained generation tasks. Our work opens up a design space of highly-parallelized Monte Carlo inference strategies that outperform standard best-of-N sampling, require no finetuning, and can be implemented automatically by existing LMs.

Inference amortization methods allow the sharing of statistical strength across related observations when learning to perform posterior inference. Generally this requires the inversion of the dependency structure in the generative model, as the modeller must design and learn a distribution to approximate the posterior. Previous methods invert the dependency structure in a heuristic way and fail to capture the dependencies in the model, therefore limiting the performance of the eventual inference algorithm. We introduce an algorithm for faithfully and minimally inverting the graphical model structure of any generative model. Such an inversion has two crucial properties: a) it does not encode any independence assertions absent from the model, and b) for a given inversion, it encodes as many true independence assertions as possible. Our algorithm works by simulating variable elimination on the generative model to reparametrize the distribution. We show with experiments how such minimal inversions can assist in performing better inference.

This paper introduces the probabilistic module interface, which allows encapsulation of complex probabilistic models with latent variables alongside custom stochastic approximate inference machinery, and provides a platform-agnostic abstraction barrier separating the model internals from the host probabilistic inference system. The interface can be seen as a stochastic generalization of a standard simulation and density interface for probabilistic primitives. We show that sound approximate inference algorithms can be constructed for networks of probabilistic modules, and we demonstrate that the interface can be implemented using learned stochastic inference networks and MCMC and SMC approximate inference programs.

This paper introduces a new technique for quantifying the approximation error of a broad class of probabilistic inference programs, including ones based on both variational and Monte Carlo approaches. The key idea is to derive a subjective bound on the symmetrized KL divergence between the distribution achieved by an approximate inference program and its true target distribution. The bound's validity (and subjectivity) rests on the accuracy of two auxiliary probabilistic programs: (i) a "reference" inference program that defines a gold standard of accuracy and (ii) a "meta-inference" program that answers the question "what internal random choices did the original approximate inference program probably make given that it produced a particular result?" The paper includes empirical results on inference problems drawn from linear regression, Dirichlet process mixture modeling, HMMs, and Bayesian networks. The experiments show that the technique is robust to the quality of the reference inference program and that it can detect implementation bugs that are not apparent from predictive performance.

The discovery, representation and reconstruction of (technical) integration networks from Network Mining (NM) raw data is a difficult problem for enterprises. This is due to large and complex IT landscapes within and across enterprise boundaries, heterogeneous technology stacks, and fragmented data. To remain competitive, visibility into the enterprise and partner IT networks on different, interrelated abstraction levels is desirable. We present an approach to represent and reconstruct the integration networks from NM raw data using logic programming based on first-order logic. The raw data expressed as integration network model is represented as facts, on which rules are applied to reconstruct the network. We have built a system that is used to apply this approach to real-world enterprise landscapes and we report on our experience with this system.