Fibration Policy Optimization

Large language models are increasingly trained as heterogeneous systems spanning multiple domains, expert partitions, and agentic pipelines, yet prevalent proximal objectives operate at a single scale and lack a principled mechanism for coupling token-level, trajectory-level, and higher-level hierarchical stability control. To bridge this gap, we derive the Aggregational Policy Censoring Objective (APC-Obj), the first exact unconstrained reformulation of sample-based TV-TRPO, establishing that clipping-based surrogate design and trust-region optimization are dual formulations of the same problem. Building on this foundation, we develop Fiber Bundle Gating (FBG), an algebraic framework that organizes sampled RL data as a fiber bundle and decomposes ratio gating into a base-level gate on trajectory aggregates and a fiber-level gate on per-token residuals, with provable first-order agreement with the true RL objective near on-policy. From APC-Obj and FBG we derive Fibration Policy Optimization (or simply, FiberPO), a concrete objective whose Jacobian is block-diagonal over trajectories, reduces to identity at on-policy, and provides better update direction thus improving token efficiency. The compositional nature of the framework extends beyond the trajectory-token case: fibrations compose algebraically into a Fibration Gating Hierarchy (FGH) that scales the same gating mechanism to arbitrary hierarchical depth without new primitives, as demonstrated by FiberPO-Domain, a four-level instantiation with independent trust-region budgets at the domain, prompt group, trajectory, and token levels. Together, these results connect the trust-region theory, a compositional algebraic structure, and practical multi-scale stability control into a unified framework for LLM policy optimization.

Learning Temporal Abstractions via Variational Homomorphisms in Option-Induced Abstract MDPs

Large Language Models (LLMs) have shown remarkable reasoning ability through explicit Chain-of-Thought (CoT) prompting, but generating these step-by-step textual explanations is computationally expensive and slow. To overcome this, we aim to develop a framework for efficient, implicit reasoning, where the model "thinks" in a latent space without generating explicit text for every step. We propose that these latent thoughts can be modeled as temporally-extended abstract actions, or options, within a hierarchical reinforcement learning framework. To effectively learn a diverse library of options as latent embeddings, we first introduce the Variational Markovian Option Critic (VMOC), an off-policy algorithm that uses variational inference within the HiT-MDP framework. To provide a rigorous foundation for using these options as an abstract reasoning space, we extend the theory of continuous MDP homomorphisms. This proves that learning a policy in the simplified, abstract latent space, for which VMOC is suited, preserves the optimality of the solution to the original, complex problem. Finally, we propose a cold-start procedure that leverages supervised fine-tuning (SFT) data to distill human reasoning demonstrations into this latent option space, providing a rich initialization for the model's reasoning capabilities. Extensive experiments demonstrate that our approach achieves strong performance on complex logical reasoning benchmarks and challenging locomotion tasks, validating our framework as a principled method for learning abstract skills for both language and control.