Technology Category
Knowledge Representation and Reasoning: Action, Change, and CausalityCognitive Modeling & Cognitive Systems: Conceptual Inference and ReasoningReasoning under Uncertainty: Sequential Decision Making
STATe-of-Thoughts: Structured Action Templates for Tree-of-Thoughts
🤖 AI Summary
Existing inference-time computation methods struggle to simultaneously achieve high output quality, diversity, and controllability of the reasoning process. This work proposes a structured and interpretable reasoning framework that replaces stochastic sampling with discrete textual interventions and introduces explicit action sequences to govern generation. The framework employs a tripartite architecture comprising a controller, a generator, and an evaluator: the controller selects high-level reasoning actions, the generator produces diverse and high-quality reasoning paths conditioned on these actions, and the evaluator guides the search toward high-potential regions. Evaluated on tasks such as argument generation, the approach significantly enhances response diversity, demonstrates strong correlation between action sequences and output quality, and enables targeted steering of the reasoning trajectory.
Technology Category
Application Category
📝 Abstract
Inference-Time-Compute (ITC) methods like Best-of-N and Tree-of-Thoughts are meant to produce output candidates that are both high-quality and diverse, but their use of high-temperature sampling often fails to achieve meaningful output diversity. Moreover, existing ITC methods offer limited control over how to perform reasoning, which in turn limits their explainability. We present STATe-of-Thoughts (STATe), an interpretable ITC method that searches over high-level reasoning patterns. STATe replaces stochastic sampling with discrete and interpretable textual interventions: a controller selects actions encoding high-level reasoning choices, a generator produces reasoning steps conditioned on those choices, and an evaluator scores candidates to guide search. This structured approach yields three main advantages. First, action-guided textual interventions produce greater response diversity than temperature-based sampling. Second, in a case study on argument generation, STATe's explicit action sequences capture interpretable features that are highly predictive of output quality. Third, estimating the association between performance and action choices allows us to identify promising yet unexplored regions of the action space and steer generation directly toward them. Together, these results establish STATe as a practical framework for generating high-quality, diverse, and interpretable text. Our framework is available at https://github.com/zbambergerNLP/state-of-thoughts.
Problem
Research questions and friction points this paper is trying to address.
Inference-Time-Compute
output diversity
reasoning control
interpretability
Tree-of-Thoughts
Innovation
Methods, ideas, or system contributions that make the work stand out.
Structured Action Templates
Tree-of-Thoughts
Inference-Time Compute
Interpretable Reasoning
Action-Guided Generation
🔎 Similar Papers
No similar papers found.
💼 Related Jobs
No related jobs found.
