Search papers, labs, and topics across Lattice.
The paper introduces STATe-of-Thoughts (STATe), a novel Inference-Time-Compute (ITC) method that replaces stochastic sampling in Tree-of-Thoughts with discrete, interpretable textual interventions guided by a controller that selects high-level reasoning actions. This structured approach enhances output diversity compared to temperature-based sampling, and allows for interpretable action sequences that are predictive of output quality. By estimating the association between performance and action choices, STATe can identify promising reasoning strategies and steer generation towards them, resulting in high-quality, diverse, and interpretable text.
Ditch the randomness: STATe-of-Thoughts lets you steer LLMs through reasoning tasks with interpretable actions, boosting diversity and quality while revealing the secrets to better outputs.
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.
