Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

$T^2$ of Thoughts: Temperature Tree Elicits Reasoning in Large Language Models

Chengkun Cai, Xu Zhao, Yucheng Du, Haoliang Liu, Lei Li

TL;DR

This paper tackles the limitations of fixed-temperature prompting in large language models by introducing T2oT, a hybrid framework that merges Tree of Thoughts with a dynamic, PSO-guided temperature update to modulate reasoning randomness. By maintaining a single-particle setup, T2oT explores multiple reasoning paths with per-step temperature adjustments, achieving higher single-solution accuracy and richer multi-solution diversity while preserving computational efficiency. Empirical results on the Game of 24 and Creative Writing tasks show that T2oT improves performance over ToT and other baselines in both objective coherency metrics and perceived quality, with dynamic temperature enabling better balance between exploration and exploitation. The work highlights a promising direction for adaptive prompting in LLMs and suggests future enhancements via learnable temperature controllers and multi-modal extensions to further broaden applicability and performance.

Abstract

Large Language Models (LLMs) have emerged as powerful tools in artificial intelligence, especially in complex decision-making scenarios, but their static problem-solving strategies often limit their adaptability to dynamic environments. We explore the enhancement of reasoning capabilities in LLMs through Temperature Tree ($T^2$) prompting via a heuristic algorithm, termed as $T^2$ of Thoughts ($T^2oT$). The primary focus is on enhancing decision-making processes by dynamically adjusting search parameters, especially temperature, to improve accuracy without increasing computational demands. We empirically validate that our hybrid $T^2oT$ approach yields enhancements in, single-solution accuracy, multi-solution generation and text generation quality. Our findings suggest that while dynamic search depth adjustments based on temperature can yield mixed results, a fixed search depth, when coupled with adaptive capabilities of $T^2oT$, provides a more reliable and versatile problem-solving strategy. This work highlights the potential for future explorations in optimizing algorithmic interactions with foundational language models, particularly illustrated by our development for the Game of 24 and Creative Writing tasks.

$T^2$ of Thoughts: Temperature Tree Elicits Reasoning in Large Language Models

TL;DR

This paper tackles the limitations of fixed-temperature prompting in large language models by introducing T2oT, a hybrid framework that merges Tree of Thoughts with a dynamic, PSO-guided temperature update to modulate reasoning randomness. By maintaining a single-particle setup, T2oT explores multiple reasoning paths with per-step temperature adjustments, achieving higher single-solution accuracy and richer multi-solution diversity while preserving computational efficiency. Empirical results on the Game of 24 and Creative Writing tasks show that T2oT improves performance over ToT and other baselines in both objective coherency metrics and perceived quality, with dynamic temperature enabling better balance between exploration and exploitation. The work highlights a promising direction for adaptive prompting in LLMs and suggests future enhancements via learnable temperature controllers and multi-modal extensions to further broaden applicability and performance.

Abstract

Large Language Models (LLMs) have emerged as powerful tools in artificial intelligence, especially in complex decision-making scenarios, but their static problem-solving strategies often limit their adaptability to dynamic environments. We explore the enhancement of reasoning capabilities in LLMs through Temperature Tree () prompting via a heuristic algorithm, termed as of Thoughts (). The primary focus is on enhancing decision-making processes by dynamically adjusting search parameters, especially temperature, to improve accuracy without increasing computational demands. We empirically validate that our hybrid approach yields enhancements in, single-solution accuracy, multi-solution generation and text generation quality. Our findings suggest that while dynamic search depth adjustments based on temperature can yield mixed results, a fixed search depth, when coupled with adaptive capabilities of , provides a more reliable and versatile problem-solving strategy. This work highlights the potential for future explorations in optimizing algorithmic interactions with foundational language models, particularly illustrated by our development for the Game of 24 and Creative Writing tasks.
Paper Structure (27 sections, 1 equation, 4 figures, 5 tables, 1 algorithm)

This paper contains 27 sections, 1 equation, 4 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Prompting and Reasoning for LLMs.
  4. Heuristic Optimization.
  5. Methodology
  6. Problem Formulation.
  7. T2 of Thoughts.
  8. Experiment
  9. Game of 24
  10. Experimental Setup
  11. Task.
  12. Dataset.
  13. Baseline and Framework Setup.
  14. Parameter Setup.
  15. Results
  16. ...and 12 more sections

Figures (4)

  • Figure 1: Our T2oT compared with IO, CoT and ToT. Rectangles of different colors represent thoughts with different evaluations. Arrows of different colors represent different temperatures in the reasoning process.
  • Figure 2: T2oT in a Game of 24. The LM is prompted for thought generation and evaluation. The temperature changes in each step of reasoning.
  • Figure 3: T2oT in Creative Writing. After generating five plans by prompting, dynamically adjust the temperature for producing passages based on the highest score.
  • Figure 4: Creative Writing results.