Abstract
Knowledge-based complex reasoning remains a significant challenge for large language models (LLMs) with in-context learning. To tackle this issue, previous studies focus on ensuring behavior fidelity, factuality, or reliability in generated reasoning processes that guide LLMs to produce solutions. However, these studies often neglect the simultaneous optimization on all these three aspects for each thought. The main challenges are the lack of comprehensive assessment mechanisms and the difficulty of efficient thought-level optimization. This paper introduces the Evolution of Thoughts (EoT) framework, which enhances the factuality, fidelity, and reliability of each thought in the reasoning process through a few LLM inferences. We propose a thought assessment method that is sensitive to knowledge and LLM behaviors, using three scorers to evaluate each thought by considering domain context, semantic alignment, and behavior impact. Additionally, we establish a self-reflective evolution mechanism to facilitate each reasoning process generation in a single-forward inference. Extensive experiments demonstrate that, for knowledge-based complex tasks, EoT improves the factuality and fidelity of reasoning processes by approximately 16.5% and 48.8%, respectively, while enhancing LLM reasoning capability by about 6.2%, outperforming advanced approaches.
Introduction
While Large Language Models (LLMs) have become adept at simple tasks, "knowledge-based complex reasoning"—tasks requiring the synthesis of long-context information and intricate logical steps—remains a hurdle. Existing prompting strategies like Chain-of-Thought (CoT) or Tree-of-Thoughts (ToT) often suffer from three specific gaps:
- Hallucinations: Thoughts are generated that aren't grounded in the provided context (low factuality).
- Explanation Discrepancy: The "thoughts" the model writes out don't actually match the logic it used to reach the answer (low fidelity).
- Incomplete Optimization: Most frameworks optimize for one of these factors but neglect the others, leading to reasoning processes that might look good but lead to incorrect answers.
The Evolution of Thoughts (EoT) framework addresses these gaps by treating reasoning as a multi-objective optimization problem at the individual thought level.
The EoT Framework
The EoT framework operates in an iterative loop consisting of three main modules: the Assessor, the Prompter, and the Generator.
1. The Assessor
This module uses three specialized scorers to evaluate the current reasoning process ($R$):
- Factuality Scorer: Uses Natural Language Inference (NLI) to determine if each thought is supported by the provided domain knowledge.
- Reliability Scorer: Introduces a novel hybrid metric called Hssim, which combines the learned metric BLEURT with NLI judgments to measure how closely the generated answer aligns with the reference solution.
- Fidelity Scorer: Employs a "thought masking" technique. It removes a specific thought and checks if the model's answer changes. If the answer stays the same without the thought, that thought has low fidelity (it didn't actually influence the model's logic).
2. The Prompter
The Prompter translates these scores into a "self-reflection" prompt. This prompt tells the LLM exactly where its reasoning failed (e.g., "Thought 2 contradicts the context") and provides instructions for multi-objective optimization.
3. The Generator
The task-solving LLM itself acts as the generator. In a single-forward inference pass per iteration, it attempts to produce a refined reasoning process that addresses the feedback from the prompter.
Experimental Evaluation
The researchers evaluated EoT on two primary benchmarks: OpsQA, containing 40 complex operational maintenance tasks with contexts averaging over 6,000 tokens, and LongBench, a subset of 60 representative questions from open-source datasets like HotpotQA.
Key Results
Extensive experiments using Qwen2-72B and GPT-4 Turbo showed that EoT consistently outperformed advanced baselines like BoT, xLLM, and CoT-dec.
As illustrated in the performance distributions, EoT achieves significant improvements across all dimensions:
- Factuality: Improved by approximately 16.5%.
- Behavior Fidelity: Improved by a massive 48.8%.
- Overall Reasoning Capability (Reliability): Boosted by about 6.2%.
Beyond performance, EoT is highly efficient. It reduced iteration overhead by about 92% compared to the Tree-of-Thoughts (BoT) approach, which requires exploring many different branches.
Conclusion and Future Work
The core insight of EoT is that factuality, fidelity, and reliability are deeply interconnected. Improving one in isolation often degrades the others; EoT is the first to treat this as a multi-objective optimization problem at the thought level. While the current framework relies on reference answers for reliability assessment, future work will explore unsupervised evolution mechanisms to broaden its applicability to open-ended scenarios.
Summary updated