PSSD: Making Large Language Models Self-denial via Human Psyche Structure

TL;DR

PSSD introduces a Freudian-inspired self-denial framework for LLMs, decomposing reasoning into three roles—id (intuition), superego (rule-guided critique), and ego (script-based execution)—to internally identify and rectify mistakes without heavy external tooling. The method includes a two-stage fine-tuning variant (PSSD-SFT) that merges roles into a single model using LoRA with losses $oldsymbol{\\mathcal{L}_1}$ and $oldsymbol{\mathcal{L}_2}$ and a final weight $W = W_0 + W_1 + W_2$. Experiments across textual and mathematical reasoning tasks show PSSD and PSSD-SFT achieve superior EM, PM, and RM compared to Stand, CoT variants, and other baselines, with strong compatibility with external tools. The work demonstrates that embedding a self-denial mechanism within LLMs improves reasoning accuracy and efficiency, offering a practical, orthogonal addition to current approaches and a pathway for future enhancements in reliability and integration with retrieval systems.

Abstract

The enhance of accuracy in reasoning results of LLMs arouses the community's interests, wherein pioneering studies investigate post-hoc strategies to rectify potential mistakes. Despite extensive efforts, they are all stuck in a state of resource competition demanding significant time and computing expenses. The cause of the situation lies in the failure of identifying the fundamental feature of the solutions in this line, coined as the self-denial of LLMs. In other words, LLMs should confidently determine the potential existence of mistakes and carefully execute the targeted correction. As the whole procedure conducts within LLMs, supporting and persuasive references are hard to acquire, while the absence of specific steps towards refining hidden mistakes persists even when errors are acknowledged. In response to the challenges, we present PSSD, which refers to and implements the human psyche structure such that three distinct and interconnected roles contribute to human reasoning. Specifically, PSSD leverages the recent multi-agent paradigm, and is further enhanced with three innovatively conceived roles: (1) the intuition-based id role that provides initial attempts based on benign LLMs; (2) the rule-driven superego role that summarizes rules to regulate the above attempts, and returns specific key points as guidance; and (3) the script-centric ego role that absorbs all procedural information to generate executable script for the final answer prediction. Extensive experiments demonstrate that the proposed design not only better enhance reasoning capabilities, but also seamlessly integrate with current models, leading to superior performance.

Figures (5)

• Figure 1: The framework of PSSD. It is mainly constituted of the id role, the superego role, and the ego role. Here we expect that the discussion of the three roles will enlighten LLMs to reason better.
• Figure 2: Sketch of the consistency distribution in different methods. Consistency denotes the models confidence on predictions, and density denotes the probability density. We present the results in AdvHotpotQA for illustration. Details in Section \ref{['sec:consistency']}.
• Figure 3: The comparison of different paradigms.
• Figure 4: We compare the results of the PSSD and Self-Contrast using two pie charts. It shows PSSD is more accurate and stable than direct Self-Contrast.
• Figure 5: Sketch of the answering type distribution between CoT-SC and PSSD in AdvHotpotQA. The presence of the missing answer indicates incorrect results including exact answers, thereby contributing to the overall count of incorrect answers. The height of each bar means the corresponding answer type in CoT-SC, and the inside detailed types come from the results of PSSD. Details in Section \ref{['sec:ans']}.