可微筛选与架构边界 v1.1 ——以早期符号逻辑为参照系解释大模型伪推理现象
Year: 2026
Type: preprint
Abstract: 现有研究大多将大语言模型的推理缺陷归因于数据规模、上下文窗口或对齐策略。本文提出一个分析性假说:在向量表征与梯度下降的优化范式之下,Transformer 在结构上只承载了早期符号逻辑体系中可离散、可微、可优化的那一部分,与之并列的另一部分认知能力并未进入架构。本文把这一结果称为 “可微筛选”,并强调它是筛选现象而非可追溯的设计决策;筛选的历史因果方向仍作为开放问题保留。在此参照框架下,架构层面缺乏三项推理的架构级保证:变量—实体跨步绑定(A)、规则泛化(B)、推理链组合一致性(C₁);与之相邻的语法格式合法性另记为 C₂。缺乏保证不等于能力为零,模型在分布内仍可偶发达成,但架构不提供强制约束,因而在需要跨步绑定的任务上,最近邻结构替换成为主导的失效模式。本文据此把主流增强技术归入三条路线:内置拟合、序列展开式计算、外挂增强,并逐类说明其作用对象与实践上限。本文给出可观测的差异化预测与可执行的判决条件,用以区分 “数据不足解释” 与 “架构缺口解释”。最后讨论与验证性缺失研究的接口,以及本框架的适用边界。 Abstract Most existing studies attribute the reasoning deficiencies of large language models to data scale, context windows, or alignment strategies. This paper proposes an analytical hypothesis: under the optimization paradigm of vector representations and gradient descent, Transformers structurally encode only the portion within early symbolic coexists with logical this systems has not that entered is discrete, the architecture. differentiable, This paper and optimizable—the refers to this outcome other portion as "differentiable of cognitive screening" capacity that and emphasizes that it is a screening phenomenon rather than a traceable design decision; the historical causal direction types of architectural-level of the screening remains guarantees an open for reasoning: problem. Within variable–entity this reference stepwise framework, binding the (A), architecture rule generalization lacks three (B), and reasoning chain compositional consistency (C₁); the adjacent grammatical format validity is denoted separately as C₂. The lack of guarantees does not imply zero capability—the model can still occasionally succeed within-distribution—but since the architecture does not impose mandatory constraints, nearest-neighbor structural substitution categorizes mainstream becomes the enhancement dominant failure techniques mode in into tasks three requiring paradigms: stepwise in-built binding. fitting, Based sequence-unfolding on this, this paper computation, and external augmentation—and elaborates for each category on their targets of action and practical limits. between This the paper "insufficient-data presents observable explanation" differential and the "architectural-gap predictions and executable explanation." decision Finally, conditions it discusses to the distinguish interface with verificationally deficient studies and the applicability boundaries of this framework.
Author: 狂潮
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236818444
