# Fuzzy, Symbolic, and Contextual: Enhancing LLM Instruction via Cognitive Scaffolding

**Authors:** Vanessa Figueiredo  
**arXiv:** [2508.21204](https://arxiv.org/abs/2508.21204)  
**Date:** 2025-08-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2508.21204v1/

## Key points

- Investigating impact of architectural inductive biases on cognitive behavior of large language models (LLMs) during instructional dialogue
- Introducing symbolic scaffolding mechanism and short-term memory schema for adaptive and structured reasoning in Socratic tutoring scenarios
- Assessing model outputs through controlled ablation experiments across five system variants using expert-designed rubrics covering aspects such as scaffolding, responsiveness, symbolic reasoning, and conversational memory
- Preliminary results show full system consistently outperforms baseline variants; removing memory or symbolic structure leads to degradation of key cognitive behaviors including abstraction, adaptive probing, and conceptual continuity
- Importance of architectural scaffolds in shaping emergent instructional strategies within LLMs supported by findings
- Making code and annotations publicly available for transparency and reproducibility upon acceptance
- Contributions include development of modular natural language boundary architecture with fuzzy, symbolic scaffolding; short-term memory schema for turn-by-turn cognitive control; prompt-level symbolic loop for real-time strategy modulation; and evaluation framework adapted from instructional science
- Discussing related work on schemas as controllers, memory-augmented LMs, fuzzy reasoning techniques, and interpretability methods
- Emphasizing foregrounding symbolic interpretability through schema-guided reasoning to advance operational cognition in LLMs with session-level coherence, dynamic adaptivity, and interpretability support

## Summary

In this study, we investigate the impact of architectural inductive biases on the cognitive behavior of large language models (LLMs) during instructional dialogue. We introduce a symbolic scaffolding mechanism and a short-term memory schema to facilitate adaptive and structured reasoning in Socratic tutoring scenarios. Through controlled ablation experiments across five system variants, we assess model outputs using expert-designed rubrics that cover aspects such as scaffolding, responsiveness, symbolic reasoning, and conversational memory. Our preliminary results show that our full system consistently outperforms baseline variants. Specifically, removing memory or symbolic structure leads to a degradation of key cognitive behaviors including abstraction, adaptive probing, and conceptual continuity. These findings support the importance of architectural scaffolds in shaping emergent instructional strategies within LLMs. Furthermore, we make our code and annotations publicly available upon acceptance for increased transparency and reproducibility. The preprint serves as a high-throughput behavioral screening method for early experimental stages. Our contributions include the development of a modular natural language boundary architecture incorporating fuzzy, symbolic scaffolding; a short-term memory schema for turn-by-turn cognitive control; a prompt-level symbolic loop for real-time strategy modulation; and an evaluation framework adapted from instructional science. We also discuss related work on schemas as controllers, memory-augmented LMs, fuzzy reasoning techniques, and interpretability methods. By foregrounding symbolic interpretability through schema-guided reasoning and embedding structure at inference time to produce interpretable responses aligned with pedagogical principles, we aim to advance operational cognition in LLMs. This approach emphasizes session-level coherence and dynamic adaptivity while supporting interpretability—a hallmark of cognitive control—rather than solely probing black-box behavior.