External Hippocampus: Topological Cognitive Maps for Guiding Large Language Model Reasoning

TL;DR

Small-language-models exhibit cognitive deadlock on multi-step reasoning, limiting reliability during inference. The External Hippocampus builds online topological cognitive maps from reasoning trajectories, trains a Learned Navigator to value transitions, and uses test-time perturbations and majority voting to guide reasoning, yielding substantial efficiency and accuracy gains. Empirical analyses reveal model-specific topologies (e.g., plate tectonics vs neural archipelago), low-entropy attractors underpinning deadlock, and a temperature-based recovery mechanism, along with domain-specific transfer limitations. The framework offers a topology-aware, data-efficient path to bolster reasoning in small models and highlights the need for domain-aligned maps and selective perturbations to shape energy landscapes during inference.

Abstract

This paper proposes the External Hippocampus framework, which models language model reasoning from a cognitive dynamics perspective as the flow of information energy in semantic space. Unlike traditional weight-space optimization methods, this framework constructs topological cognitive maps through dimensionality reduction projection, enabling precise navigation and intervention of energy flow at test time while avoiding substantial computational requirements and demonstrating predictable intervention patterns. The method effectively addresses the cognitive deadlock problem in multi-step reasoning for small models. Experiments on models <=7B parameters show: map-guided methods achieve 81.20% accuracy on 500 challenging problems (relative baseline +16.80%), reduce reasoning time by >= 15x, with key findings revealing that reasoning stagnation manifests as "Cognitive Vortex" and low-entropy potential wells, while temperature perturbations effectively restart energy flow. The framework requires no additional training, possesses autonomous growth capability, and provides an efficient and controllable topological-aware solution for small model reasoning.

Figures (7)

• Figure 1: System architecture of the External Hippocampus framework. The system consists of three core components: (1) Cognitive Map Builder that extracts semantic embeddings and constructs topological graphs from reasoning trajectories; (2) Learned Navigator that trains an MLP scorer to predict transition values; (3) Intervention Engine that dynamically injects hints or perturbations based on map states. The entire system updates online at test time for continuous learning.
• Figure 2: Cognitive map topology of Qwen-2.5-3B: "Tectonic Plates" structure with separated high/low-confidence clusters.
• Figure 3: Cognitive map topology of Phi-3-mini: "Neural Archipelago" structure with distributed high-trust islands.
• Figure 4: Token entropy distribution across cognitive states. Deadlock states (trust $<0.2$) show entropy $0.326$, significantly lower than normal states ($0.582$), confirming the low-entropy attractor. Perturbation restores entropy to $0.612$.
• Figure 5: Inverted-U curve of intervention intensity (Phi-3-mini and Qwen-2.5-3B). Balanced strategy ($P<0.5$) achieves the best performance (Phi-3: 68.00%, Qwen-3B: 66.60%), while conservative and aggressive strategies underperform, validating the inverted-U pattern.