Query

**Objective:** Conduct a comprehensive, deep literature review and meta-analysis of empirically validated research investigating the cognitive architecture of current Large Language Models (LLMs). The analysis must be strictly grounded in specific theoretical frameworks: the **Complementary Learning Systems (CLS) theory** (specifically the dichotomy between Hippocampal fast/relational learning and Neocortical slow/statistical learning) and the **Dual-Process Theory** (System 1 automatic/associative processing vs. System 2 controlled/logical reasoning). **Core Research Question:** Do current LLMs possess genuine Hippocampus-like capabilities (rapid knowledge binding, bidirectional relational memory) and System 2 capabilities (deliberate planning, rule-based reasoning, logical control), or does empirical evidence prove they are fundamentally confined to Neocortex-like properties (slow statistical pattern matching) and System 1 mechanisms (associative priming and heuristics)? **Scope and Constraints for the Search:** 1. **Empirical Validation Required:** Exclude purely theoretical, philosophical, or opinion-based papers. Only include studies that conduct rigorous empirical experiments, cognitive evaluations, behavioral testing, or mechanistic interpretability on LLMs (e.g., GPT-4, Claude 3, o1, etc.). 2. **Reference Benchmarks:** Look for specific experimental paradigms analogous to human cognitive testing. Examples include "The Reversal Curse" (for testing hippocampal relational binding), "Blocksworld" or multi-step planning tasks (for testing System 2 logic), and evaluations of the true mechanistic nature of Chain-of-Thought (CoT). 3. **Timeframe:** Prioritize the most recent, peer-reviewed papers or highly cited preprints from top-tier conferences (NeurIPS, ICLR, ICML, ACL) published between 2023 and 2026. **Required Output Structure:** **1. Executive Summary:** * A brief overview of the current academic consensus regarding LLMs' cognitive capabilities through the lens of CLS and Dual-Process theories. **2. Evaluation of Hippocampal (CLS) Capabilities in LLMs:** * Detail empirical studies testing whether LLMs can perform fast, relational, and bidirectional learning (Hippocampus) vs. strictly directional, statistical weight updates (Neocortex). * Summarize the methodology and quantitative results of these experiments (e.g., failure modes in relational reasoning, Continual Learning limits without replay). **3. Evaluation of System 2 (Dual-Process) Capabilities in LLMs:** * Detail empirical studies testing whether LLMs possess true System 2 reasoning (planning, explicit rule execution, mental simulation) or merely simulate it using System 1 pattern matching. * Critically analyze studies investigating "Inference-time compute" and "Chain-of-Thought." Do these mechanisms mathematically constitute System 2, or are they proven to be amplified System 1 priming effects? Include experimental proofs. **4. Synthesis and Future Directions:** * Summarize the structural limits of purely autoregressive models based on the gathered empirical evidence. * Briefly identify any emerging architectures (Neuro-symbolic, true external episodic memory buffers, etc.) that aim to authentically implement the missing Hippocampal or System 2 functions.

Run d2defa2d · deepest · started 6/1/2026, 6:00:41 AM