PLDR-LLM and Emergent Reasoning

A recent study published on arXiv introduces PLDR-LLMs (Pretrained Language Models with Deductive Reasoning), large language models trained under self-organized criticality. The research indicates that these models are capable of exhibiting reasoning abilities during the inference phase.

Criticality and Phase Transitions

The characteristics of PLDR-LLM deductive outputs under critical conditions show similarities to second-order phase transitions. Under such conditions, the correlation length diverges, and the deductive outputs reach a metastable steady state. This behavior suggests that the models learn representations equivalent to scaling functions, universality classes, and renormalization groups from the training data, leading to generalization and reasoning capabilities.

Order Parameters and Reasoning Ability

The study defines an order parameter based on the global statistics of the model's deductive output parameters during inference. The reasoning capabilities of a PLDR-LLM are superior when its order parameter is close to zero under critical conditions. This observation is supported by the benchmark results obtained from models trained in near-critical and sub-critical conditions.

Implications

The results provide a self-contained explanation of how reasoning manifests in large language models. Furthermore, the ability to reason can be quantified solely from the global parameter values of the model at steady state, without the need to evaluate curated benchmark datasets through inductive output for reasoning and comprehension.