How Do LLMs Use Their Depth?

2025-10-21

AI Summary

Is deep utilization of large language models (LLMs) uniform across layers? Existing studies lack fine-grained characterization of layer-wise prediction dynamics. Method: We propose a “Guess-then-Refine” computational framework, integrating intermediate representation tracing on open-weight models, lexical frequency and part-of-speech (POS) analysis, multi-step reasoning decomposition, and response format identification—across diverse case studies. Contribution/Results: We present the first empirical evidence of LLMs’ hierarchical prediction mechanism: early layers generate coarse initial guesses primarily grounded in high-frequency tokens, while deeper layers specialize in context-sensitive refinement. Over 70% of initial predictions are successfully corrected in later layers. Furthermore, we identify a transferable, stage-specific depth-wise division of labor across POS tagging, factual recall, and multiple-choice tasks—revealing consistent functional specialization across architectural depths and task types.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Planning, Routing, and Scheduling: Planning with Language Models

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Growing evidence suggests that large language models do not use their depth uniformly, yet we still lack a fine-grained understanding of their layer-wise prediction dynamics. In this paper, we trace the intermediate representations of several open-weight models during inference and reveal a structured and nuanced use of depth. Specifically, we propose a"Guess-then-Refine"framework that explains how LLMs internally structure their computations to make predictions. We first show that the top-ranked predictions in early LLM layers are composed primarily of high-frequency tokens, which act as statistical guesses proposed by the model early on due to the lack of appropriate contextual information. As contextual information develops deeper into the model, these initial guesses get refined into contextually appropriate tokens. Even high-frequency token predictions from early layers get refined>70% of the time, indicating that correct token prediction is not"one-and-done". We then go beyond frequency-based prediction to examine the dynamic usage of layer depth across three case studies. (i) Part-of-speech analysis shows that function words are, on average, the earliest to be predicted correctly. (ii) Fact recall task analysis shows that, in a multi-token answer, the first token requires more computational depth than the rest. (iii) Multiple-choice task analysis shows that the model identifies the format of the response within the first half of the layers, but finalizes its response only toward the end. Together, our results provide a detailed view of depth usage in LLMs, shedding light on the layer-by-layer computations that underlie successful predictions and providing insights for future works to improve computational efficiency in transformer-based models.

Problem

• Analyzing layer-wise prediction dynamics in large language models
• Investigating how LLMs refine early guesses using contextual depth
• Examining computational depth usage across linguistic tasks

Innovation

• Proposed Guess-then-Refine framework for layer dynamics
• Traced intermediate representations during model inference
• Analyzed depth usage across three linguistic case studies