# How Do LLMs Use Their Depth?

arXiv: https://arxiv.org/abs/2510.18871 · HTML: https://mlsanity.com/paper/2510.18871/

## Key facts

- Cited by 16 papers in the map (data through 14 Oct 2026).
- Builds on 31 of its 43 references that are in the map.
- First posted on arXiv 21 Oct 2025.

## Builds on (most cited first, 30 of 31)

- [Measuring Massive Multitask Language Understanding](https://mlsanity.com/paper/2009.03300/) (2020)
- [Layer Normalization](https://mlsanity.com/paper/1607.06450/) (2016)
- [Locating and Editing Factual Associations in GPT](https://mlsanity.com/paper/2202.05262/) (2022)
- [Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling](https://mlsanity.com/paper/2304.01373/) (2023)
- [Root Mean Square Layer Normalization](https://mlsanity.com/paper/1910.07467/) (2019)
- [Transformer Feed-Forward Layers Are Key-Value Memories](https://mlsanity.com/paper/2012.14913/) (2020)
- [An Explanation of In-context Learning as Implicit Bayesian Inference](https://mlsanity.com/paper/2111.02080/) (2021)
- [Transformers learn in-context by gradient descent](https://mlsanity.com/paper/2212.07677/) (2022)
- [Eliciting Latent Predictions from Transformers with the Tuned Lens](https://mlsanity.com/paper/2303.08112/) (2023)
- [Transformer Feed-Forward Layers Build Predictions by Promoting Concepts in the Vocabulary Space](https://mlsanity.com/paper/2203.14680/) (2022)
- [Datasets: A Community Library for Natural Language Processing](https://mlsanity.com/paper/2109.02846/) (2021)
- [Dissecting Recall of Factual Associations in Auto-Regressive Language Models](https://mlsanity.com/paper/2304.14767/) (2023)
- [ShortGPT: Layers in Large Language Models are More Redundant Than You Expect](https://mlsanity.com/paper/2403.03853/) (2024)
- [MQuAKE: Assessing Knowledge Editing in Language Models via Multi-Hop Questions](https://mlsanity.com/paper/2305.14795/) (2023)
- [Depth-Adaptive Transformer](https://mlsanity.com/paper/1910.10073/) (2019)
- [Label Words are Anchors: An Information Flow Perspective for Understanding In-Context Learning](https://mlsanity.com/paper/2305.14160/) (2023)
- [The Unreasonable Ineffectiveness of the Deeper Layers](https://mlsanity.com/paper/2403.17887/) (2024)
- [Why Can GPT Learn In-Context? Language Models Implicitly Perform Gradient Descent as Meta-Optimizers](https://mlsanity.com/paper/2212.10559/) (2022)
- [Linearity of Relation Decoding in Transformer Language Models](https://mlsanity.com/paper/2308.09124/) (2023)
- [Adaptive Neural Networks for Efficient Inference](https://mlsanity.com/paper/1702.07811/) (2017)
- [Future Lens: Anticipating Subsequent Tokens from a Single Hidden State](https://mlsanity.com/paper/2311.04897/) (2023)
- [Language Models Implement Simple Word2Vec-style Vector Arithmetic](https://mlsanity.com/paper/2305.16130/) (2023)
- [Not All Layers of LLMs Are Necessary During Inference](https://mlsanity.com/paper/2403.02181/) (2024)
- [Jump to Conclusions: Short-Cutting Transformers With Linear Transformations](https://mlsanity.com/paper/2303.09435/) (2023)
- [Do Language Models Use Their Depth Efficiently?](https://mlsanity.com/paper/2505.13898/) (2025)
- [Do language models plan ahead for future tokens?](https://mlsanity.com/paper/2404.00859/) (2024)
- [Evidence of Learned Look-Ahead in a Chess-Playing Neural Network](https://mlsanity.com/paper/2406.00877/) (2024)
- [DecoderLens: Layerwise Interpretation of Encoder-Decoder Transformers](https://mlsanity.com/paper/2310.03686/) (2023)
- [Unlocking the Future: Exploring Look-Ahead Planning Mechanistic Interpretability in Large Language Models](https://mlsanity.com/paper/2406.16033/) (2024)
- [Looking Beyond The Top-1: Transformers Determine Top Tokens In Order](https://mlsanity.com/paper/2410.20210/) (2024)

## Built on by (most cited first, 16 of 16)

- [Cornerstones or Stumbling Blocks? Deciphering the Rock Tokens in On-Policy Distillation](https://mlsanity.com/paper/2605.09253/) (2026)
- [Inverse Depth Scaling From Most Layers Being Similar](https://mlsanity.com/paper/2602.05970/) (2026)
- [Good Arguments Against the People Pleasers: How Reasoning Mitigates (Yet Masks) LLM Sycophancy](https://mlsanity.com/paper/2603.16643/) (2026)
- [Deterministic Differentiable Structured Pruning for Large Language Models](https://mlsanity.com/paper/2603.08065/) (2026)
- [Reasoning-preserved Efficient Distillation of Large Language Models via Activation-aware Initialization](https://mlsanity.com/paper/2605.29327/) (2026)
- [Deeper is Not Always Better: Mitigating the Alignment Tax via Confident Layer Decoding](https://mlsanity.com/paper/2606.21906/) (2026)
- [PAS : Prelim Attention Score for Detecting Object Hallucinations in Large Vision--Language Models](https://mlsanity.com/paper/2511.11502/) (2025)
- [RBCorr: Response Bias Correction in Language Models](https://mlsanity.com/paper/2602.12445/) (2026)
- [OISD: On-Policy Internal Self-Distillation of Language Models](https://mlsanity.com/paper/2605.29089/) (2026)
- [Attention Degradation, Function Token Anchoring, and the Limits of Attention-Based Intervention in Large Language Models](https://mlsanity.com/paper/2607.20524/) (2026)
- [Do Transformers Use their Depth Adaptively? Evidence from a Relational Reasoning Task](https://mlsanity.com/paper/2604.12426/) (2026)
- [Child-directed speech facilitates production, not comprehension, in BabyLMs](https://mlsanity.com/paper/2606.01045/) (2026)
- [From Execution to Education: A Bloom-Aligned Framework for Measuring Educational Control in LLMs](https://mlsanity.com/paper/2607.08009/) (2026)
- [The Hard Decision Layer: Evidence for Committed Inference in Transformers](https://mlsanity.com/paper/2607.21613/) (2026)
- [Deep Learning Models Also Recall Features](https://mlsanity.com/paper/2608.20970/) (2026)
- [More Than Words: Compositional Tokenization for Efficient Language Models](https://mlsanity.com/paper/2610.05597/) (2026)

## Abstract (arXiv)

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 due to the lack of contextual information. As contextual information develops deeper into the model, these initial guesses get refined into contextually appropriate tokens. We then examine the dynamic usage of layer depth through three case studies. (i) Multiple-choice task analysis shows that the model identifies appropriate options within the first half of the model and finalizes the response in the latter half. (ii) Fact recall task analysis shows that in a multi-token answer, the first token requires more computational depth than the rest. (iii) Part-of-speech analysis shows that function words are, on average, the earliest to be predicted correctly. To validate our results, we supplement probe-based analyses with causal manipulations in the form of activation patching and early-exiting experiments. 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.