Tapered Language Models

Hugging Face Daily Papers Papers

Summary

This paper introduces Tapered Language Models (TLMs), an architecture principle that allocates more parameters to earlier layers and fewer to later layers, consistently improving perplexity and downstream performance across multiple architectures without extra cost.

Modern language models, including transformer, recurrent, and memory-based variants, share a common chassis: a stack of identical layers in which parameters are allocated uniformly across depth. This is a default inherited from the original transformer and largely unchanged since, yet a growing body of evidence suggests that layers contribute non-uniformly to the final output, with later layers refining the residual stream rather than transforming it. We ask whether parameter capacity should reflect this asymmetry. Our controlled experiment shows that, under a fixed budget, allocating more capacity to earlier layers and less to later layers improves perplexity over a uniform-width baseline, while the reverse allocation hurts. Building on this result, we introduce Tapered Language Models (TLMs), an architectural principle in which a parameter-bearing component is monotonically tapered across depth under a fixed total budget. MLPs are the natural site for this instantiation: they dominate parameter count across all modern LM families and expose width as a single, clean axis of variation. Across three model scales and four architectures (Transformer, Gated Attention, Hope-attention, and Titans), tapering MLP width via a smooth cosine schedule consistently improves perplexity and downstream benchmark performance over uniform baselines, at no additional parameter or compute cost. These findings establish depth-aware capacity allocation as a simple, architecture-agnostic axis of language model design, a free lever hidden in plain sight.
Original Article
View Cached Full Text

Cached at: 06/23/26, 05:40 AM

Paper page - Tapered Language Models

Source: https://huggingface.co/papers/2606.23670

Abstract

Tapered language models allocate more parameters to earlier layers and fewer to later layers, improving performance without increasing total parameters or compute costs.

Modern language models, includingtransformer,recurrent, andmemory-based variants, share a common chassis: a stack ofidentical layersin which parameters are allocated uniformly across depth. This is a default inherited from the originaltransformerand largely unchanged since, yet a growing body of evidence suggests that layers contribute non-uniformly to the final output, with later layers refining the residual stream rather than transforming it. We ask whether parameter capacity should reflect this asymmetry. Our controlled experiment shows that, under a fixed budget, allocating more capacity to earlier layers and less to later layers improvesperplexityover a uniform-width baseline, while the reverse allocation hurts. Building on this result, we introduce Tapered Language Models (TLMs), an architectural principle in which a parameter-bearing component is monotonically tapered across depth under a fixed total budget.MLPsare the natural site for this instantiation: they dominate parameter count across all modern LM families and expose width as a single, clean axis of variation. Across three model scales and four architectures (Transformer, Gated Attention, Hope-attention, and Titans), tapering MLP width via a smoothcosine scheduleconsistently improvesperplexityand downstream benchmark performance over uniform baselines, at no additional parameter or compute cost. These findings establishdepth-aware capacity allocationas a simple,architecture-agnosticaxis of language model design, a free lever hidden in plain sight.

View arXiv pageView PDFAdd to collection

Get this paper in your agent:

hf papers read 2606\.23670

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2606.23670 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2606.23670 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2606.23670 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Pretraining Language Models on Historical Text

arXiv cs.CL

This paper introduces TypewriterLM, a 7.24B parameter language model trained exclusively on English text predating 1913, along with TypewriterCorpus (a 54B-token cleaned historical corpus) and instruction-tuning datasets to avoid temporal leakage and lookahead bias. It also presents a benchmark suite, History-Event, for evaluating temporal grounding and leakage.