Hidden Decoding at Scale: Latent Computation Scaling for Large Language Models

2026-07-09

AI Summary

This work addresses the challenge of further enhancing large language model performance without modifying existing powerful Transformer backbones. It proposes Hidden Decoding (HD), a method that, during continued pretraining, expands each token into *n* independent computational streams. HD leverages multi-stream embeddings, stream-decomposed attention mechanisms, and key-value cache reuse to increase computation along the sequence length dimension, thereby augmenting model capacity. This approach establishes the first effective scaling pathway under a fixed backbone architecture, validated on hundred-billion-scale Mixture-of-Experts (MoE) models. Experimental results demonstrate that WeLM-HD4-80B and WeLM-HD4-617B significantly outperform their non-HD baselines, with performance consistently improving as the expansion factor *n* increases.

Abstract

Scaling Large Language Models (LLMs) has been driven mainly by enlarging the Transformer backbone, but for an already-strong model this requires another round of costly pretraining. We study whether an existing backbone can keep improving by allocating more computation to each token while leaving the Transformer backbone fixed. Depth-recurrent (looped) Transformers pursue this goal but are hard to scale, because looped computation does not fit naturally with the pipeline parallelism used to train the largest models. We add computation along the sequence-length dimension, where the extra computation is simply a longer input and stays compatible with standard large-model training. We propose Hidden Decoding, a sequence-length scaling method applied during continued pretraining (CPT). It expands each token into n streams with independent embedding tables and keeps the intermediate streams' key-value cache as context, so each token performs more internal computation without adding or widening Transformer layers. To keep this affordable at scale, we introduce Stream-Factorized Attention, in which most layers attend only within each stream and only a few layers mix across streams, reducing the attention cost from quadratic to roughly linear in n. Experiments support two scaling results. At frontier scale, we train WeLM-HD4-80B and WeLM-HD4-617B at n=4 and improve their matched non-HD baselines, making Hidden Decoding the first demonstrated sequence-length scaling method at the 100B+ MoE scale. Across expansion factors, the gains grow as n increases, showing that sequence-length expansion is a practical fixed-backbone scaling path for frontier-scale LLMs.

Problem

• computation scaling
• sequence-length scaling
• fixed-backbone scaling
• large language models
• latent computation

Innovation

• Hidden Decoding
• sequence-length scaling
• Stream-Factorized Attention
• fixed-backbone scaling
• continued pretraining