Residual Stream Analysis with Multi-Layer SAEs

• URL: http://arxiv.org/abs/2409.04185v3
• Date: Mon, 24 Feb 2025 09:18:36 GMT
• Title: Residual Stream Analysis with Multi-Layer SAEs
• Abstract summary: We introduce the multi-layer SAE (MLSAE), a single SAE trained on the residual stream activation vectors from every transformer layer.<n>We find that individual latents are often active at a single layer for a given token or prompt, but the layer at which an individual latent is active may differ for different tokens or prompts.<n>Our results represent a new approach to understanding how representations change as they flow through transformers.
• Score: 21.142967037533175
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Sparse autoencoders (SAEs) are a promising approach to interpreting the internal representations of transformer language models. However, SAEs are usually trained separately on each transformer layer, making it difficult to use them to study how information flows across layers. To solve this problem, we introduce the multi-layer SAE (MLSAE): a single SAE trained on the residual stream activation vectors from every transformer layer. Given that the residual stream is understood to preserve information across layers, we expected MLSAE latents to 'switch on' at a token position and remain active at later layers. Interestingly, we find that individual latents are often active at a single layer for a given token or prompt, but the layer at which an individual latent is active may differ for different tokens or prompts. We quantify these phenomena by defining a distribution over layers and considering its variance. We find that the variance of the distributions of latent activations over layers is about two orders of magnitude greater when aggregating over tokens compared with a single token. For larger underlying models, the degree to which latents are active at multiple layers increases, which is consistent with the fact that the residual stream activation vectors at adjacent layers become more similar. Finally, we relax the assumption that the residual stream basis is the same at every layer by applying pre-trained tuned-lens transformations, but our findings remain qualitatively similar. Our results represent a new approach to understanding how representations change as they flow through transformers. We release our code to train and analyze MLSAEs at https://github.com/tim-lawson/mlsae.

Related papers

• LLM Layers Immediately Correct Each Other [42.36086603213463]
  Transformer Layer Correction Mechanism (TLCM) appears in 5 out of 7 major open-source model families and activates across nearly all tokens in diverse texts.<n>We show that TLCM emerges during pretraining, operates most strongly on contextually dependent tokens, and adaptively calibrates its correction strength based on the preceding layer's output.<n>This dynamic suggests that the residual stream at any layer contains transient proposals alongside persistent features.
  arXiv  Detail & Related papers  (2026-09-07T18:36:39Z)
• ReSAE: Residualized Sparse Autoencoders for Multi-Layer Transformer Interventions [10.217703037524824]
  We introduce Residualized Sparse Autoencoders (ReSAEs)<n>ReSAEs fit an affine map between selected layers and train each later-layer SAE on the unexplained residual rather than on the full activation.<n>On Pythia-1.4B and Gemma-2-9B, residualization reduces decoder redundancy and improves sparse probing and targeted perturbations.
  arXiv  Detail & Related papers  (2026-05-27T01:27:53Z)
• One Model, Many Budgets: Elastic Latent Interfaces for Diffusion Transformers [80.19461768457622]
  Elastic Latent Interface Transformer (ELIT) is a drop-in, DiT-compatible mechanism that decouples input image size from compute.<n>Read and Write cross-attention layers move information between spatial tokens and latents and prioritize important input regions.<n>On ImageNet-1K 512px, ELIT delivers an average gain of $35.3%$ and $39.6%$ in FID and FDD scores.
  arXiv  Detail & Related papers  (2026-03-12T17:57:04Z)
• Do All Individual Layers Help? An Empirical Study of Task-Interfering Layers in Vision-Language Models [51.754991950934375]
  In a pretrained VLM, all layers are engaged by default to make predictions on downstream tasks.<n>We find that intervening on a single layer, such as by zeroing its parameters, can improve the performance on certain tasks.<n>We propose TaLo, a training-free, test-time adaptation method that dynamically identifies and bypasses the most interfering layer for a given task.
  arXiv  Detail & Related papers  (2026-02-01T11:37:05Z)
• A BERTology View of LLM Orchestrations: Token- and Layer-Selective Probes for Efficient Single-Pass Classification [2.0069888187253615]
  Production LLM systems often rely on separate models for safety and other classification-heavy steps.<n>We instead reuse computation already paid for by the serving LLM: we train lightweight probes on its hidden states and predict labels in the same forward pass used for generation.
  arXiv  Detail & Related papers  (2026-01-19T18:40:29Z)
• Direct Multi-Token Decoding [24.347862297812977]
  We introduce Direct Multi-Token Decoding (DMTD) as an inference paradigm for large language models (LLMs)<n>Unlike speculative decoding, our method introduces no additional parameters, auxiliary routines, or post-generation verification.<n>A fine-tuned DMTD Qwen3-4B model has already demonstrated promising results, achieving up to a 2x speedup with only minor performance loss.
  arXiv  Detail & Related papers  (2025-10-13T21:42:37Z)
• Diversity of Transformer Layers: One Aspect of Parameter Scaling Laws [42.926341529639274]
  Transformers deliver outstanding performance across a wide range of tasks.<n>Their task-solving performance is improved by increasing parameter size.<n>This study focuses on layers and their size, which mainly decide the parameter size of Transformers.
  arXiv  Detail & Related papers  (2025-05-29T21:13:31Z)
• Intermediate Layer Classifiers for OOD generalization [17.13749013546228]
  In this work, we question the use of last-layer representations for out-of-distribution (OOD) generalisation. We discover that intermediate layer representations frequently offer substantially better generalisation than those from the penultimate layer. Our analysis suggests that intermediate layers are less sensitive to distribution shifts compared to the penultimate layer.
  arXiv  Detail & Related papers  (2025-04-07T19:50:50Z)
• Adaptive Layer-skipping in Pre-trained LLMs [27.938188248731038]
  FlexiDepth is a method that dynamically adjusts the number of Transformer layers used in text generation. By incorporating a plug-in router and adapter, FlexiDepth enables adaptive layer-skipping in large language models.
  arXiv  Detail & Related papers  (2025-03-31T07:20:58Z)
• Multimodal Latent Language Modeling with Next-Token Diffusion [111.93906046452125]
  Multimodal generative models require a unified approach to handle both discrete data (e.g., text and code) and continuous data (e.g., image, audio, video) We propose Latent Language Modeling (LatentLM), which seamlessly integrates continuous and discrete data using causal Transformers.
  arXiv  Detail & Related papers  (2024-12-11T18:57:32Z)
• FIRP: Faster LLM inference via future intermediate representation prediction [54.897493351694195]
  FIRP generates multiple tokens instead of one at each decoding step. We conduct extensive experiments, showing a speedup ratio of 1.9x-3x in several models and datasets.
  arXiv  Detail & Related papers  (2024-10-27T15:53:49Z)
• Value Residual Learning For Alleviating Attention Concentration In Transformers [14.898656879574622]
  stacking multiple attention layers leads to attention concentration. One natural way to address this issue is to use cross-layer attention, allowing information from earlier layers to be directly accessible to later layers. We propose Transformer with residual value (ResFormer) which approximates cross-layer attention through adding a residual connection from the values of the the first layer to all subsequent layers.
  arXiv  Detail & Related papers  (2024-10-23T14:15:07Z)
• Unveiling Induction Heads: Provable Training Dynamics and Feature Learning in Transformers [54.20763128054692]
  We study how a two-attention-layer transformer is trained to perform ICL on $n$-gram Markov chain data. We prove that the gradient flow with respect to a cross-entropy ICL loss converges to a limiting model.
  arXiv  Detail & Related papers  (2024-09-09T18:10:26Z)
• Implicit Multimodal Alignment: On the Generalization of Frozen LLMs to Multimodal Inputs [63.29737699997859]
  Large Language Models (LLMs) have demonstrated impressive performance on multimodal tasks, without any multimodal finetuning. In this work, we expose frozen LLMs to image, video, audio and text inputs and analyse their internal representation.
  arXiv  Detail & Related papers  (2024-05-26T21:31:59Z)
• Manifold-Preserving Transformers are Effective for Short-Long Range Encoding [39.14128923434994]
  Multi-head self-attention-based Transformers have shown promise in different learning tasks. We propose TransJect, an encoder model that guarantees a theoretical bound for layer-wise distance preservation between a pair of tokens.
  arXiv  Detail & Related papers  (2023-10-22T06:58:28Z)
• JoMA: Demystifying Multilayer Transformers via JOint Dynamics of MLP and Attention [36.737750120893516]
  We propose Joint/Attention (JoMA) dynamics, a novel mathematical framework to understand the training procedure of multilayer Transformers. JoMA predicts that the attention first becomes sparse (to learn salient tokens), then dense (to learn less salient tokens) in the presence of nonlinear activations. We leverage JoMA to explain how tokens are combined to form hierarchies in multilayer Transformers, when the input tokens are generated by a latent hierarchical generative model.
  arXiv  Detail & Related papers  (2023-10-01T01:21:35Z)
• Learnable Polyphase Sampling for Shift Invariant and Equivariant Convolutional Networks [120.78155051439076]
  LPS can be trained end-to-end from data and generalizes existing handcrafted downsampling layers. We evaluate LPS on image classification and semantic segmentation.
  arXiv  Detail & Related papers  (2022-10-14T17:59:55Z)
• Layer Reduction: Accelerating Conformer-Based Self-Supervised Model via Layer Consistency [31.572652956170252]
  Transformer-based self-supervised models are trained as feature extractors and have empowered many downstream speech tasks to achieve state-of-the-art performance. We experimentally achieve 7.8X parameter reduction, 41.9% training speedup and 37.7% inference speedup while maintaining comparable performance with conventional BERT-like self-supervised methods.
  arXiv  Detail & Related papers  (2021-04-08T08:21:59Z)
• Transformer Feed-Forward Layers Are Key-Value Memories [49.52087581977751]
  We show that feed-forward layers in transformer-based language models operate as key-value memories. We show that the learned patterns are human-interpretable, and that lower layers tend to capture shallow patterns, while upper layers learn more semantic ones.
  arXiv  Detail & Related papers  (2020-12-29T19:12:05Z)
• Understanding Self-supervised Learning with Dual Deep Networks [74.92916579635336]
  We propose a novel framework to understand contrastive self-supervised learning (SSL) methods that employ dual pairs of deep ReLU networks. We prove that in each SGD update of SimCLR with various loss functions, the weights at each layer are updated by a emphcovariance operator. To further study what role the covariance operator plays and which features are learned in such a process, we model data generation and augmentation processes through a emphhierarchical latent tree model (HLTM)
  arXiv  Detail & Related papers  (2020-10-01T17:51:49Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.