LLM Layers Immediately Correct Each Other

• URL: http://arxiv.org/abs/2609.07876v1
• Date: Mon, 07 Sep 2026 18:36:39 GMT
• Title: LLM Layers Immediately Correct Each Other
• Abstract summary: 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.
• Score: 42.36086603213463
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Recent methods in language model interpretability employ techniques such as sparse autoencoders to decompose residual stream contributions into linear, semantically meaningful features. Such methods are commonly interpreted as identifying features that persist in the residual stream and that subsequent layers build upon. We challenge this view by identifying the Transformer Layer Correction Mechanism (TLCM), wherein adjacent transformer layers systematically counteract portions of each other's contributions. TLCM appears in 5 out of 7 major open-source model families and activates across nearly all tokens in diverse texts. 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. Using the layer Jacobian, we further show that TLCM selectively corrects specific subspaces while reinforcing others, which we interpret through a ``propose-and-reject'' framework in which layers propose candidate features and subsequent layers selectively remove inappropriate ones. This dynamic suggests that the residual stream at any layer contains transient proposals alongside persistent features, helping explain why SAE feature descriptions often have low specificity, why effective model steering requires extreme feature amplification, and why transcoders hold a theoretical advantage over SAEs.

Related papers

• Adaptive Depth Sparse Framework: Similarity-Driven Resource Allocation for Pre-Trained LLMs [18.457250148137494]
  We present an Adaptive Depth Sparse Framework (AdaDSF) that converts off-the-shelf pre-trained large language models into depth-sparse models without retraining.<n>On GPT-NeoX and Qwen2.5 over language modeling and commonsense reasoning, AdaDSF substantially reduces inference FLOPs while preserving performance close to dense counterparts.
  arXiv  Detail & Related papers  (2026-07-23T13:13:04Z)
• Analysis-by-Proxy: Localization Signals in VLMs Operating as Condition Encoders [54.06522068041745]
  Vision-Language Models (VLMs) are increasingly utilized as the conditioning backbone for diffusion-based image editing.<n>While standalone VLMs demonstrate strong localization capabilities, editing pipelines frequently struggle to maintain this accuracy.<n>In this work, we investigate this performance gap, hypothesizing that it stems from treating the VLM as a condition encoder.
  arXiv  Detail & Related papers  (2026-07-07T16:11:13Z)
• Deeper is Not Always Better: Mitigating the Alignment Tax via Confident Layer Decoding [92.1521161575198]
  We introduce Confident Decoding, a training-free decoding strategy that dynamically selects the most reliable near-final layer.<n> Experiments across dense and Mixture-of-Experts LLMs demonstrate consistent gains on challenging reasoning benchmarks.
  arXiv  Detail & Related papers  (2026-06-20T07:03:26Z)
• DifFRACT: Diffusion Feature Reconstruction and Attribution for Circuit Tracing [8.306039957986394]
  We extend transcoder-based circuit tracing to multimodal diffusion transformers.<n>We find mechanisms underlying attribute binding and cross-stream semantic propagation.<n>Our results provide causal explanations for systematic generation errors.
  arXiv  Detail & Related papers  (2026-06-14T13:02:44Z)
• Towards Fine-Grained Vision-Language Alignment for Few-Shot Anomaly Detection [65.29550320117526]
  We propose a novel framework named FineGrainedAD to improve anomaly localization performance.<n> Experiments demonstrate that the proposed FineGrainedAD achieves superior overall performance in few-shot settings.
  arXiv  Detail & Related papers  (2025-10-30T13:09:00Z)
• Multimodal LLM-Guided Semantic Correction in Text-to-Image Diffusion [52.315729095824906]
  MLLM Semantic-Corrected Ping-Pong-Ahead Diffusion (PPAD) is a novel framework that introduces a Multimodal Large Language Model (MLLM) as a semantic observer during inference.<n>It performs real-time analysis on intermediate generations, identifies latent semantic inconsistencies, and translates feedback into controllable signals that actively guide the remaining denoising steps.<n>Extensive experiments demonstrate PPAD's significant improvements.
  arXiv  Detail & Related papers  (2025-05-26T14:42:35Z)
• Analyzing Finetuning Representation Shift for Multimodal LLMs Steering [56.710375516257876]
  We propose to map hidden states to interpretable visual and textual concepts.<n>This enables us to more efficiently compare certain semantic dynamics, such as the shift from an original and fine-tuned model.<n>We also demonstrate the use of shift vectors to capture these concepts changes.
  arXiv  Detail & Related papers  (2025-01-06T13:37:13Z)
• Residual Stream Analysis with Multi-Layer SAEs [21.142967037533175]
  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.
  arXiv  Detail & Related papers  (2024-09-06T11:01:55Z)
• Language Models as Hierarchy Encoders [22.03504018330068]
  We introduce a novel approach to re-train transformer encoder-based LMs as Hierarchy Transformer encoders (HiTs) Our method situates the output embedding space of pre-trained LMs within a Poincar'e ball with a curvature that adapts to the embedding dimension. We evaluate HiTs against pre-trained LMs, standard fine-tuned LMs, and several hyperbolic embedding baselines.
  arXiv  Detail & Related papers  (2024-01-21T02:29:12Z)
• Layer-wise Representation Fusion for Compositional Generalization [26.771056871444692]
  A key reason for failure on compositional generalization is that the syntactic and semantic representations of sequences in both the uppermost layer of the encoder and decoder are entangled. We explain why it exists by analyzing the representation evolving mechanism from the bottom to the top of the Transformer layers. Inspired by this, we propose LRF, a novel textbfLayer-wise textbfRepresentation textbfFusion framework for CG, which learns to fuse previous layers' information back into the encoding and decoding process.
  arXiv  Detail & Related papers  (2023-07-20T12:01:40Z)

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.