Research map / Architectures and efficiency / RNNs and normalization
LLM Layers Immediately Correct Each Other
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Identifies the Transformer Layer Correction Mechanism, where adjacent layers partially reverse each other's residual-stream contributions, and characterizes it across model families, training checkpoints, tokens, and via Jacobian analysis.
Problem
Feature-based interpretability (SAEs, transcoders, logit lens) assumes residual stream contributions build additively across layers; the paper tests whether adjacent layers instead systematically cancel each other's contributions.
Approach
Adjacent transformer layers show systematic negative cosine similarity between their residual-stream contributions (layer i+1 partially reverses layer i). Causal scaling interventions and layer Jacobian eigendecomposition show this correction is adaptive and selective, motivating a propose-and-reject interpretation.
Key results
- TLCM appears in 5 of 7 open-weight model families, with adjacent-layer cosine similarities typically -0.3 to -0.15 and as low as -0.4.
- TLCM is absent at OLMo 7B step 33,000 and emerges around step 135,500 (about 0.5T tokens) during pretraining.
- On Llama 3.1 8B, the newline token averages 12.41 TLCM activations versus 7.63 for 'sustainability'; activations grow about 1 per 1000 tokens of context.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Compare adjacent transformer-layer contributions during ordinary inference | Open-weight LLMs · WikiText | adjacent-layer cosine similarity: ≈ -0.2; typically -0.3 to -0.15, as low as -0.4 | Mixed |
| Train through pretraining | OLMo 7B · OLMo 7B pretraining checkpoints | TLCM presence: step 33,000 (5% of training) -> step 135,500 (21% of training, approximately 0.5 trillion tokens) | Helped |
| Increase available context | Llama 3.1 8B · WikiText context positions | TLCM activations per token: every 1000 tokens of additional context corresponds to approximately 1 additional TLCM activation | Helped |
| Use context-dependent tokens rather than common English words | Llama 3.1 8B · Generated responses | mean TLCM activations per token: 12.41 (± 0.10 SEM) vs 7.63 (± 0.23 SEM) | Mixed |
| Scale the preceding layer's contribution upward | TLCM-active adjacent layers · Llama 3.1 8B | correction strength: α ∈ [-1, 1]; α = 1 | Helped |
| Remove the preceding layer's contribution with α = -1 | TLCM-active adjacent layers · Llama 3.1 8B | cosine similarity between adjacent-layer contributions: α = -1 | Mixed |
| Project the preceding-layer contribution onto TLCM Jacobian directions | TLCM interactions · TLCM Jacobian pairs | eigenvalue: near λ = -1; around λ = 0.5 | Mixed |
| Replace pre-LayerNorm with post-LayerNorm and QK norm | OLMo 2 · OLMo 2 models | TLCM presence | No clear effect |
| Add post-LayerNorm to pre-LayerNorm | Gemma 2 · Gemma 2 | TLCM presence | No clear effect |
| Use GPT-2 or Microsoft Phi rather than model families with TLCM | Open-weight LLMs · WikiText | TLCM presence: 5 of 7 model families; TLCM is absent in GPT-2 and the Microsoft Phi models | Hurt |
Surprising findings
- Adjacent-layer contributions are negatively correlated in the residual stream, contrary to the common additive-enrichment view.
- TLCM correction capacity diminishes when the prior layer is amplified beyond about 2x.
Limitations
- The propose-and-reject framework is conceptual and not explicitly tested.
- Jacobian analysis is a local linearization with about 5% error only for small perturbations.
- Causal cause of TLCM (architectural vs learned) is not fully established; LayerNorm blindness alone does not explain it.
Experimental setup
Datasets: WikiText PyTorch instructional content (about 500 tokens)
Metrics: cosine similarity between layer contributions TLCM activation count per token Jacobian eigenvalue distribution Welch's t-test
Baselines: GPT-2 Phi
Builds on: sparse autoencoders cross-layer transcoders self-repair logit lens
Closest papers by meaning
- Analyze Feature Flow to Enhance Interpretation and Steering in Language Models (ICML 2025 · similarity 0.91)
- Layer by Layer: Uncovering Hidden Representations in Language Models (ICML 2025 · similarity 0.89)
- Sparsity-Guided Holistic Explanation for LLMs with Interpretable Inference-Time Intervention (AAAI 2024 · similarity 0.89)
- Remarkable Robustness of LLMs: Stages of Inference? (NeurIPS 2025 · similarity 0.89)
- Talking Heads: Understanding Inter-Layer Communication in Transformer Language Models (NeurIPS 2024 · similarity 0.88)
- The Truth is in There: Improving Reasoning in Language Models with Layer-Selective Rank Reduction (ICLR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.