LLM Layers Immediately Correct Each Other
Arjun Patrawala, Jiahai Feng, Erik Jones, Jacob Steinhardt
cs.CL · cs.LG
Quick verdict
LLM Layers Immediately Correct Each Other claims adjacent-layer propose-and-reject; the evidence is careful and broad, but 'correction' remains interpretation, not established function.
Summary
This NeurIPS 2025 paper challenges the view that transformer layers additively build persistent residual-stream features. It identifies TLCM: in 5 of 7 open-weight families (Llama 3, OLMo, Mistral, Gemma, Qwen2), adjacent layers' contribution vectors have negative cosine similarity (typically −0.3 to −0.15, as low as −0.4), indicating partial reversal of the previous layer's write. The phenomenon is absent in GPT-2 and Phi, appears across nearly all tokens, emerges gradually during OLMo pretraining checkpoints, fires most on contextually dependent tokens (numbers, dates, punctuation; newline averages 12.41 activations versus 7.63 for "sustainability"), and rises with context position (~1 extra activation per 1000 tokens). Perturbing the prior layer's contribution scales the next layer's correction roughly linearly, with compensation at −1 and saturation beyond ~2×. Layer-Jacobian eigendecomposition shows selective correction (eigenvalue mode near −1) and reinforcement (mode near 0.5) rather than uniform attenuation. The authors propose a propose-and-reject framework connecting TLCM to SAE feature low specificity, steering's extreme amplification demands, and transcoder advantages. Main limitations: propose-and-reject is explicitly untested, corrections are not semantically decomposed, and the mechanism is absent in GPT-2 and Phi families; the training-dynamics analysis relies on OLMo checkpoints with roughly 500 tokens each.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- In 5 of 7 open-weight model families, adjacent transformer layers write systematically opposing contribution vectors (adjacent-layer cosine similarity typically −0.3 to −0.15, as low as −0.4), a phenomenon the authors name TLCM; it is absent in GPT-2 and Phi.
- TLCM is learned during pretraining, not present at initialization: it is absent at OLMo 7B step 33,000 (5% of training) and emerges around step 135,500 (~21% of training, ~0.5T tokens).
- Layer i+1's correction adaptively depends on layer i's output: perturbing the input by α·a_t changes correction strength approximately linearly, with compensatory behavior at α=−1 and diminishing correction beyond ~2× amplification.
- Jacobian eigendecomposition shows selective correction (eigenvalue mode near −1) plus reinforcement (mode near 0.5) rather than uniform attenuation, with roughly 3000 of 4096 eigenvalues negative.
- The propose-and-reject framework explains three interpretability observations: low SAE feature-description specificity, ~10× steering amplification requirements, and the advantage of transcoders over SAEs.
Strengths
Simple, well-executed observational core with broad model coverage
The contribution-cosine matrices M[i,j] (Sec. 4.1, Fig. 1) are computed over ~100,000 WikiText tokens across 7 open-weight families, revealing consistent adjacent-layer negative similarity (−0.3 to −0.15, as low as −0.4) plus a bimodal per-pair distribution with a valley at −0.1 used to define activations.
Learned-versus-architectural question addressed directly
Sec. 4.2 shows no negative super-diagonal at OLMo 7B step 33,000 (5% of training), onset by step 135,500 (~0.5T tokens), strengthening through the final checkpoints across OLMo 7B, OLMo 2 1B, and OLMo 2 7B, ruling out pure initialization artifact and post-training origin.
Genuine causal intervention beyond correlation
Statistical calibration and reporting discipline
The Jacobian analysis is a genuine methodological step
Weaknesses
The central 'propose-and-reject' interpretation is untested
Geometric confounds not fully excluded
Token-level analysis lacks frequency/position controls
Key quantitative results rest on modest samples
Model-family absences remain unexplained; code availability is inconsistent
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper identifies and characterizes the Transformer Layer Correction Mechanism (TLCM): in 5 of 7 open-weight model families (Llama 3, OLMo, Mistral, Gemma, Qwen2), the marginal contribution vectors of adjacent layers have systematically negative cosine similarity (−0.3 to −0.15 typically, as low as −0.4; Sec. 4.1, Fig. 1), which the authors read as layer i+1 partially reversing layer i's write. TLCM is absent in GPT-2 and Phi. The claims span four registers: an observational phenomenon (well supported), a training-dynamics account (supported on OLMo checkpoints only), a causal adaptivity result (well supported), and a functional interpretation (propose-and-reject) that the authors explicitly leave untested (Sec. 5.3). The discussion links TLCM to three interpretability puzzles: SAE description low specificity (Templeton et al. 2024), ~10× steering amplification, and transcoder advantages (Paulo et al. 2025).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the Transformer Layer Correction Mechanism (TLCM) in one sentence?
It is the paper's name for the finding that in 5 of 7 open-weight LLM families (Llama 3, OLMo, Mistral, Gemma, Qwen2), each transformer layer's output systematically partially negates what the immediately preceding layer wrote, visible as negative cosine similarity (typically −0.3 to −0.15) between adjacent layers' contribution vectors — challenging the assumption that layers purely add features that later layers build upon.
Why do GPT-2 and Phi models not show TLCM, and does that undermine the finding?
The authors cannot fully explain their absence; they speculate it relates to architectural differences such as dropout placement after sublayers (Phi builds on GPT-2-style design despite a Llama 2 base). This means TLCM is not universal across transformer LLMs, so treat the finding as broad but not universal.
Is the 'propose-and-reject' hypothesis proven in this paper?
No. It is offered as a conceptual framework consistent with the evidence — layers write candidate features that the next layer, using attention, selectively removes. The authors explicitly state they did not test it, so it should be read as interpretation rather than established mechanism.
What does TLCM mean practically for steering or SAE interpretability work?
The paper suggests TLCM may explain why SAE feature descriptions are often unspecific and why steering requires extreme amplification (up to ~10×): corrections may actively remove amplified features. It proposes amplifying along directions the layer Jacobian shows are not corrected, though this remedy is not tested in the paper.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The provided text is a sampled excerpt with two [...sampling gap...] spans; hidden middle content may exist, and at least one sampled figure caption (Fig. A1, post-LayerNorm comparison) is present only as a caption with its figure omitted, so figure-level judgments should be treated cautiously.
- PDF extraction garbles some math notation; ranges such as α ∈ [−1, 1] and the 0.1-step grid in Fig. A4 were reconstructed from broken character sequences.
- Tables A1/A2 note the authors removed a few hundred mid-distribution token rows themselves, which is distinct from extraction truncation.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.