Pith. sign in

REVIEW 7 cited by

NormFormer: Improved Transformer Pretraining with Extra Normalization

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2110.09456 v2 pith:W2MQDI53 submitted 2021-10-18 cs.CL cs.AI

classification cs.CLcs.AI
keywords normformerlayerperformanceperplexitypretrainingcomputeextrafairseq
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

During pretraining, the Pre-LayerNorm transformer suffers from a gradient magnitude mismatch: gradients at early layers are much larger than at later layers. These issues can be alleviated by our proposed NormFormer architecture, which adds three normalization operations to each layer: a Layer Norm after self attention, head-wise scaling of self-attention outputs, and a Layer Norm after the first fully connected layer. The extra operations incur negligible compute cost (+0.4% parameter increase), but improve pretraining perplexity and downstream task performance for both causal and masked language models ranging from 125 Million to 2.7 Billion parameters. For example, adding NormFormer on top of our strongest 1.3B parameter baseline can reach equal perplexity 24% faster, or converge 0.27 perplexity better in the same compute budget. This model reaches GPT3-Large (1.3B) zero shot performance 60% faster. For masked language modeling, NormFormer improves fine-tuned GLUE performance by 1.9% on average. Code to train NormFormer models is available in fairseq https://github.com/pytorch/fairseq/tree/main/examples/normformer .

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 7 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. OpenAlex reports about 28 citations worldwide. Full citation record

  1. A Controlled Study of Attention-Only Transformers

    cs.LG 2026-07 conditional novelty 7.0 of 10

    Attention-only transformers match standard transformers within 0.006 nats of loss at matched parameter count, with the residual gap localized to low-context parametric recall.

  2. Transforming Rank: How Architecture Navigates the Spectral Pathologies of Depth

    cs.LG 2026-07 conditional novelty 7.0 of 10

    Rank survival in Transformer blocks is governed by a branch-to-skip ratio law (βα^M√L), a mean-spike coherence c_ℓ=E[σ]²/E[σ²], and a Marchenko–Pastur width threshold m/d=1/p(σ).

  3. Dual Attention Residuals

    cs.CL 2026-07 conditional novelty 6.0 of 10

    Reciprocal cross-stream depth retrieval improves Transformer pretraining loss across dense and MoE models from 0.1B to 7B.

  4. Enhancing next token prediction based pre-training for jet foundation models

    hep-ph 2025-12 conditional novelty 6.0 of 10

    Using continuous particle features as input and combining next-token with masked-token pre-training markedly improves classification accuracy of the OmniJet jet foundation model without visibly hurting its generative quality.

  5. SpanNorm: Reconciling Training Stability and Performance in Deep Transformers

    cs.CL 2026-01 conditional novelty 5.0 of 10

    SpanNorm—a block-level residual with PostNorm-style normalization—trains deeper transformers more stably and outperforms PreNorm and hybrid normalization on LM benchmarks.

  6. Residual Matrix Transformers: Scaling the Size of the Residual Stream

    cs.LG 2025-06 conditional novelty 5.0 of 10

    The RMT uses an outer product memory as the residual stream, allowing the residual stream to be scaled up at negligible cost and improving language model efficiency.

  7. Understanding Transformer from the Perspective of Associative Memory

    cs.LG 2025-05 conditional novelty 4.0 of 10

    Frames the Transformer as associative memory, derives retrieval SNR for linear, softmax, ReLU, and SoLU kernels, and proposes DeltaFormer, a softmax-plus-delta-rule variant claimed to exceed TC0 expressivity.

Pith tools