Pith. sign in

REVIEW 5 cited by

What Would Elsa Do? Freezing Layers During Transformer Fine-Tuning

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 1911.03090 v1 pith:IQZWZ7AY submitted 2019-11-08 cs.CL

What Would Elsa Do? Freezing Layers During Transformer Fine-Tuning

classification cs.CL
keywords layersfinalfine-tunedlanguagemodelsneedtasksacross
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Pretrained transformer-based language models have achieved state of the art across countless tasks in natural language processing. These models are highly expressive, comprising at least a hundred million parameters and a dozen layers. Recent evidence suggests that only a few of the final layers need to be fine-tuned for high quality on downstream tasks. Naturally, a subsequent research question is, "how many of the last layers do we need to fine-tune?" In this paper, we precisely answer this question. We examine two recent pretrained language models, BERT and RoBERTa, across standard tasks in textual entailment, semantic similarity, sentiment analysis, and linguistic acceptability. We vary the number of final layers that are fine-tuned, then study the resulting change in task-specific effectiveness. We show that only a fourth of the final layers need to be fine-tuned to achieve 90% of the original quality. Surprisingly, we also find that fine-tuning all layers does not always help.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 5 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. A Causal Language Modeling Detour Improves Encoder Continued Pretraining

    cs.CL 2026-05 conditional novelty 7.0

    A temporary CLM phase followed by MLM decay during encoder continued pretraining outperforms standard MLM on biomedical tasks by 0.3-2.8pp across languages and model sizes.

  2. BID-LoRA: A Parameter-Efficient Framework for Continual Learning and Unlearning

    cs.LG 2026-04 unverdicted novelty 6.0

    BID-LoRA uses bi-directional low-rank adapters with retain/new/unlearn pathways and escape unlearning to enable continual learning and unlearning while minimizing knowledge leakage and parameter updates.

  3. Vision Transformer Finetuning Benefits from Non-Smooth Components

    cs.LG 2026-02 conditional novelty 6.0

    For vision transformers, components with higher input-output sensitivity (attention and feedforward layers) yield better and more stable fine-tuning accuracy than smoother LayerNorm components.

  4. Task-agnostic Low-rank Residual Adaptation for Efficient Federated Continual Fine-Tuning

    cs.LG 2025-05 unverdicted novelty 6.0

    Fed-TaLoRA uses task-agnostic low-rank residual adaptation with post-aggregation calibration to enable efficient federated continual fine-tuning across sequential tasks under non-IID conditions.

  5. Landsat-Sentinel-2 Algal Bloom Mapping Using Vision Transformers: Model Description, Implementation, and Examples

    cs.CV 2026-06 unverdicted novelty 5.0

    Vision transformers trained on a new global dataset of Landsat-Sentinel-2 patches detect floating coastal algal blooms with 8-65% omission/commission error and outperform spectral indices under cloud and glint conditions.