Pith. sign in

REVIEW 1 cited by

Rethinking the Value of Transformer Components

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 2011.03803 v1 pith:SLAUXIZT submitted 2020-11-07 cs.CL

Rethinking the Value of Transformer Components

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

Transformer becomes the state-of-the-art translation model, while it is not well studied how each intermediate component contributes to the model performance, which poses significant challenges for designing optimal architectures. In this work, we bridge this gap by evaluating the impact of individual component (sub-layer) in trained Transformer models from different perspectives. Experimental results across language pairs, training strategies, and model capacities show that certain components are consistently more important than the others. We also report a number of interesting findings that might help humans better analyze, understand and improve Transformer models. Based on these observations, we further propose a new training strategy that can improves translation performance by distinguishing the unimportant components in training.

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. EvolKV: Evolutionary KV Cache Compression for LLM Inference

    cs.LG 2025-09 conditional novelty 6.0

    CMA-ES search over per-layer KV cache budgets beats uniform and pyramidal compression heuristics on LongBench, NIAH, RULER, and GSM8K, and edges past the full cache on one code dataset at 1.5% of the budget.