Pith. sign in

REVIEW 4 cited by

Jump to Conclusions: Short-Cutting Transformers With Linear Transformations

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 2303.09435 v2 pith:DVVDNV45 submitted 2023-03-16 cs.CL

Jump to Conclusions: Short-Cutting Transformers With Linear Transformations

classification cs.CL
keywords representationslayershiddenmethodfinalbertearlygpt-2
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-based language models create hidden representations of their inputs at every layer, but only use final-layer representations for prediction. This obscures the internal decision-making process of the model and the utility of its intermediate representations. One way to elucidate this is to cast the hidden representations as final representations, bypassing the transformer computation in-between. In this work, we suggest a simple method for such casting, using linear transformations. This approximation far exceeds the prevailing practice of inspecting hidden representations from all layers, in the space of the final layer. Moreover, in the context of language modeling, our method produces more accurate predictions from hidden layers, across various model scales, architectures, and data distributions. This allows "peeking" into intermediate representations, showing that GPT-2 and BERT often predict the final output already in early layers. We then demonstrate the practicality of our method to recent early exit strategies, showing that when aiming, for example, at retention of 95% accuracy, our approach saves additional 7.9% layers for GPT-2 and 5.4% layers for BERT. Last, we extend our method to linearly approximate sub-modules, finding that attention is most tolerant to this change. Our code and learned mappings are publicly available at https://github.com/sashayd/mat.

discussion (0)

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

Forward citations

Cited by 4 Pith papers

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

  1. The Right Answer, the Wrong Direction: Why Transformers Fail at Counting and How to Fix It

    cs.LG 2026-05 unverdicted novelty 7.0

    Transformers encode counts correctly internally but fail to read them out due to misalignment with digit output directions, fixable by updating 37k output parameters or small LoRA on attention.

  2. The Right Answer, the Wrong Direction: Why Transformers Fail at Counting and How to Fix It

    cs.LG 2026-05 accept novelty 7.0

    Transformers store count information internally but cannot read it out as digits due to near-orthogonal alignment with output-head rows; updating digit rows or applying LoRA to attention layers improves constrained an...

  3. Eliciting Latent Predictions from Transformers with the Tuned Lens

    cs.LG 2023-03 accept novelty 7.0

    Training per-layer affine probes on frozen transformers yields more reliable latent predictions than the logit lens and enables detection of malicious inputs from prediction trajectories.

  4. How Transformers Reject Wrong Answers: Rotational Dynamics of Factual Constraint Processing

    cs.CL 2026-02 reject novelty 6.0

    Correct and incorrect single-token continuations of factual queries are separated by rotation of displacement vectors in transformer hidden states, with larger models also suppressing the correct token when forced to ...