# Training, Reading, and Editing Legible Transformers

arXiv: https://arxiv.org/abs/2607.08946 · HTML: https://mlsanity.com/paper/2607.08946/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Builds on 63 of its 75 references that are in the map.
- First posted on arXiv 9 Jul 2026.

## Builds on (most cited first, 30 of 63)

- [Bootstrap your own latent: A new approach to self-supervised Learning](https://mlsanity.com/paper/2006.07733/) (2020)
- [Exploring Simple Siamese Representation Learning](https://mlsanity.com/paper/2011.10566/) (2020)
- [Locating and Editing Factual Associations in GPT](https://mlsanity.com/paper/2202.05262/) (2022)
- [Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation](https://mlsanity.com/paper/1308.3432/) (2013)
- [Barlow Twins: Self-Supervised Learning via Redundancy Reduction](https://mlsanity.com/paper/2103.03230/) (2021)
- [GLU Variants Improve Transformer](https://mlsanity.com/paper/2002.05202/) (2020)
- [Sparse Autoencoders Find Highly Interpretable Features in Language Models](https://mlsanity.com/paper/2309.08600/) (2023)
- [Transformer Feed-Forward Layers Are Key-Value Memories](https://mlsanity.com/paper/2012.14913/) (2020)
- [Representation Engineering: A Top-Down Approach to AI Transparency](https://mlsanity.com/paper/2310.01405/) (2023)
- [Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small](https://mlsanity.com/paper/2211.00593/) (2022)
- [Editing Models with Task Arithmetic](https://mlsanity.com/paper/2212.04089/) (2022)
- [Steering Llama 2 via Contrastive Activation Addition](https://mlsanity.com/paper/2312.06681/) (2023)
- [VICReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learning](https://mlsanity.com/paper/2105.04906/) (2021)
- [Concept Bottleneck Models](https://mlsanity.com/paper/2007.04612/) (2020)
- [Refusal in Language Models Is Mediated by a Single Direction](https://mlsanity.com/paper/2406.11717/) (2024)
- [Mass-Editing Memory in a Transformer](https://mlsanity.com/paper/2210.07229/) (2022)
- [The LAMBADA dataset: Word prediction requiring a broad discourse context](https://mlsanity.com/paper/1606.06031/) (2016)
- [Toy Models of Superposition](https://mlsanity.com/paper/2209.10652/) (2022)
- [The Linear Representation Hypothesis and the Geometry of Large Language Models](https://mlsanity.com/paper/2311.03658/) (2023)
- [KAN: Kolmogorov-Arnold Networks](https://mlsanity.com/paper/2404.19756/) (2024)
- [Eliciting Latent Predictions from Transformers with the Tuned Lens](https://mlsanity.com/paper/2303.08112/) (2023)
- [Towards Faithfully Interpretable NLP Systems: How should we define and evaluate faithfulness?](https://mlsanity.com/paper/2004.03685/) (2020)
- [BLiMP: The Benchmark of Linguistic Minimal Pairs for English](https://mlsanity.com/paper/1912.00582/) (2019)
- [Scaling and evaluating sparse autoencoders](https://mlsanity.com/paper/2406.04093/) (2024)
- [Transformer Feed-Forward Layers Build Predictions by Promoting Concepts in the Vocabulary Space](https://mlsanity.com/paper/2203.14680/) (2022)
- [Towards Robust Interpretability with Self-Explaining Neural Networks](https://mlsanity.com/paper/1806.07538/) (2018)
- [Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models](https://mlsanity.com/paper/2403.19647/) (2024)
- [Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2](https://mlsanity.com/paper/2408.05147/) (2024)
- [Towards Best Practices of Activation Patching in Language Models: Metrics and Methods](https://mlsanity.com/paper/2309.16042/) (2023)
- [In-Context Learning Creates Task Vectors](https://mlsanity.com/paper/2310.15916/) (2023)

## Built on by (most cited first, 1 of 1)

- [Through the Looking Glass: Directly Reading and Writing Transformers](https://mlsanity.com/paper/2609.10210/) (2026)

## Abstract (arXiv)

A transformer can be built from operators that are legible by construction -- bounded, named units that read as fuzzy set operations rather than dense activations -- but legibility must be pressed for during training, and the pressure has a failure mode. A crispness penalty meant to sharpen a bounded operator into a decisive detector instead collapses it into a dead constant. An identity, E[v(1-v)] = mu(1-mu) - var, shows why -- the penalty is a variance-minimizer blind to the difference between a live detector and a constant -- and names the fix: a per-channel variance floor, the target legibility metric written as a loss, which recovers both legibility and quality. A learned per-unit fraction then retires the hand-set reserved-GELU partition of prior work: given the choice the model keeps no unit as pure GELU and routes 87% of its load-bearing computation through crisp operators. The result is the most legible transformer we have built -- 78% of its feed-forward operands and 50% of its attention value channels are crisp-and-contextual detectors, and per-head legibility rises from 18% in shallow layers to 78% in deep ones. Read in the correct rotated per-layer frame, these units separate a clean detection (what a unit responds to) from a harder naming (what its output decodes to); and because the objective makes each unit crisp and sparse, edits to them are far more local -- 50-184x in the deep layers where the edit sites concentrate -- and can target explicit conjunctions a single neuron cannot express. Finally, a between-unit decorrelation pressure exposes a legibility dial: it trades a circuit's reuse for independence at no quality cost, turning concepts into single, surgically editable units and a prediction into a short explanation read off a handful of named operations. Quality holds at parity with a conventional baseline throughout.