# Through the Looking Glass: Directly Reading and Writing Transformers

arXiv: https://arxiv.org/abs/2609.10210 · HTML: https://mlsanity.com/paper/2609.10210/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Builds on 37 of its 60 references that are in the map.
- First posted on arXiv 9 Sep 2026.

## Builds on (most cited first, 30 of 37)

- [Locating and Editing Factual Associations in GPT](https://mlsanity.com/paper/2202.05262/) (2022)
- [Efficient Streaming Language Models with Attention Sinks](https://mlsanity.com/paper/2309.17453/) (2023)
- [Sparse Autoencoders Find Highly Interpretable Features in Language Models](https://mlsanity.com/paper/2309.08600/) (2023)
- [Inference-Time Intervention: Eliciting Truthful Answers from a Language Model](https://mlsanity.com/paper/2306.03341/) (2023)
- [Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small](https://mlsanity.com/paper/2211.00593/) (2022)
- [Steering Llama 2 via Contrastive Activation Addition](https://mlsanity.com/paper/2312.06681/) (2023)
- [Steering Language Models With Activation Engineering](https://mlsanity.com/paper/2308.10248/) (2023)
- [Eliciting Latent Predictions from Transformers with the Tuned Lens](https://mlsanity.com/paper/2303.08112/) (2023)
- [Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models](https://mlsanity.com/paper/2403.19647/) (2024)
- [Massive Activations in Large Language Models](https://mlsanity.com/paper/2402.17762/) (2024)
- [Attribution Patching Outperforms Automated Circuit Discovery](https://mlsanity.com/paper/2310.10348/) (2023)
- [Transcoders Find Interpretable LLM Feature Circuits](https://mlsanity.com/paper/2406.11944/) (2024)
- [Transformer-Patcher: One Mistake worth One Neuron](https://mlsanity.com/paper/2301.09785/) (2023)
- [All Bark and No Bite: Rogue Dimensions in Transformer Language Models Obscure Representational Quality](https://mlsanity.com/paper/2109.04404/) (2021)
- [The Hydra Effect: Emergent Self-repair in Language Model Computations](https://mlsanity.com/paper/2307.15771/) (2023)
- [Analyzing Transformers in Embedding Space](https://mlsanity.com/paper/2209.02535/) (2022)
- [AtP*: An efficient and scalable method for localizing LLM behaviour to components](https://mlsanity.com/paper/2403.00745/) (2024)
- [Rewriting a Deep Generative Model](https://mlsanity.com/paper/2007.15646/) (2020)
- [Tracr: Compiled Transformers as a Laboratory for Interpretability](https://mlsanity.com/paper/2301.05062/) (2023)
- [Augmenting Self-attention with Persistent Memory](https://mlsanity.com/paper/1907.01470/) (2019)
- [Copy Suppression: Comprehensively Understanding an Attention Head](https://mlsanity.com/paper/2310.04625/) (2023)
- [Codebook Features: Sparse and Discrete Interpretability for Neural Networks](https://mlsanity.com/paper/2310.17230/) (2023)
- [Weight-sparse transformers have interpretable circuits](https://mlsanity.com/paper/2511.13653/) (2025)
- [Explorations of Self-Repair in Language Models](https://mlsanity.com/paper/2402.15390/) (2024)
- [Optimal ablation for interpretability](https://mlsanity.com/paper/2409.09951/) (2024)
- [Language Model Circuits Are Sparse in the Neuron Basis](https://mlsanity.com/paper/2601.22594/) (2026)
- [Interpreting Context Look-ups in Transformers: Investigating Attention-MLP Interactions](https://mlsanity.com/paper/2402.15055/) (2024)
- [An Adversarial Example for Direct Logit Attribution: Memory Management in GELU-4L](https://mlsanity.com/paper/2310.07325/) (2023)
- [Does Knowledge Localization Hold True? Surprising Differences Between Entity and Relation Perspectives in Language Models](https://mlsanity.com/paper/2409.00617/) (2024)
- [Using the Path of Least Resistance to Explain Deep Networks](https://mlsanity.com/paper/2502.12108/) (2025)

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

- [Every Ablation Is a Dose: Counterweights and the Semblance of Self-Repair](https://mlsanity.com/paper/2610.02173/) (2026)

## Abstract (arXiv)

How many of a transformer's components decide a token? Counted by the absolute value of each unit's and channel's contribution to the logit, one prediction rests on thousands to hundreds of thousands of them. But contributions are signed, and across eighteen models the mass pushing away from the predicted token is a median of seven times the mass carrying it. Divide by the net and the count is dozens: on the baseline, 53 components carry ninety percent of a prediction, 13 it cannot survive losing, and 8 suffice to produce it alone. Across twelve models trained elsewhere, 124M to 7B parameters, the sufficient set runs from two components to sixteen, and what a prediction draws on, followed all the way back, is one to three percent of the model, a share that does not grow with size. Three quarters of a layer's update is a fixed linear map of the state it received. Everything is read from the model's own parameters and activations, with nothing trained or fitted, and it names a component on both sides: what it writes, from the predictions it drives, reaching close to half of every model; what it reads, from its weights in the frame of its own layer, at 58.9 percent above chance over its eight strongest inputs. Sorting the remainder by upstream source yields grammatical categories the embedding cannot see. A name can be acted on. An association the model does not hold installs into one spare unit, key and value read from the weights, for a quarter of a percent of held-out loss, a fortieth of what a rank-one update costs. An installed attention head and a unit two layers above it make an edit fire only where a token occurred earlier in the context, and a unit the model trained for itself is driven from two layers upstream, 86 percent of the effect passing through it. An order-preserving activation puts a unit's inputs at the instrument's ceiling, at the price of a two-part install.