# Gathered, Not Admitted: How Attention Brings a Latent Variable into Verbalizable Form

arXiv: https://arxiv.org/abs/2608.15022 · HTML: https://mlsanity.com/paper/2608.15022/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 22 of its 24 references that are in the map.
- First posted on arXiv 15 Aug 2026.

## Builds on (most cited first, 22 of 22)

- [Locating and Editing Factual Associations in GPT](https://mlsanity.com/paper/2202.05262/) (2022)
- [No Language Left Behind: Scaling Human-Centered Machine Translation](https://mlsanity.com/paper/2207.04672/) (2022)
- [Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small](https://mlsanity.com/paper/2211.00593/) (2022)
- [Eliciting Latent Predictions from Transformers with the Tuned Lens](https://mlsanity.com/paper/2303.08112/) (2023)
- [Dissecting Recall of Factual Associations in Auto-Regressive Language Models](https://mlsanity.com/paper/2304.14767/) (2023)
- [Towards Best Practices of Activation Patching in Language Models: Metrics and Methods](https://mlsanity.com/paper/2309.16042/) (2023)
- [Function Vectors in Large Language Models](https://mlsanity.com/paper/2310.15213/) (2023)
- [Do Llamas Work in English? On the Latent Language of Multilingual Transformers](https://mlsanity.com/paper/2402.10588/) (2024)
- [Localizing Model Behavior with Path Patching](https://mlsanity.com/paper/2304.05969/) (2023)
- [Is This the Subspace You Are Looking for? An Interpretability Illusion for Subspace Activation Patching](https://mlsanity.com/paper/2311.17030/) (2023)
- [Verbalizable Representations Form a Global Workspace in Language Models](https://mlsanity.com/paper/2607.15495/) (2026)
- [Separating Tongue from Thought: Activation Patching Reveals Language-Agnostic Concept Representations in Transformers](https://mlsanity.com/paper/2411.08745/) (2024)
- [Causal Head Gating: A Framework for Interpreting Roles of Attention Heads in Transformers](https://mlsanity.com/paper/2505.13737/) (2025)
- [Prompt-Activation Duality: Improving Activation Steering via Attention-Level Interventions](https://mlsanity.com/paper/2605.10664/) (2026)
- [Endogenous Resistance to Activation Steering in Language Models](https://mlsanity.com/paper/2602.06941/) (2026)
- [Where Computation Lives Inside TabPFN: Causal Localisation of Attention Head Function](https://mlsanity.com/paper/2606.12917/) (2026)
- [How Alignment Routes: Localizing, Scaling, and Controlling Policy Circuits in Language Models](https://mlsanity.com/paper/2604.04385/) (2026)
- [Silent Alarm: A J-Space Protocol for Comparing Danger Recognition Across Models and Quantization Levels](https://mlsanity.com/paper/2607.12792/) (2026)
- [J-CoT: Chain-of-Thought in J-Space](https://mlsanity.com/paper/2607.21981/) (2026)
- [Learning to Chain Operations by Routing Information Through a Global Workspace](https://mlsanity.com/paper/2503.01906/) (2025)
- [When Activation Oracles Learn Not to Read: Concept-Specific Blind Spots in Fine-Tuned Oracles](https://mlsanity.com/paper/2607.23379/) (2026)
- [A Theory of Conditional Collapse under Low-Rank Weight-Space Ablations: I. The Single-Block Theory and Synthetic Validation](https://mlsanity.com/paper/2608.03620/) (2026)

## Abstract (arXiv)

Language models hold latent quantities in a form they can report on, and more of a quantity is present in that form when the task requires reusing it flexibly. What causes a representation to enter that form is open, and the word workspace invites an admission story: a gate that decides what gets in. Testing it on open-weight models with Jacobian lenses, over a benchmark whose five arms share an identical context, we find no gate where it predicts one. Demand raises a concept's lens visibility beyond what applying an operator to a supplied value produces: +0.050 [+0.045, +0.057] in percentile rank on our primary checkpoint, positive on all four we measure, though that arm answers at ceiling and the accuracymatched contrast is stronger under that readout. At the same time one shared linear map decodes the variable from every arm, the control included, at 6.4-9.0x its selection-corrected floor. What produces the later readable form at the queried position is attention-mediated gathering inside a mid-depth window: separating patch depth from readout depth puts transport there at least 17x above anywhere shallower under non-saturating readouts, with no tested MLP output contributing positively inside it. Under the saturating percentile rank the same grid does not localise the window, which is a fact about that measure. An arm that needs the variable for nothing concentrates sevenfold less, so the window is demand-specific. That window has two measured edges, a survival failure below and destruction above, and it falls at the same fractional depth in a 64-layer hybrid and a 62-layer dense model from another family. We localise where the variable is installed and read, not the route from the passage, which transports nothing. But the readout is not a calibrated measure of use: three components move it to within 12% of one another and differ 7.4x in what they do to the answer.