💬 NLP

One mechanism for many mental spaces: a shared router over a value slot in language models

This paper demonstrates that transformer language models implement a unified mechanism for diverse mental spaces (such as beliefs, fictions, and counterfactuals) using a shared, low-rank router that selects values from a single reusable slot, rather than employing distinct logical structures for each context type.

Original authors: Oliver Steele, Jiangtao Wen, Yuxing Han

Published 2026-07-14

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Oliver Steele, Jiangtao Wen, Yuxing Han

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a language model as a super-organized librarian who doesn't just store books, but stores entire worlds inside a single library. In this library, a flower can be red in the "Real World" section, purple in the "Portrait" section, and green in "Alice's Dream" section. The big question was: Does the librarian keep these worlds in totally separate, magical rooms with different rules for each? Or does the librarian use one clever, reusable system to handle them all?

This paper suggests the librarian uses one shared system for everything.

The Magic Tag and the Router

Think of the flower's color as a piece of data sitting on a shelf. The paper finds that the model uses a special, low-rank "tag" (called a router or space index) to select which world to read. This tag is separate from the data itself. The data sits in a neutral storage area called a value slot, which holds the color but doesn't care if it belongs to reality or a dream.

But how does the model know which color to pick when you ask, "What color is the flower in the portrait?" It uses the router. You can think of this router as a smart, causally manipulable index. When you ask about the portrait, the router selects the "Portrait" setting, and the model reads the purple flower from the shared slot. When you ask about reality, the router selects "Reality," and it reads the red flower.

The paper shows that this router isn't a different machine for every type of world. Whether the world is a belief (what Alice thinks), a memory (yesterday), a hypothetical (what if...), or a fiction (in a movie), the model uses the same type of router.

One Switch to Rule Them All

The researchers tested this by training the model to control just one type of world, like "hypotheticals" (things that could happen). Then, they tried using that same control mechanism on other worlds, like "beliefs" or "movies."

The result? The switch worked for everything.

When they trained the model to flip the color in a hypothetical scenario, that same training flipped the color in a belief scenario about 71% to 89% of the time (depending on the model family).
This suggests that the model doesn't have a special, secret "belief machine" that is totally different from its "movie machine." Instead, it has one shared router that handles all these different contexts.

What It's NOT (The Ruled-Out Ideas)

The paper is very careful to say what this system is not:

It's not just remembering words: The model isn't just memorizing that the word "red" appeared near "reality" and "purple" near "portrait." The system actually carries a code for the context itself.
It's not a separate room for beliefs: Formal logic often treats "beliefs" as a totally different, opaque category where rules change. This paper argues that, mechanically, the model treats beliefs just like any other space. The "belief" tag isn't special or separated from the "hypothetical" tag in the model's wiring.
It's not a static sticky note: The model doesn't write the color down on a permanent note and leave it there. Instead, it's more like a retrieval system. When you ask a question, the model goes back to the source text and re-reads the color based on the current "space" setting. It's a "just-in-time" read, not a permanent storage slot.

The "Report" vs. "Reasoning" Split

Here is a fun twist the paper discovered. The model has two different paths for the same information:

The Report Path: This is what the model says out loud.
The Reasoning Path: This is what the model uses to solve a puzzle.

The researchers found they could control these paths separately. They could train the model to say "The flower is green" (flipping the report) while still making the model solve a logic puzzle as if the flower were "blue" (leaving the reasoning unchanged). This proves the model keeps the "what I say" and "what I think" in slightly different parts of its brain, even though they come from the same source.

Building New Worlds

When the model encounters a complex sentence like "Alice used to believe the cup was blue," it's building a new space (a mix of "past" and "belief"). The paper shows that the model doesn't build a whole new library for this. Instead, it reuses the same value slot (the cup's color) but mints a fresh router (a new index setting) specifically for this new, combined world. It's like taking an existing book and giving it a new, temporary cover label without rewriting the whole story.

How Sure Are We?

The paper is quite confident about these findings, but with specific limits:

Proven in the lab: The results are based on direct measurements of the model's internal "activations" (its electrical signals) in three different families of models (Qwen, Pythia, and Falcon3). The researchers didn't just guess; they physically intervened in the model's code to prove the router causes the behavior.
Not a magic bullet: The paper admits that while the router is shared, it's not perfectly identical for every single type of world. For example, "painting" or "movie" contexts sometimes act a little differently than "belief" contexts. Also, the model's ability to handle these worlds is "retrieval-shaped," meaning it relies on reading the text again rather than holding a static memory, which makes it slightly "leaky" in some tests.
Scale matters: These mechanical details were found in smaller models (3 billion parameters). However, the behavior (the ability to handle these worlds) was confirmed to exist in massive, frontier models (like GPT-4o and 72B models), suggesting the mechanism scales up even if we can't peek inside those giant models as easily.

In short, the paper suggests that language models have a unified, reusable "space router" that lets them juggle reality, dreams, beliefs, and movies using the same mechanical tool, rather than needing a different tool for every single type of imagination.

Technical Summary: One Mechanism for Many Mental Spaces

Problem Statement
Language models (LMs) must manage discourse contexts that diverge from reality, such as beliefs, memories, hypotheticals, and depictions (e.g., a flower that is red in reality but purple in a portrait). Formal semantics treats these as distinct logical domains (modal, temporal, doxastic, depictive) with different inference rules. Conversely, Fauconnier's mental-space theory posits a unified "space-building" operation where discourse creates partial spaces and binds values to them. The central question is whether transformer LMs implement these as separate semantic modules or as a unified mechanistic operation. Specifically, does the model maintain a dedicated code for the local discourse context (a "space index") that is causally distinct from the entity values themselves, and is this code shared across different types of non-actual contexts?

Methodology
The study employs mechanistic interpretability techniques on open-weight transformer models (primarily Qwen2.5-3B, with replication on Pythia-2.8b and Falcon3-3B) using controlled natural-language stimuli.

Stimuli: Items present an entity with two conflicting attribute values (e.g., color) across a "base" world (reality) and an "alternative" space (e.g., a painting, a belief, a past state). The model is queried at a later token to retrieve the value for a specific space.
Localization: The authors use linear probing to verify that both base and alternative values are decodable from residual stream activations at the query token.
Causal Intervention:
- Distributed Alignment Search (DAS): To overcome the failure of hand-built steering vectors, the authors train low-rank orthonormal subspaces ( $R \in \mathbb{R}^{d \times k}$ ) using an interchange objective. This subspace is optimized to flip the model's output from the base value to the alternative value when the subspace component is spliced from a donor item.
- Double Dissociation: The authors test whether the value the model reports and the value it uses for reasoning (applying a sorting rule) are controlled by distinct subspaces.
- Cross-Type Transfer: A subspace trained to control one builder type (e.g., hypothetical) is applied unchanged to other builder types (e.g., belief, temporal) to test for a shared mechanism.
- Composition: The study examines nested spaces (e.g., "Alice used to believe...") to see if a new space reuses the base's value slot while minting a new router.

Key Contributions and Results

Existence of a Low-Rank Shared Subspace (The Space Index):
- The model represents coexisting values for the same entity in a later mention.
- The mechanism is a low-rank subspace (effective rank $\sim$ 4–5, causal rank $\le$ 4) that acts as a "router" or "space index."
- This index composes additively with entity identity (the representation of "Entity E in Space S" factorizes into Entity + Space).
- It is property-general (works for color, height, wealth, etc.) and entity-type-general (works for people and objects).
- The binding is retrieval-shaped: values are not stored in static global slots but are re-read from source tokens via late-layer attention heads, resembling a "lookback" mechanism.
Causal Unification Across Space Types:
- A DAS subspace trained on one builder type (e.g., counterfactuals) causally controls other types (belief, temporal, fictional) with high transfer indices (0.71–0.89 across model families).
- This unification holds even for belief, which formal semantics treats as a distinct, hyperintensional case. The model does not separate the belief tag from other space tags; it shares the same router mechanism.
- The unification is not a flat identity; a coarse semantic taxonomy (attitude, intensional, temporal, fictional) survives within the shared subspace, but belief is not specially separated.
Dissociation of Report and Reasoning:
- The value the model reports (emits as text) and the value it reasons with (applies a rule to) are controlled by distinct subspaces.
- Training a subspace on the "color report" readout flips the reported color but leaves the rule-derived consequent unchanged, and vice versa. This provides a causal mechanism for the "faithfulness gap" where a model's output may not reflect its internal computation.
Construction as a Binding Operation:
- When a sentence builds a composite space (e.g., "past belief"), the model reuses the base's value slot (high transfer of the value-binding subspace) but mints a fresh router (the query-token subspace for the composite space is distinct and orthogonal to the base router).
- This confirms that space-building is an operation on the binding machinery itself, not a separate faculty.

Significance and Claims
The paper establishes that transformer language models implement a mechanistic version of Fauconnier's unified mental-space theory. Rather than maintaining separate logical modules for modality, tense, and belief, the model uses a single, causally manipulable, low-rank "space index" to tag which discourse space a value belongs to.

Representational Format: The model carries a compact key for the discourse context, separate from the entity and the value.
Generality: This mechanism is general across the semantic inventory of non-actual contexts. The formal semantic distinction between belief (referentially opaque) and other intensional contexts is not reflected in a separate representational mechanism; the "belief tag" is not privileged.
Mechanistic Insight: The findings suggest that the "faithfulness" problem (where reasoning and reporting diverge) arises because these processes operate on distinct subspaces within the same bound token.
Scope: The paper claims this as evidence for a shared representational format in LMs, not as an adjudication of the semantics of English or a proof that LMs possess human-like theory of mind. The results are robust across three model families and scale to frontier models behaviorally, though mechanistic access is limited to smaller open models.

The authors explicitly state that the "space index" is distinct from entity-binding IDs (which track entities) and character-object-state tracking (which tracks specific triples). The contribution is the localization of the discourse space variable itself as a shared, low-rank, causally compact router.