- The paper demonstrates that language models use a unified, low-rank router to tag mental spaces, enabling robust separation of discourse contexts.
- Controlled interventions like DAS reveal high causal control with flip rates over 0.9 and near-orthogonal additive composition across entity representations.
- The paper establishes a factorization between value slots and routers, providing insights into report/reasoning dissociation and modular semantic binding.
One Mechanism for Many Mental Spaces: A Shared Router over a Value Slot in LLMs
Introduction
This paper investigates the mechanistic basis by which Transformer-based LMs represent and separate multiple discourse contexts—so-called "mental spaces"—such as beliefs, counterfactuals, fictional depictions, temporal contexts, and more. It draws on the distinction established in formal semantics, where modal, doxastic, depictive, and temporal "spaces" have distinct logics, and contrasts this with Fauconnier’s mental spaces theory, which posits a unified cognitive operation for context differentiation. The central empirical finding is that LMs implement a mechanistically unified architecture: a shared, low-rank subspace ("space index" or "router") systematically tags which discourse space any attributed value belongs to, controlling semantic separation in a way that is both causally general and compositional across space-building operations.
Methods
Three distinct Transformer model families (Qwen2.5-3B, Pythia-2.8B, Falcon3-3B) were interrogated using controlled, structured natural language stimuli. The stimuli comprised assertions about entities and their attributes (primarily color), instantiated within "base" contexts (reality) and a range of alternative mental spaces (belief, painting, dream, movie, temporal, counterfactual, etc.). Linear probing and, critically, causally optimized interventions (Distributed Alignment Search; DAS) in residual streams were applied to locate, manipulate, and probe the functional role and causal status of putative space indices. A constellation of controls and variants systematically ruled out confounds related to surface form, entity and property identity, positional coding, and spurious correlations in color or lexical choices.
Key Findings
Spatial Separation and Representation
The core representational result is that LMs maintain robust, separable encodings for coexisting values across mental spaces. At query points, both the base value and the alternative context value are reliably decodable, with high accuracy maintained even in the face of distractors or surface transformations. The "space index" operates as a low-rank shared subspace, empirically of effective rank 4–5, and shows strong additive compositionality with entity identity (additive R2 ~0.92–0.95; principal cosines 0.29–0.34), indicating near-orthogonality and factorization.
Causal Control and Subspace Characteristics
Supervised steering along intuitive or diff-of-means directions yields no significant causal control above matched random subspace baselines, emphasizing the necessity for on-manifold, learned subspaces via DAS. Intervention in the identified low-rank subspace at late layers (empirically L34 for Qwen, L19/L13 for Pythia/Falcon3) allows for complete and selective control (flip rates > 0.9 with rank-1 or rank-2 subspaces) over which space-specific value is output, fully dissociating functionality from non-mechanistic leakage.
Factorization: Slot and Router
The representational system factorizes into a value slot (which holds the attributed content) and a router (the space index), which selects which discourse space’s value is read during inference or output. The router is localized at the entity/query token and is formally separate from the value token. Compositionally, nested or built-up spaces (e.g., "Alice used to believe...") are implemented by minting fresh routers atop the shared value slot, demonstrating that space-building is itself a binding operation in the model.
Cross-Type Generalization
A subspace (router) trained or optimized on one space type (e.g., counterfactuals) directly transfers to control over other spaces (belief, temporal, fictional, etc.), with strong normalized transfer indices (0.71–0.89) across all tested architectures. This causal-unification holds for both geometric structure (subspace recovery) and causal control, indicating that the representational protocol is not partitioned by discourse type, but shared and abstract.
Report vs. Reasoning Double Dissociation
A double dissociation is uncovered between the “reported” value (what the model emits as output) and the "reasoned about" value (what is actually used in downstream computational steps, e.g., rule application). Distinct DAS-optimized subspaces are responsible for controlling each, with negligible alignment (principal cosine near zero), substantiating robust mechanistic separation. This gives concrete, subspace-level reality to the behavioral faithfulness problem in LLMs: reported content may not correspond to the reasoning content.
Taxonomic Residual Structure
Despite the shared format, a residual coarse taxonomy of builder types survives within the shared subspace: modal and counterfactual, doxastic (belief), fictional/depictive, and temporal classes show internal clustering, while the partition between belief and counterfactual (referential opacity per formal semantics) is not more pronounced than other class distinctions. This supports a mechanistic realization of Fauconnier’s unifying mental-spaces hypothesis, with semantic distinctions (as per formal semantics) being non-privileged overlays rather than hard partitions at the representational level.
Implications
The results have immediate implications for mechanistic interpretability, cognitive modeling, and formal semantics. They demonstrate that current LMs instantiate a binding architecture that generalizes across what formal semantics treats as different logical phenomena. This impacts how representational modularity, compositional reasoning, and embedding-level binding are to be interpreted in LMs and may have direct downstream consequences for technical advances in prompt engineering, activation steering, and in-context reasoning. Of particular note is the direct instantiation of the report/inference faithfulness distinction as orthogonal subspaces, which is central for epistemic reliability, model auditing, and explainability.
On a theoretical level, the results constrain possible mapping between linguistic/cognitive theories and implemented mechanisms in neural LMs. The lack of a mechanistically privileged status for belief (and the opacity ascribed to doxastic contexts in formal semantics) challenges the naturalness of certain classic logical taxonomies in reflecting LM computation.
Future Directions
Several routes for extension and refinement are evident. These include mechanistic comparison of the space router with entity binding and belief tracking paths, further generalization to naturalistic (non-template) text and scenarios, mapping of nested/compound router compositionality, and behavioral mapping at frontier scale under more demanding multi-space and inference-capacity conditions. The connection to high-level causal models and the possible abstraction of router mechanisms in larger or more highly fine-tuned models remain open questions.
Conclusion
A single, causally manipulable, low-rank router mechanism implements the separation of multiple mental spaces in today’s LMs, with additive composition over entities and cross-space generality across modal, temporal, fictional, and doxastic contexts. The architecture cleanly separates the value assigned in a context from the representation of the context itself, and modularizes reported and reasoned values at the level of distributed subspaces. The representational format aligns with the predictions of mental spaces theory, with semantic distinctions arising as soft taxonomies within an otherwise unified protocol. These findings advance the understanding of neural LM internals and invite reevaluation of how semantic compositionality is mapped to architectures in practice.
Reference:
"One mechanism for many mental spaces: a shared router over a value slot in LLMs" (2607.10248)
