- Rating
- 3
- Confidence
- 1
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 3
Read review
Paper summary
The authors ask whether LLMs organize internal representations into reusable modular computational interfaces (hidden APIs) by moving out of immediate output decodability to define hidden-state equivalence through forked futures, where states are evaluated on the response distributions over downstream operations revealed only after state caching. Their primary contributions are (i) formulating an empirical causal quotient from these future causal signatures, (ii) establishing a prequential causal Minimum Description Length (MDL) competition among Shared, Local, Mixture, and Distributed interface topologies at matched behavioral fidelity, and (iii) introducing a mechanistic and ground-truth validation framework to verify downstream consumption. Using several models, the authors show that a Shared interface is statistically favored, outperforming alternative architectures and baselines (e.g., SAEs, PCA) across OOD shifts. Mechanically, role-aligned transplantation achieves high target correctness and locality.
Critical strengths
The paper’s foremost strength is its methodological shift from present-token decodability to counterfactual future sufficiency via forked futures, which solves a fundamental limitation in mechanistic interpretability by testing how internal representations serve multiple unrevealed downstream consumers rather than a single immediate output. Equally consequential is the convergent, multi-level causal validation, which does not rely solely on description-length compression: pairing prequential causal MDL with role-aligned transplantation, path-blocked mediation ( mediated fraction vs. for nulls), and blind recovery of ground-truth model organisms provides rigorous evidence that the recovered shared interface reflects actual downstream causal consumption inside the frozen model rather than an external fitting artifact.
Critical weaknesses
The most critical weakness is the reliance on synthetic, highly templated task families (relations, toy programs, and proof graphs) rather than naturalistic or open-domain language generation: because these synthetic domains possess explicit, hand-engineered discrete state transitions, they may naturally bias the model toward a compact "Shared" interface, leaving it unproven whether such shared causal APIs genuinely emerge in complex real-world behaviors like open-ended reasoning, long-context retrieval, or multi-turn dialogue. Furthermore, this limitation is compounded by the modest cross-family transfer degradation and limited causal capacity i.e., transplantation success noticeably drops from 0.88 within-family to 0.78 cross-family, and dual-task interference spikes as concurrent causal classes increase. This suggests the recovered interface may capture narrow, task-system-specific state representations rather than a truly universal, reusable computational interface.
Questions for authors
Please check weaknesses above.