acceptodds
Under review as a conference paper at ICLR 2027

Hidden APIs in Language Models: Discovering Reusable Causal Interfaces from Forked Futures

Abstract

Identical language-model answers can arise from hidden states that support different future computations, so present-answer decodability leaves the organization of reusable state unresolved. We introduce forked futures: future operations are sampled only after a prefix state has formed, and the resulting response distributions define a future causal signature over that state. Shared, Local, Mixture, and Distributed interfaces then compete under prequential causal description length with matched capacity and future-signature fidelity. Shared achieves the lowest held-out description length on Qwen2.5-1.5B and Llama-3-8B, with Sharedness Gains of 0.216 and 0.294 nats, while preserving comparable or lower mean, worst-family, and tail distortion; the preference remains positive across a five-backbone sweep. Role-aligned transplantation gives Shared the strongest joint target-correctness, locality, and copy-preservation profile, and API-aligned paths mediate 0.749 of the target effect versus 0.150 for matched null paths. In blind model organisms with known routing, the procedure recovers 14/16 architectures. Together, these results identify a compact reusable causal interface through convergent evidence from description length, held-out generalization, transplantation, mediation, and ground-truth architecture recovery.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, language models, causal interventions, representation learning, minimum description length, activation patching, reusable representations

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_5xgTOpenReview ↗
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 (0.7490.749 mediated fraction vs. 0.1500.150 for nulls), and blind recovery of 14/1614/16 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.

Reviewer_5KAoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework for discovering reusable causal interfaces in language models through forked futures. By defining hidden states based on their future response distributions, the authors compare four candidate architectures (Shared, Local, Mixture, and Distributed) using prequential causal MDL. The framework is evaluated through architecture comparisons, activation transplantation, causal mediation, and model-organism experiments.

Critical strengths
  1. Interesting research direction. The paper investigates reusable internal representations from the perspective of future computations, providing an interesting alternative to conventional probing methods.

  2. Comprehensive experimental evaluation. The authors conduct extensive experiments across multiple language models, including architecture comparisons, causal interventions, generalization tests, and controlled model-organism studies.

  3. Causal validation. The transplantation and mediation experiments provide additional evidence that the discovered representations are involved in downstream computations, rather than merely correlating with model outputs.

Critical weaknesses
  1. The main claim is that LLMs contain reusable causal interfaces. However, the fact that an externally trained Shared encoder achieves lower MDL does not necessarily establish that the original model internally implements a shared computational interface. Even distributed representations may admit compact shared abstractions. Although the causal interventions provide supporting evidence, the distinction between an extracted representation and an intrinsic computational interface remains unclear.

  2. The architecture selection depends on predefined candidate structures and MDL complexity penalties. Since Shared is structurally simpler than the alternatives, its advantage may partly reflect the chosen coding scheme rather than genuine architectural reuse. The paper would benefit from stronger sensitivity analyses or comparisons with more flexible hierarchical and nonlinear architectures.

  3. Limited generalization to realistic tasks. The experiments primarily focus on synthetic structured tasks involving relations, program states, and logical reasoning. These tasks have well-defined states and operations, which may naturally favor shared representations. It remains unclear whether the findings extend to more realistic settings, such as open-ended reasoning, long-context interactions, and tool use.

Questions for authors
  1. How can the proposed framework distinguish an externally extracted shared representation from a genuinely shared computational interface used by the original model?

  2. How sensitive are the architecture rankings to the MDL complexity penalties? Would the conclusions hold with more expressive hierarchical or nonlinear alternatives?

  3. Can the authors provide additional evidence of generalization to less structured, real-world language-model tasks?

  4. What causes the Local architecture misclassifications in the model-organism experiments, and how might these errors affect the conclusions on pretrained LLMs?

Reviewer_eYy8OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

NA

Critical strengths

None

Critical weaknesses

Unfortunately, I think this article should be desk rejected. I found the paper to be unreadable. In particular, the over-formalization and poor explanation seems like LLM generated text to me.

Questions for authors

NA

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.