Hidden APIs in Language Models: Discovering Reusable Causal Interfaces from Forked Futures
This work addresses the problem of identifying reusable internal interfaces in language models for researchers and developers, providing an empirical method to characterize causal relationships without requiring predefined latent labels. It represents an incremental step in understanding LLM internals.
The paper introduces "forked futures" to discover reusable causal interfaces within language models by comparing hidden states based on their induced response distributions to future operations. They found that a "Shared" interface consistently had the lowest held-out description length, achieving gains of 0.216 nats on Qwen2.5-1.5B and 0.294 nats on Llama-3-8B, while also demonstrating strong transplantation performance and recovering 14/16 architectures in a blind test.
Identical language-model answers can arise from hidden states that support different future computations, so current-answer probes do not establish a reusable internal interface. We introduce forked futures: future operations are sampled only after a prefix state has formed, and states are compared through the response distributions induced by those operations. This yields an empirical causal quotient over hidden states without requiring researcher-specified latent labels. Shared, Local, Mixture, and Distributed interfaces then compete under prequential causal description length subject to future-signature fidelity and matched capacity constraints. In the two detailed model evaluations, Shared has the lowest held-out description length, with gains of 0.216 nats on Qwen2.5-1.5B and 0.294 nats on Llama-3-8B, while maintaining tightly clustered mean future-signature distortion; a five-backbone sweep preserves the positive direction of Sharedness Gain. The figure-aligned transplantation analysis gives Shared the strongest joint target-correctness, locality, copy-preservation, and composite profile, and API-aligned paths mediate 0.749 of the target effect versus 0.150 for matched null paths. In the blind four-class model-organism test, 14/16 architectures are recovered, with one observed non-Shared to Shared error among 12 non-Shared organisms. These results support an economical reusable causal interface within the tested operation banks, while keeping the claim explicitly conditional on the candidate architectures, interventions, and held-out futures.