TL;DR
This paper introduces a formal mathematical framework for "succession"—the process of determining if a modified artificial system (e.g., one that has been fine-tuned, distilled, or copied) is the same entity as its predecessor. By mapping AI transformation pipelines to nondeterministic automata and flow networks, the author establishes the computational limits of auditing AI lineages and provides a rigorous way to separate structural facts from governance decisions.
What Problem Does This Paper Address?
In the world of AI, models are rarely static. They undergo fine-tuning, distillation, quantization, and cloning. This raises a "Ship of Theseus" problem: if you fine-tune a model to be more honest, is the resulting model still the "same" entity that holds the original model's legal obligations or alignment guarantees?
Currently, we lack a formal language to discuss these transitions. We often conflate four distinct questions:
- Continuation: What properties can the representation track?
- Compatibility: Which descendants are consistent with a specific modification?
- Recognition: Which descendants should we treat as carrying the identity?
- Realization: Which descendants actually exist in the real world?
The paper seeks a "weak but strong" mathematical structure—one general enough to cover any AI pipeline but rigorous enough to support theorems about symmetry, complexity, and recoverability.
Methods
The author models succession using a combination of category theory, automata theory, and graph theory:
- State/Modification Graphs: The history of a system is a directed graph where vertices are states (checkpoints) and edges are modifications (training runs).
- Relational Lifts: Instead of functions, modifications are treated as relations between "continuation spaces" (sets of behaviors or circuits). This allows the framework to handle fission (one model becoming two) and loss (a property disappearing).
- Functoriality and Additivity: The framework assumes that modifications compose along paths and that they act on sets of continuations element-by-element.
- Automata Mapping: The structure is formally equivalent to a nondeterministic automaton (NFA), where system states index the fibers of the automaton's state space.
- Flow Networks: To measure robustness, the paper uses Menger’s Theorem to calculate a "cut number"—the minimum number of failures required to "kill" a specific lineage.
Datasets
Not discussed in the paper. As a theoretical framework paper, it focuses on formal definitions and complexity proofs rather than empirical training on specific datasets.
Key Insights and Ideas
- Structural Underdetermination: A major "aha moment" is the proof that if a representation contains a symmetry (e.g., two identical copies of a model), no purely structural rule can decide which one inherits the "identity." The decision must come from an external governance act (symmetry-breaking).
- Separation of Governance and Structure: The framework treats the "Recognition Criterion" (the rule for identity) as an exogenous input. This allows researchers to study the cost of auditing a lineage without needing to agree on what defines "identity" first.
- The Three Fates: Every continuation along a path has exactly one of three structural statuses: Preserved (unique descendant), Diluted (multiple descendants/branching), or Dead (no descendants).
- Correctable Exclusion: The framework distinguishes between a model being "dead" (structurally impossible to recover) and "excluded" (a governance decision to set it aside). It provides a formal way to check if a "wrongful exclusion" can be corrected later by finding a new path to a valid descendant.
Key Results
- Computational Complexity: The paper establishes a baseline for the "cost of auditing":
- Determining if a lineage can become diluted is NL-complete.
- Determining if a lineage can die or be recovered is NP-complete for acyclic pipelines and PSPACE-complete for general structures with cycles.
- The Cut Number: The author introduces $\kappa_n(\gamma_0)$, a quantitative measure of robustness. It represents the maximum number of vertex-disjoint paths to a future state, effectively telling you how many independent "realization failures" a model's lineage can survive.
- Orbit Invariance: The paper proves that any "structural" recognition criterion must be constant on the orbits of the system's symmetry group.
Limitations and Future Work
- Stipulated Structure: The framework assumes the continuation spaces and relations are given (stipulated). It does not yet provide a way to automatically derive these relations from raw model weights or tensors.
- Static Criteria: The current model uses a static recognition criterion. The author notes that a "history-indexed" criterion (one that changes over time) is a logical next step.
- Plural Succession: The current work focuses on one-to-many branching (fission). The author suggests that "multi-source joins" (merging two models into one) would require extending the framework to higher-arity relations.
- Continuous Settings: The paper uses discrete sets for continuations. Future work could involve thresholding metrics on behaviors to create these spaces in a continuous domain.
Summary updated