Causal Abstraction and Interventional Consistency

This direction is split into other directions: interventional consistency. Its last version stays readable and citable here.

Version 1 of 172 papers145 conceptsfrozen 2026-10-04

In the natural and social sciences, we rarely model the world at its most fundamental level. A biologist might study the causal relationship between a drug and a tumor's growth without modeling every individual molecular collision. A psychologist might study how "intentions" cause "actions" without tracking every neuron. These high-level models are abstractions—simplified representations of a more complex, underlying reality.

The central challenge in this research direction is ensuring that these simplifications remain "causally consistent." If a high-level model predicts that an intervention will have a certain effect, that prediction must align with what would happen if we performed the corresponding intervention on the detailed, low-level system.

The Commutativity of Interventions

The core intuition behind a valid causal abstraction is commutativity. Imagine two paths to predict the state of a system after an intervention:

  1. The Micro Path: You start with a detailed micro-level state, apply a low-level intervention (e.g., deleting a specific gene), and then "coarsen" or abstract the resulting state into a high-level description.
  2. The Macro Path: You abstract the initial micro-state into a high-level macro-state first, and then apply a high-level intervention (e.g., "disabling the pathway").

As formalized by Paul K. Rubenstein and Bernhard Schölkopf in Causal Consistency of Structural Equation Models, a macro-level model is consistent only if these two paths yield the same result. This ensures that the high-level model isn't just a statistical summary, but a faithful causal proxy that "commutes" with the underlying physics.

Defining the Mapping

To make this work, we need more than just a way to group variables; we need a dual mapping. Sander Beckers and Joseph Halpern argue in Abstracting Causal Models that a valid abstraction requires both a variable transformation (how micro-states map to macro-states) and an intervention transformation (how high-level actions map to low-level implementations).

Crucially, they show that high-level interventions cannot be arbitrary. To avoid "abstraction gaps" where a macro-model makes impossible predictions, the mapping of high-level interventions must be naturally induced by the mapping of low-level states. This prevents a researcher from "cherry-picking" a macro-model that only looks correct under specific, limited conditions.

From Exact to Approximate Abstractions

In practice, perfect consistency is almost impossible to achieve. Most scientific models are "lossy." Sander Beckers and Joseph Halpern addressed this in Approximate Causal Abstraction by introducing distance metrics. These metrics quantify the "error" of an abstraction, allowing researchers to navigate the trade-off between a model's simplicity and its causal accuracy.

This is particularly relevant in fields like Reinforcement Learning (RL). In Learning Nonlinear Causal Reductions to Explain Reinforcement Learning Policies, Armin Kekić and colleagues use these principles to distill high-dimensional agent behaviors into simple causal "explanations." By "poking" a policy with perturbations and measuring the reward impact, they identify the macroscopic factors that actually drive success, ensuring the explanation is interventionally consistent with the agent's actual code.

Learning Abstractions from Data

While early work focused on proving consistency for known models, recent research focuses on discovering these mappings from raw data.

  • COTA (Causal Optimal Transport of Abstractions): Fabio Massimo Zennaro and Theodoros Damoulas developed COTA, which uses Optimal Transport to learn the bridge between micro and macro models. By embedding do-calculus constraints into the learning process, they ensure the learned map respects causal logic.
  • Interventionally Consistent Surrogates: In the realm of complex simulations (like pandemic modeling), Joel Dyer has proposed frameworks to build fast "surrogate" models. As detailed in Interventionally Consistent Surrogates for Agent-based Simulators, these surrogates don't just mimic the output of a slow simulator; they are trained to respect the same causal "rules," ensuring they remain accurate even when a policymaker changes the system's parameters.

Handling Determinism and Cycles

Traditional causal graphs (DAGs) often struggle with two things: feedback loops and variables that are exact functions of each other (e.g., a macro-variable that is the sum of micro-variables).

To solve the first, Joris M. Mooij and Jonas Peters explored Causal Marginalization, showing that complex cyclic systems can be simplified into functional units that preserve interventional distributions.

To solve the second, Scott Garrabrant and colleagues introduced Factored Space Models. This framework moves beyond simple graphs to "structural independence," providing a way to reason about causality even when variables at different levels of abstraction are deterministically linked.

Why It Works

Causal abstraction works because it treats "level of detail" as a formal mathematical transformation rather than a subjective choice. By enforcing that high-level models must commute with low-level interventions, researchers can build simplified models that are not just easier to understand, but are guaranteed to provide reliable answers to "what-if" questions across different scales of reality.

Papers in this version

Showing 60 of 72 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_causal_abstraction_and_interventional_co_34b27843,
  title = {Causal Abstraction and Interventional Consistency},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-abstraction-and-interventional-consistency/rdr_34b278434c564dd78ab11693128078f6},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_causal_abstraction_and_interventional_co_34b27843_v1,
  title = {Causal Abstraction and Interventional Consistency},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-abstraction-and-interventional-consistency/rdr_34b278434c564dd78ab11693128078f6/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}