Graded Causal Reasoning via Normality and Strength

56 concepts32 papers1 proposal

Traditional causal frameworks rely heavily on binary logic: either an event is a cause, or it is not. Under the classic counterfactual definition ("but-for" causality), an event $A$ causes $B$ if, had $A$ not occurred, $B$ would not have occurred.

However, this binary view fails to capture how humans actually reason. It suffers from two fundamental flaws: it cannot separate routine background conditions from active causes, and it cannot measure the intensity of a causal relationship. For example, if a fire breaks out, binary logic treats both the arsonist's match and the presence of oxygen as equal causes, since the fire would fail to ignite without either. Furthermore, it treats a minor contributing factor and a massive systemic failure as structurally identical.

To bridge this gap, researchers have developed frameworks for graded causal reasoning. This research direction enhances structural causal models by introducing two core mechanisms: normality rankings to filter out implausible counterfactuals, and continuous or ordinal strength metrics to quantify causal impact.

Filtering Counterfactuals with Normality and Defaults

To resolve why human intuition blames the match-striker rather than the presence of oxygen, extended causal models incorporate theories of defaults and normality. As shown by Joseph Halpern et al. in Defaults and Normality in Causal Structures, structural equations can be augmented with a ranking function that assigns typicality or normality values to possible worlds.

When evaluating a counterfactual ("What if the match hadn't been struck?"), the model restricts its analysis to alternative worlds that are no more exceptional than the actual world. Striking a match is an abnormal intervention; not striking it is a return to the default, normal state. Conversely, imagining a world without oxygen is highly exceptional and far-fetched. By filtering out these highly abnormal counterfactual contingencies, the model stabilizes causal claims and aligns with human judgment, preventing non-causes from retroactively becoming causes when new variables are introduced (Appropriate Causal Models and Stability of Causation).

[Actual World: Match + Oxygen = Fire]
       /                        \
      / (Test Counterfactuals)   \
     v                            v
[World A: No Match + Oxygen]   [World B: Match + No Oxygen]
  * Normal default state         * Highly exceptional state
  * VALID comparison             * FILTERED OUT as far-fetched

This integration of normality allows systems to differentiate between prescriptive norms, statistical frequencies, and unexpected deviations. For instance, Sander Beckers et al. in Combining Probabilistic, Causal, and Normative Reasoning in CP-logic demonstrate that causal responsibility can be modeled by weighting a counterfactual outcome by the initial atypicality of the cause. Rare or norm-violating actions are naturally prioritized as primary causes. Without these constraints, abstract frameworks struggle to prevent irrelevant or pathological formulas from polluting causal explanations (Causality Without Causal Models).

The second major limitation of traditional models is their inability to distinguish between a minor contributing factor and a dominant root cause. In complex domains like medicine or telecommunications, relationships exist along a spectrum of intensity.

Early foundations for non-binary representation were laid by Robert L. Blum in Representation of Empirically Derived Causal Relationships, which used multivariate linear models to capture the continuous variance one biological variable accounts for in another. In modern structural frameworks, this is often handled by quantifying degrees of responsibility. Joseph Halpern et al. in Causality, Responsibility, and Blame: A Structural-Model Approach formalize this by measuring how many external factors must be modified to transform a variable into a decisive "but-for" cause. This approach successfully accounts for over-determined outcomes—such as a vote won by a landslide—where individual contributions are graded based on how close they were to being pivotal.

In natural language processing and computer vision, evaluating whether models understand this gradation has become a major focus. Traditional benchmarks often collapse causality into binary classification, but recent datasets introduce ordinal or continuous scales:

  • Narrative Causality: Margaret Meehan et al. developed an ordinal scoring system to capture the varying intensities of narrative causality in fiction, demonstrating that collapsing links into binary categories overlooks how stories are interconnected (Causality Mining in Fiction).
  • Real-World Events: The CRAB benchmark, introduced by Angelika Romanou et al., uses a continuous intensity scale to evaluate how large language models (LLMs) navigate complex news networks, revealing that models often hallucinate stronger causal links than humans perceive (CRAB: Assessing the Strength of Causal Relationships Between Real-world Events).

The Challenge of Causal Intensity in Modern AI

Why does quantifying causal strength matter for current AI systems? Empirical evidence shows a stark performance gap: modern machine learning models can easily identify the direction or polarity of a causal effect, but they fail miserably at estimating its magnitude.

In the CEBaB benchmark, Eldar David Abraham et al. discovered that while NLP explanation methods can predict whether a concept positively or negatively impacts model behavior, they cannot accurately quantify the specific magnitude of that influence (CEBaB: Estimating the Causal Effects of Real-World Concepts on NLP Model Behavior). This is echoed by Shaobo Cui 0006 et al. in Nuance Matters: Probing Epistemic Consistency in Causal Reasoning, who found that advanced LLMs exhibit high proficiency in binary polarity classification (e.g., distinguishing whether a factor supports or defeats a claim) but fail to consistently rank fine-grained degrees of causal intensity.

Direction/Polarity (What AI can do):
[Factor A] ----(+)----> [Outcome]  (Identifies that A increases Outcome)

Intensity/Magnitude (Where AI struggles):
[Factor A] ----(0.12)----> [Outcome]  (Fails to quantify the exact strength)

This limitation becomes critical when modeling defeasibility—the way new information strengthens or weakens an existing causal link. As Shaobo Cui et al. note in Exploring Defeasibility in Causal Reasoning, moving from qualitative, directional assessments to precise numerical or probabilistic shifts remains an open challenge. When models treat causality as a binary state, they fail to replicate the nuanced, graded judgments humans make naturally when observing physical, legal, or narrative events (CLEVRER-Humans: Describing Physical and Causal Events the Human Way).

Key Foundational Work

Representation of Empirically Derived Causal Relationships (1983) — contributed 1 concept elements to this direction.
An Approach to Default Reasoning Based on a First-Order Conditional Logic (1987) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_graded_causality_d3e74d46,
  title = {graded causality},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/graded-causality/rdr_d3e74d46e2054eb19abf13fc2ea26c7b},
  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_graded_causality_d3e74d46_v1,
  title = {graded causality},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/graded-causality/rdr_d3e74d46e2054eb19abf13fc2ea26c7b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake