Imagine a model trained on a large corpus of daily logs. It notices that every time an alarm clock rings, the sun rises shortly after. Because the conditional probability $P(\text{Sunrise} \mid \text{Alarm})$ is exceptionally high, the model concludes that to bring about the sunrise, it simply needs to trigger the alarm.
This is the post hoc ergo propter hoc fallacy—the assumption that because one event follows another, the first must have caused the second. In sequence modeling, this temporal precedence fallacy is a pervasive trap. Because real-world data is inherently chronological, purely statistical models frequently conflate temporal succession with causal necessity.
Why Sequence Models Become "Causal Parrots"
Standard deep learning architectures—whether transformers processing text or recurrent networks tracking event logs—are optimized to minimize prediction error over sequences. They excel at capturing temporal regularities. However, as Chi Gao et al. observe in their work on Noether Embedding: Efficient Learning of Temporal Regularities, identifying temporal regularities is fundamentally distinct from discovering genuine causal relationships. Coincidental sequences or shared underlying confounders can easily masquerade as valid predictive rules.
This vulnerability is highly visible in Large Language Models (LLMs). When trained on massive text corpora, LLMs absorb "script knowledge"—our shared understanding of prototypical event sequences (such as ordering food $\rightarrow$ eating). As Noah Weber et al. point out in Causal Inference of Script Knowledge, traditional methods for learning these sequences rely on raw co-occurrence frequencies. This approach cannot isolate the direct, functional influence of one event on another, leading models to capture noisy, coincidental associations. Consequently, systems routinely struggle to distinguish strict logical entailment from simple sequential actions, a failure mode documented by Chikara Hashimoto et al. in Large-Scale Verb Entailment Acquisition from the Web.
When evaluated on their reasoning abilities, models often act as "causal parrots." They mimic causal language based on statistical timing rather than an underlying understanding of functional mechanisms. As Nitish Joshi et al. demonstrate in LLMs Are Prone to Fallacies in Causal Inference, when LLMs are presented with counterfactual or synthetic scenarios where chronological order is decoupled from causality, they fall back on position heuristics and the post hoc fallacy, assuming the first-mentioned or chronologically prior event is always the cause.
The Illusion Across Modalities
The conflation of time and cause spans multiple domains and data structures:
- Video Reasoning: In video understanding, models frequently fall victim to "illusory causality." Tieyuan Chen et al. note in MECD+: Unlocking Event-Level Causal Graph Discovery for Video Reasoning that a model might mistake the consistent presence of a background object or a preceding action for a functional trigger, failing to verify if the first event actually necessitated the second.
- Temporal Knowledge Graphs (TKGs): TKG models predict future facts by looking at historical patterns. However, Qian Chen et al. argue in Beyond Entity Correlations: Disentangling Event Causal Puzzles in Temporal Knowledge Graphs that these models prioritize entity-level correlations over actual causal drivers. This makes them brittle when historical repetitions break down. Similarly, Zixuan Li et al. show in Search from History and Reason for Future: Two-stage Reasoning on Temporal Knowledge Graphs that models over-rely on the assumption that the future will simply repeat the past, failing to capture complex causal sequences where an event leads to a completely different, logically necessitated outcome.
- Process Mining: When analyzing event logs, traditional process discovery algorithms often rely exclusively on temporal ordering. Andrea Zecchini argues in Beyond Temporal Relationships: Causal Support in Declarative Process Modeling that assuming a relationship exists merely because one event consistently precedes another leads to models that misinterpret the underlying process logic, confusing coincidental sequences with operational dependencies.
To expose these systemic gaps, researchers have developed rigorous benchmarks. Martina Miliani et al. introduced ExpliCa: Evaluating Explicit Causal Reasoning in Large Language Models, revealing that state-of-the-art models show a heavy bias toward events presented in chronological order, treating chronological succession as causal consequence. To test everyday tasks, Abhinav Joshi et al. designed COLD: Causal reasOning in cLosed Daily activities, finding that while LLMs can fluidly describe sequential steps, they struggle significantly with the functional "why" behind them. A critical review by Linying Yang et al. in A Critical Review of Causal Reasoning Benchmarks for Large Language Models highlights that older evaluation frameworks often failed to penalize the temporal precedence fallacy, thereby inflating historical assessments of AI causal reasoning.
Mechanisms for Disentanglement
To move models past passive observation and toward true causal reasoning, researchers are developing frameworks that explicitly separate time from cause. The core intuition behind these solutions involves two strategies: counterfactual intervention and logical grounding.
Counterfactual Intervention and Masking
How do we prove that the alarm clock does not cause the sunrise? We perform an intervention: we turn off the alarm and observe if the sun still rises. If the outcome remains unchanged, the causal link is broken.
The ROCK framework, developed by Hongming Zhang et al. in ROCK: Causal Inference Principles for Reasoning about Commonsense Causality, applies this potential-outcomes framework to text. To prevent the model from relying on chronological precedence, ROCK uses "temporal propensities" to balance confounding prior events. It isolates true causal relationships by comparing the likelihood of an outcome against a counterfactual scenario where the hypothesized cause is replaced by a contextually similar alternative.
Passive Observation (Conflation Trap):
Event A (Alarm Rings) ----> Event B (Sun Rises) ==> Model infers A causes B
Counterfactual Intervention (Disentanglement):
Intervention: Mask/Remove Event A
Alternative World: [No Alarm] ----> Event B (Sun Rises)
Result: Event B still occurs; therefore, A does not cause B
The authors of MECD+: Unlocking Event-Level Causal Graph Discovery for Video Reasoning implement a similar interventionist mechanism in the visual domain using their Video Granger Causality Model (VGCM). Instead of allowing a network to guess causality based on the chronological flow of video frames, VGCM systematically masks specific events within the video. By measuring how the erasure of an early event affects the prediction of a subsequent one, the model mathematically isolates functional dependencies from mere visual sequence.
Logical and Ontological Grounding
An alternative strategy is to prevent the model from learning spurious temporal correlations by constraining its hypothesis space with hard logical rules.
In Neural Datalog Through Time: Informed Temporal Modeling via Logical Specification, Hongyuan Mei et al. introduce a neural-symbolic framework (NDTT) that uses logic programming to structure deep recurrent models. By encoding domain knowledge as Datalog rules, the system forces the neural architecture to respect hard relational constraints. The model is physically blocked from associating two events purely because of temporal proximity if a logical rule dictates they are independent.
Similarly, Cogan Shimizu et al. address this at the data-modeling layer in A Pattern for Modeling Causal Relations Between Events. They propose an Ontology Design Pattern that explicitly separates abstract event types ("the recipe") from concrete executions ("the execution"). By forcing data architectures to explicitly define provenance and the structural nature of a relationship, it prevents downstream systems from using simple temporal ordering or seasonality as a proxy for causation.
Architectural Equilibrium
Finally, networks can be designed to dynamically weigh temporal clues without treating them as definitive proof of causality. Xiaosong Yuan et al. propose this in TC-GAT: Graph Attention Network for Temporal Causality Discovery. TC-GAT combines Graph Attention Networks with a specialized "equilibrium mechanism." While the model utilizes the temporal flow of events as an important contextual clue, the equilibrium mechanism acts as a mathematical counterweight, balancing time-based signals against structural graph features to discover true directional influence without falling into the post hoc trap.
Key Foundational Work
Neural Datalog Through Time: Informed Temporal Modeling via Logical Specification (2020) — contributed 1 concept elements to this direction.
Causal Inference of Script Knowledge (2020) — contributed 1 concept elements to this direction.
Large-Scale Verb Entailment Acquisition from the Web (2009) — contributed 1 concept elements to this direction.