Human language is inherently linear—one word must follow another on a page or in a stream of speech. However, the real world is structural; events unfold across complex timelines and interconnected causal webs. When large language models (LLMs) process narratives, they frequently conflate these two dimensions. They fall back on a powerful but fragile heuristic: assuming that the order in which events are mentioned in text directly mirrors their chronological and causal reality. This phenomenon is known as the narrative order bias.
Why does this happen? In standard human communication, we naturally prefer "iconic ordering"—recounting events in the sequence they occurred. Furthermore, as noted in a foundational study on text-based causal discovery by Mateo Rojas-Carulla, the prevalence of the active voice in natural language corpora creates a strong baseline where the subject (the cause) routinely precedes the object (the effect). Because autoregressive language models are trained to optimize next-token prediction on these corpora, they internalize this statistical correlation as a shortcut for reasoning.
When Timelines Run Backward
When a narrative departs from a strict forward-chronological flow, the limits of this shortcut become glaringly obvious. If a story uses flashbacks, reverse-chronological updates, or complex discourse structures, LLMs routinely misinterpret the timeline.
For instance, Xin Wu et al. demonstrated that when models are presented with time-sensitive evidence in reverse chronological order, their performance drops sharply compared to when the same information is delivered in a standard forward sequence. Instead of building an abstract, robust timeline of events, the models rely heavily on the immediate narrative flow.
This dependency means that explicit linguistic markers—such as before, after, or previously—are often completely overridden by the physical position of the text. Work by Hongming Zhang et al. shows that autoregressive models exhibit a strong narrative bias, systematically assuming chronological order matches mention order even when explicit connectives state the exact opposite. Similarly, Chunkit Chan et al. evaluated these behaviors across multiple datasets, finding that models consistently fail temporal tasks when linguistic structures decouple the narrative order from temporal reality. The model is not reading the clock; it is reading the page layout.
Conflating Sequence with Causality
The bias extends far beyond simple timelines; it fundamentally corrupts how models infer cause and effect. In logic, assuming that "Event B happened after Event A, therefore Event A caused Event B" is known as the post hoc fallacy. LLMs are highly susceptible to this error because of their position-based heuristics.
When evaluating explicit causal reasoning, Nitish Joshi et al. found that LLMs frequently adopt a strict mention-order heuristic: they assume the first entity mentioned is the cause and the second is the effect. When the presentation order of entities is reversed or randomized, the models' causal understanding completely collapses. This pattern is reinforced by Khurram Yamin et al., who observed that this bias persists even when a text explicitly describes a reverse causal relationship where the effect is introduced before the cause.
Iconic Narrative (Easy for LLMs):
[Cause: The window broke.] -> [Effect: The room got cold.]
Result: Model correctly identifies the cause.
Non-Iconic Narrative (Hard for LLMs):
[Effect: The room got cold] because [Cause: the window broke.]
Result: Model often misidentifies the cold room as the cause.
This sequence-to-causality pipeline manifests in procedural reasoning as well. In a benchmark tracking dependencies in plans, Yash Kumar Lal et al. revealed that LLMs treat the linear sequence of instructions as a proxy for causality. If Step B follows Step A in the text, the model assumes Step B depends on Step A, failing to realize when steps are actually independent and could be executed in parallel.
As Martina Miliani et al. summarize in their evaluation of explicit causal reasoning, accuracy drops precipitously whenever the linguistic presentation deviates from an iconic, chronological structure. The model lacks an abstract framework for logical necessity, relying instead on linear temporal patterns.
The Illusion of Understanding via Linguistic Shortcuts
If LLMs are so heavily biased by text order, why do they often appear competent at temporal and causal tasks in standard benchmarks? The answer lies in superficial linguistic shortcuts.
Models are highly adept at exploiting surface-level patterns to guess the correct sequence without truly understanding the underlying events. For example, Kent Fujiwara et al. found that motion-language models leverage pronouns and definite articles to deduce event order in descriptions, masking their lack of genuine physical or temporal comprehension.
Furthermore, Yingjin Song et al. demonstrated that the mere presence of explicit temporal markers like then, after, or finally allows models to solve sequential ordering tasks through text processing alone. When evaluating multimodal models on image sequences, these textual cues act as a crutch. The model does not need to reason about the visual progression of time because it can simply follow the narrative breadcrumbs left by sequential language.
Breaking the Linear Dependency
To move past these superficial heuristics, researchers are developing benchmarks and training methodologies that force models to decouple text order from logical structure. Benchmarks like CaT-Bench and ExpliCa intentionally structure evaluations to penalize models that rely on position heuristics.
On the mitigation side, Hongming Zhang et al. proposed Counterfactual Data Augmentation (CDA), which deliberately trains models on "alternative reality" narratives where timelines are inverted. By exposing models to diverse narrative structures during training, the community can begin to break the rigid reliance on linear sequence, pushing models toward a more robust, abstract understanding of time and causality.
Key Foundational Work
Getting Sick After Seeing a Doctor? Diagnosing and Mitigating Knowledge Conflicts in Event Temporal Reasoning (2024) — contributed 3 concept elements to this direction.
LLMs Are Prone to Fallacies in Causal Inference (2024) — contributed 3 concept elements to this direction.
Exploring the Potential of ChatGPT on Sentence Level Relations: A Focus on Temporal, Causal, and Discourse Relations (2024) — contributed 2 concept elements to this direction.