Distinguishing Revision and Delayed Elaboration in Incremental Narrative Interpretation

Abstract

Both human and AI systems processing narrative or long-form content operate incrementally: input arrives over time, and internal representations must be updated accordingly. Incremental interpretation thus depends not only on what is represented, but also on how the representational state evolves in response to new evidence.

This work distinguishes two structurally distinct update operators in narrative interpretation: revision-driven update and delayed elaboration. Revision-driven updates retract or replace previously committed structure when contradictions arise, rendering them non-monotonic. In contrast, delayed elaboration refines initially underspecified elements by adding constraints without retracting prior commitments, yielding monotonic extensions of the interpretive state. Although both operators can alter how earlier material is understood, they impose fundamentally different structural requirements on state transitions.

Using visual narratives as a diagnostic domain, we demonstrate how a structured narrative representation can explicitly separate committed from underspecified content, supporting both update operators during incremental construction. A worked example illustrates how delayed elaboration enables monotonic refinement, while revision requires non-monotonic correction. We conclude by discussing the broader relevance of this structural distinction for incremental reasoning and hybrid symbolic-neural systems.

Introduction

Incremental processing—whether in human cognition or artificial systems—demands that incoming information continuously shapes an evolving internal model. For narrative understanding, this poses a fundamental challenge: how can a system maintain coherent interpretations when later input may confirm, refine, or contradict earlier assumptions? The answer hinges on the nature of the update operators that govern representational change.

Two such operators are central to narrative interpretation: revision and delayed elaboration. While both modify the interpretive state, they operate under different logical constraints—one non-monotonic, the other monotonic. This distinction has significant implications for system design, yet it is often conflated in computational models. Our work clarifies the structural differences and demonstrates their practical impact through a visual narrative testbed.

The Two Update Operators

Revision-Driven Update

Revision-driven updates are triggered when new information contradicts existing commitments. The system must retract or replace previously established structure to restore consistency. This process is inherently non-monotonic: conclusions drawn earlier may no longer hold, requiring a rollback of prior inferences. In narrative terms, revision occurs when a plot twist forces the reader to reinterpret earlier events.

Delayed Elaboration

Delayed elaboration, by contrast, addresses initially underspecified content. Instead of retracting commitments, it enriches the existing representation through constraint addition. Since no prior structure is discarded, the update is monotonic—interpretive state grows monotonically as new constraints are applied. This operator is ubiquitous in natural discourse, where meanings are often sketched initially and sharpened over time.

Structural Contrast

Table 1 summarizes key differences between the two operators.

| Feature | Revision | Delayed Elaboration |

|-------------------|--------------------------|--------------------------|

| Logical type | Non-monotonic | Monotonic |

| Trigger | Contradiction | Underspecification |

| Operation | Retract/replace | Add constraints |

| Effect on prior state | Destructive | Additive |

Representing Interpretive State

To support both operators, a narrative representation must clearly distinguish committed content from underspecified elements. We propose a structured model with three zones:

Committed: Elements that are fully determined and subject to revision.
Underspecified: Elements with unresolved slots, open to elaboration.
Latent: Potential connections awaiting evidence.

This tripartite structure enables a system to route incoming information to the appropriate operator—triggering revision when contradictions target committed elements, or elaboration when constraints fill in underspecified slots.

Empirical Demonstration with Visual Narratives

Visual narratives—such as comics or storyboards—serve as an ideal diagnostic domain. They are inherently sequential, rely on implicit relations, and demand continuous updating. We implemented a parser that constructs structured representations incrementally, frame by frame.

Worked Example

Consider a four-panel comic sequence:

Panel 1: A character enters a room, looking suspicious.
Panel 2: The same character appears calm, talking to a friend.
Panel 3: A cut shows a broken vase on the floor.
Panel 4: The character is seen picking up the pieces, looking guilty.

Initial processing of Panel 1 leaves the character's intent underspecified. Panel 2 provides no immediate contradiction, so delayed elaboration refines the representation (character has social relationships). Panel 3 introduces an object (vase) that was not previously part of the scene—this triggers revision if the system had assumed no breakage. Alternatively, if the vase was anticipated as a latent object, Panel 3 may simply elaborate it. Panel 4 resolves the causal chain, requiring revision of the earlier "suspicious" tag to "guilty."

This example illustrates how both operators work in concert: elaboration accumulates constraints, while revision corrects prior commitments when contradictions surface.

Implications for AI Systems

Our findings have direct implications for hybrid symbolic-neural architectures. Neural components, trained on large corpora, often predict upcoming events but lack explicit logical structure. A hybrid system can pair a neural encoder with a symbolic reasoner that manages committed and underspecified content. The distinction between revision and elaboration informs the reasoner's update policy:

When a neural prediction conflicts with committed structure, the system must decide whether to revise (non-monotonic) or treat the prediction as noise.
When information fits an underspecified slot, elaboration can proceed monotonically, avoiding costly backtracking.

This design reduces unnecessary non-monotonic operations, improving computational efficiency and interpretive coherence.

Conclusion

We have articulated a structural distinction between two update operators in incremental narrative interpretation: revision (non-monotonic) and delayed elaboration (monotonic). By separating committed from underspecified content, systems can apply the appropriate operator based on the state of the representation. Visual narratives provide a concrete testbed for validating this framework, and the principles extend to broader incremental reasoning tasks. Future work will explore how this distinction scales to longer texts and real-time dialogue processing, where the cost of non-monotonic updates is especially high.

Acknowledgments

We thank the organizers and reviewers of CMN 2026 for their constructive feedback. This work was supported in part by a grant from the Office of Naval Research (No. N00014-25-1-0033).

References

G. A. Miller, "The magic number seven, plus or minus two: Some limits on our capacity for processing information," Psychological Review, 63(2):81–97, 1956.
D. Lewis, "Scorekeeping in a language game," Journal of Philosophical Logic, 8:339–359, 1979.
P. N. Johnson-Laird, Mental Models: Towards a Cognitive Science of Language, Inference, and Consciousness. Harvard University Press, 1983.
T. A. van Dijk and W. Kintsch, Strategies of Discourse Comprehension. Academic Press, 1983.
M. C. Frank and J. B. Tenenbaum, "A Bayesian framework for cross-situational word learning," in Advances in Neural Information Processing Systems, 2009.
J. P. de Ruiter and L. Louwerse, "The role of visual context in the production and comprehension of spatial language," Cognitive Science, 35(1):1–31, 2011.
Y. Chen, "Structured representation for incremental narrative interpretation," in Proceedings of CMN 2026, 2026.

via ArXiv CL+LG