acceptodds
Under review as a conference paper at ICLR 2027

StoryHARN: Repairing Conflicting Actions in Interactive Stories

Abstract

Language-model storytellers must reconcile open-ended player actions with persistent world rules. Yet state consistency alone cannot distinguish a useful repair from a rejection that brings play to a halt. We introduce StoryHARN, a runtime layer that checks proposed state changes before they enter the story and handles conflicting actions through three mechanisms: rule-compatible alternatives, isolated what-if branches, and deferred intentions requiring later confirmation. Its decisions are recorded for inspection and replay. We evaluate five configurations over 1,200 turns across 12 authored worlds, separating structured-state validity, action coverage, and a proxy for story continuation. In the initial comparison, StoryHARN achieves 67.5% state validity, compared with 39.2% for rule prompting and 90.8% for hard rejection. Trace analysis attributes the observed StoryHARN state failures to a monitor gap and an overly restrictive evidence rule. A separately versioned development follow-up records 1,020 committed turns out of 1,200 planned turns. StoryHARN and hard rejection exhibit no monitored state violations on committed turns, although a continuation advantage over rejection is not established. A separate human study collects 200 ratings of 100 three-turn traces to examine narrative consistency and causal responsiveness; it does not assess the repair comparison or fidelity to player intent. Together, these studies provide an inspectable framework for conflict handling and identify the evidence needed to establish whether valid repairs also preserve meaningful interaction.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
interactive stories, language model agents, runtime verification, evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_SQLaOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies how to handle player actions that violate world rules in LLM-driven interactive stories, and proposes StoryHARN, which checks each action against the rules before it takes effect and, on a violation, rewrites, defers, or branches it. On 12 Chinese story worlds built by the authors, the method violates rules significantly more often than simply rejecting violating actions (67.5% vs. 90.8% state validity), and the two perform on par after the authors fix problems in their own rules.

Critical strengths

None. The authors honestly report their negative results and the problems in their own implementation, but this does not substitute for a contribution.

Critical weaknesses
  1. The method underperforms the simplest baseline. Directly rejecting violating actions reaches 90.8% state validity, while StoryHARN reaches only 67.5%; after the rules are fixed, the two perform on par. The branching and deferral mechanisms that the paper emphasizes are almost never triggered in the experiments.
  2. The method lacks novelty. In Section 2 the authors themselves acknowledge that checking and replacing a violating action before it takes effect "is not itself a new contribution."
  3. The comparison is unfair. The methods give the model different prompts, so the differences in results cannot be attributed to the methods themselves.
Questions for authors
  1. If all methods used the same prompt, would the gap between StoryHARN and direct rejection still exist?
  2. Is there any example from natural runs where branching or deferral was triggered and led to a better outcome than direct rejection? If not, where does the value of these two mechanisms lie?
Reviewer_QGj5OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents StoryHARN, a runtime framework for handling execution conflicts in LLM-based interactive storytelling. The system validates proposed state updates against predefined rules and handles conflicts through translation, what-if branching, and deferred intentions. The authors evaluate the framework on 12 synthetic story worlds under five configurations, including rule prompting and hard rejection. The experiments report state validity and continuation metrics, along with implementation diagnostics and a revised runtime evaluation.

Critical strengths

The paper provides some implementation records and failure analyses, although I find their scientific contribution rather limited.

Critical weaknesses

I have significant concerns about this submission, which reads more like a text-game development report than a scientific research paper.

Lack of scientific novelty. The paper mainly discusses engineering implementations of state validation and conflict handling without establishing a sufficiently novel or generalizable scientific contribution. Related problems, including narrative redirection and alternatives to hard rejection, have already been studied in previous work such as SENNA (Jørgensen et al., IUI 2026). The authors do not clearly demonstrate what scientific advancement their proposed mechanisms offer beyond existing approaches.

Claim-evaluation mismatch. Although the work is motivated by improving narrative conflict handling, its evaluation primarily measures structured-state validity rather than the quality of narrative repairs. These metrics cannot establish whether the proposed interventions produce more coherent stories or better preserve player intentions. This substantially limits the conclusions that can be drawn from the experiments.

Insufficient experimental coverage. Two of the proposed mechanisms, What-if Branching and Deferral, lack convincing end-to-end evaluation. The original experiments contain no what-if branches or successfully activated deferred intentions. The subsequent diagnostics and revised experiments still do not establish their effectiveness in naturally evolving stories.

Lack of demonstrated effectiveness and experimental rigor. The original StoryHARN achieves substantially lower state validity than simple hard rejection (67.5% versus 90.8%), without demonstrating a meaningful continuation advantage. The later evaluation is conducted on the same development worlds after inspecting the original results, while simultaneously changing the implementation, rules, and prompts. This does not constitute an independent or controlled validation of the proposed method.

Poor presentation and organization. The manuscript devotes excessive space to implementation revisions, debugging records, execution logs, and engineering details that do not directly support its scientific contributions. The paper would benefit from a clearer research formulation and a more conventional scientific structure emphasizing hypotheses, controlled experiments, and evidence.

Questions for authors

What is the fundamental scientific contribution of StoryHARN beyond existing narrative redirection and rule-based intervention frameworks such as SENNA?

How do the authors justify using structured-state validity as the primary evaluation of narrative conflict handling?

Can the authors demonstrate the effectiveness of What-if Branching and Deferral through complete, naturally triggered interaction trajectories?

Can the authors provide an independent evaluation on held-out worlds with controlled experimental conditions?

Reviewer_5iXVOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes StoryHARN, a runtime framework for handling execution conflicts in LLM-based interactive storytelling. To maintain consistency with predefined world rules, StoryHARN introduces three repair strategies: translation, what-if branching, and deferral. The authors evaluate the framework across 12 synthetic worlds against four baseline configurations.

Critical strengths
  1. Flexible conflict-handling framework for interactive storytelling. The paper proposes an interesting approach to maintaining world consistency in LLM-based interactive storytelling by introducing three repair strategies: translation, what-if branching, and deferral. Rather than simply rejecting actions that violate predefined world rules, the proposed framework aims to preserve narrative consistency while allowing more flexible story progression.

  2. Transparent execution tracing and systematic failure analysis. The proposed framework provides traceable records of state transitions, rule violations, and repair decisions, enabling systematic analysis of its behavior. In particular, the authors go beyond reporting experimental results by identifying and analyzing specific causes of performance degradation, including limitations in the rule-checking mechanism and issues with rule specifications.

Critical weaknesses
  1. Lack of clarity and logical flow. The paper introduces numerous technical concepts in rapid succession without providing sufficient explanations or motivation. Moreover, the logical connections between sentences and paragraphs are often unclear, making the paper difficult to follow. A more structured presentation with clearer transitions between concepts would significantly improve the readability and overall clarity of the paper.

  2. Limited empirical advantages over the baselines. The experimental results raise concerns about the effectiveness of the proposed StoryHARN framework. In the initial experiments (v0.2), StoryHARN achieves substantially lower state validity than the hard-rejection baseline (67.5% vs. 90.8%). Although the authors analyze the underlying causes and introduce several improvements in v0.3, StoryHARN still does not demonstrate a clear advantage over hard rejection. In particular, the revised framework achieves a slightly lower safe-turn rate (80.83% vs. 81.25%), while its improvement in the continuation proxy remains statistically inconclusive. Given the additional complexity introduced by the proposed repair mechanisms, the experimental results do not sufficiently justify their benefits over a simpler rejection-based approach. Further methodological improvements and stronger empirical evidence are needed to establish the effectiveness of the proposed framework.

Questions for authors
  1. Although the authors address the identified implementation issues in v0.3, StoryHARN still does not demonstrate a clear advantage over the hard-rejection baseline. Could the authors clarify what specific benefits the proposed repair mechanisms provide over hard rejection and provide additional evidence supporting these benefits?

  2. The experiments provide limited evidence regarding the effectiveness of the three proposed repair strategies, particularly what-if branching and deferral. How can the authors demonstrate that these mechanisms contribute to meaningful story progression and preserve player intent, rather than merely maintaining valid states?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.