acceptodds
Under review as a conference paper at ICLR 2027

Taxonomy for Hallucination Localization and Attribution in LLMs

Abstract

While hallucination is a well-documented failure mode of LLMs, localizing and attributing its occurrence within generated text remains ambiguous. Existing span-level hallucination annotations often exhibit substantial disagreement because annotators select different textual spans for the same underlying semantic error. This ambiguity leads to noisy training labels, unreliable evaluation of hallucination detectors, and obscures where hallucinations emerge, knowledge that is essential for building effective mitigation strategies. In this work, we argue that this localization challenge originates due to the absence of a semantic model the defines the minimal unit of hallucination. To address this limitation, we introduce a diagnostic taxonomy for hallucination localization, grounded in Conceptual Dependency Theory (CDT), which categorizes hallucinations according to failures in five semantic components: , , , , and . A deterministic decision procedure guides annotators in identifying the minimal semantic component responsible for each hallucination, reducing ambiguity in span localization and semantic labelling. The taxonomy was developed through the analysis of 2,000 LLM hallucinations and validated in a held-out replication study in which five independent human annotators applied the frozen taxonomy and decision procedure to 300 previously unseen atomic claims drawn from four benchmark datasets. The taxonomy coupled with the decision tree achieved an average pairwise Cohen's kappa of 0.865, demonstrating that it reduces ambiguity in both hallucination span annotation and failure attribution.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Hallucination, Taxonomy, LLM, Fine-grained, Localization, Attribution, Inter-annotator Agreement.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_VivjOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a span-level hallucination taxonomy, categorizing hallucinations into 5 semantic components, including actor, state, temporal, object, and action. The authors collect 2000 hallucinated model generations and ask 5 annotators to code each sample and annotate hallucinations at the claim level. Through an iterative process, the authors form a hallucination taxonomy and decision procedure to annotate hallucinations with high inter-annotator agreement. The experimental results suggest that the proposed taxonomy achieves higher agreement than taxonomies used in prior work.

Critical strengths
  1. Annotating hallucination at the span level is usually costly and noisy. This paper proposes a new taxonomy, grounded in Conceptual Dependency Theory, making the annotation more reliable.
  2. The paper provides detailed definitions for each hallucination type and a clear taxonomy decision procedure, making it easy for the community to follow and adopt the pipeline for other hallucination datasets.
  3. The paper is generally well written and easy to read.
Critical weaknesses
  1. Lack of details in the methodology section: Although the paper provides details of the hallucination taxonomy and decision procedure, the discussion of open coding and quality assessment is insufficient, making it hard to assess the soundness of the annotation results. Specifically,
    • How many iterations of open coding are performed? The paper mentions only that the open coding process was conducted iteratively (Line 158), without specifying the number of iterations or the termination criteria.
    • How are the identified themes connected to CDT and hallucination taxonomy? The 5 themes identified in Appendix A.2 do not directly map to the hallucination taxonomy. The paper should provide more details on their connection and how annotators are instructed to annotate the taxonomy after open coding.
    • How is Cohen's kappa computed on the development set? The paper states that the development set was annotated by one human annotator and two LLMs (Line 149). Is the agreement measured between the human and LLM? How do the authors ensure that the LLM's annotation is reliable?
    • How is pairwise Cohen’s kappa performed on the testing set? The annotation is done by 5 annotators, while pairwise Cohen’s kappa is designed for 2 annotators. It is unclear how the authors compute inter-annotator agreement.
    • How is the quality of localization measured? The paper measures the quality of hallucination classification at the claim level, whereas hallucination localization is done at the span level. The paper should also discuss how the authors measure the quality of hallucination localization.
  2. The experiment is biased toward the taxonomy proposed in this paper: The paper compares its own taxonomy against taxonomies derived from seven hallucination categorization schemes used in prior work. However, apart from taxonomy design, there are many other reasons to make annotators have higher agreement on its own taxonomy. Specifically,
    • Unlike its own taxonomy, where the annotators spend many iterations on forming and refining the themes, it is natural to expect that annotators may be less familiar with taxonomies used in prior work.
    • The agreement of its own taxonomy is obtained after multiple iterations on the same development set. Thus, the agreement is natural to be higher than annotating taxonomies without iterative refinement.
  3. The evaluation and analysis are too thin:
    • The refinement analysis should naturally be put in Section 3, as it is part of the rationale for the design of the decision tree. Other than that, there is no analysis in Section 5.
    • The only evaluations are Cohen's kappa on the testing set and comparison against prior work on the development set. The limitation of the comparison is discussed in the previous point.
    • I expect more analyses and experiments based on the annotation results of the proposed taxonomy. For example, how do LLMs hallucinate each type of error; which type is the most hallucinated; and what is the inter-annotator agreement across error types? Some analyses discussed in the appendix are worth putting in the main paper with more details (e.g., there is no table or figure for confusion analysis in the current appendix).
    • The paper lists automatic hallucination detection as future work, which is reasonable. But I think it is fair to run some simple hallucination detection algorithms (e.g., LLM-as-a-Judge and uncertainty-based methods [1, 2, 3]) on the testing set to see how existing algorithms perform on detecting span-level hallucinations.
  4. While the paper proposes a hallucination annotation pipeline to obtain high agreement across annotators, the process seems expensive in terms of time and cost. It is unclear how easy others can adopt the same pipeline to a larger hallucination dataset. It would be more informative if the authors could report the time and money spent on annotating these data.

[1]: Zhang et al. 2023. Enhancing uncertainty-based hallucination detection with stronger focus. [2]: Guerreiro et al. 2023. Looking for a needle in a haystack: A comprehensive study of hallucinations in neural machine translation. [3]: Malinin and Gales. 2021. Uncertainty estimation in autoregressive structured prediction.

Questions for authors
  1. In Section 3.1, the author mentions that they collect 2000 hallucinated model outputs. However, in Section 3.2, the 2000 samples are then divided into 1550 model outputs for training, and 150 and 300 hallucinated atomic claims for development and testing, respectively. Does each model output contain only one atomic claim? Or what is the unit of the collected samples?
  2. Do the authors know why Cohen's kappa on the testing set is higher than on the development set? Since the themes are refined based on the evaluation results on the development set, it is expected to see higher agreement on the same set.
2 comments

1. Open Coding and Termination

Open coding involved four training-set rounds until no new categories emerged and boundaries stabilized. We will clarify this in Sec. 3.1.2. This interpretive process depends on annotators, example order and judgments. Table 3 shows development-set taxonomy variants; Sec. 5.1 reports disagreements prompting refinement. Reproducibility concerns final definitions, guidelines and procedure, not identical exploratory iterations.

2. CDT and the Final Taxonomy

CDT does not map one-to-one to the taxonomy. Its inventory (PP, PA subtypes, ACT, STATE and actor/object/recipient roles) was explored as candidate nodes. Open coding identified output themes; CDT grounded them theoretically, and agreement guided selection and consolidation. Table 3, Sec. 5 and Appendix A.4 document this process. Primitives without empirical separability were not retained.

Five categories map as follows:

  • ACTOR_ERROR: ACTOR role, supported by PP.
  • OBJECT_ERROR: OBJECT role, supported by PP.
  • ACTION_ERROR: ACTION.
  • STATE_ERROR: STATE, including identity and classification errors.
  • TEMPORAL_ERROR: TIME, a PA subtype.

PP represents ACTOR and OBJECT entities. LOCATION and PROPERTY were subsumed under ACTION_ERROR and OBJECT_ERROR. Finer-grained hierarchy has more mappings but put as future work.

3. Development-Set Kappa

Agreement averages kappa across human–LLM 1, human–LLM 2 and LLM 1–LLM 2 pairs. LLMs are not ground truth. Final reliability uses independent humans on the held-out test set.

4. Test-Set Pairwise Kappa

For five annotators, average pairwise kappa is the arithmetic mean of ten unique coefficients ((\binom{5}{2}=10)), not a single multi-annotator calculation. We will clarify this in the methodology.

5. Localization Quality

The study operates on atomic claims. The 450 development/test claims were decomposed using GPT-4o with human validation (Sec. 3.1.2), then fixed before annotation. Each claim is segmented into ACTOR, ACTION and OBJECT using the deterministic left-to-right rule (Sec. 3.6.2); the ordered procedure identifies the first failing component, label and location.

Five annotators labelled the same 300 test claims; κ=0.865 measures agreement on the failing component and span within each claim. We did not measure agreement on decomposition or mapping spans to original outputs (Sec. 6). Decomposition is treated as an established factuality-evaluation step; FActScore and SAFE offer automated approaches. We focus on attribution within claims, where disagreement was observed (Sec. 5.1). This is an assumption, not an empirical finding: results establish consistency within atomic claims, not end-to-end localization in unrestricted text.

6. Prior-Work Taxonomies

Baseline taxonomies used published definitions and trained annotators on the same set under standardized guidelines. Our advantage is cognitive grounding and deterministic resolution of category overlap.

7. Independent Validation

Five final-validation annotators, independent of taxonomy development, evaluated the held-out test set.

8. Refinement Analysis

The appendix refinement analysis, referenced in the main text, documents stages supporting the final procedure.

9. Scope of Evaluation

The manuscript reports overall agreement and confusion patterns. Category-level distributions across LLMs and agreement by error category remain useful extensions.

10. Additional Analyses

Further analyses were omitted due to ICLR page limits.

11. Automatic Detection

We develop a taxonomy for localization; automatic detection is distinct. The taxonomy provides span-level annotations for training and evaluating detection systems. LLM-as-a-Judge and uncertainty-based approaches require alignment with span-level localization for meaningful comparison. These are future directions, not validation prerequisites.

12. Annotation Cost and Scalability

Manual annotation requires identifying, localizing and categorizing claims. Our procedure improves consistency and reproducibility, not effort. Time was not systematically recorded because annotators worked independently with different tools and environments. We will clarify this limitation; standardized measurement of time, throughput and costs remains future work.

13. Samples and Atomic Claims

The 2,000 samples are model outputs, each potentially containing multiple claims; 1,550 outputs were allocated to training. Development and test comprise 150 and 300 atomic claims. These 450 claims were decomposed using GPT-4o with human validation (Sec. 3.1.2) and fixed before annotation.

14. Development- and Test-Set Kappa

Development agreement involved one human and two LLMs; testing involved five independent humans. Values are not directly comparable given differences in annotators, samples and taxonomy maturity. Test agreement was higher on unseen claims using the frozen taxonomy and procedure; the test set was not used for refinement.

I thank the authors for their response. However, I don't think my concerns are properly addressed and I will keep my score

Reviewer_2Pm3OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a hallucination taxonomy based on Conceptual Dependency Theory, categorizing errors into five semantic types. A deterministic decision procedure improves annotation consistency, achieving a Cohen’s kappa of 0.865 across five annotators. However, the study lacks methodological novelty and fails to demonstrate whether improved annotation agreement translates into meaningful downstream benefits.

Critical strengths

The paper addresses a practical challenge in hallucination annotation by introducing a structured decision procedure that reduces ambiguity and improves consistency among human annotators.

Critical weaknesses
  1. Limited Novelty and Scientific Contribution. The proposed taxonomy largely reorganizes existing semantic error categories using manually designed rules, without introducing new algorithms, theoretical insights, or previously unknown hallucination phenomena.
  2. Insufficient Validation of Taxonomy Quality. The evaluation focuses primarily on inter-annotator agreement, which does not establish whether the proposed categories accurately capture meaningful hallucination characteristics. High agreement alone does not guarantee scientific validity.
  3. Unfair Comparison with Existing Taxonomies. The baselines were designed for different purposes and use different category structures. Comparing them solely through annotation agreement does not demonstrate the superiority of the proposed taxonomy.
  4. Lack of Downstream Utility. The paper does not demonstrate whether the taxonomy improves hallucination detection, localization, mitigation, or model debugging, leaving its practical value to the ML community unclear.
  5. Questionable Error Attribution. The deterministic decision procedure forces potentially overlapping errors into mutually exclusive categories. This may artificially improve agreement while discarding meaningful information about complex hallucinations.
Questions for authors

NA

2 comments

Thank you for taking the time to review.

1. Scientific Contribution and Novelty

The contribution is not the discovery of new hallucination phenomena or a novel detection algorithm. It is the development of a theoretically grounded semantic taxonomy and an operational decision procedure for hallucination localization.

Prior work has proposed various hallucination categorization schemes. However, these are largely descriptive (e.g., intrinsic vs. extrinsic, or broad token/sentence categories) rather than systematic frameworks for consistently identifying and classifying hallucinated spans. Such schemes also suffer from low inter-annotator agreement (often below κ=0.40).

We address this gap through semantic grounding, explicit category definitions, a deterministic decision procedure and independent empirical validation. Specifically, we operationalize Conceptual Dependency Theory (CDT) into a reproducible procedure for span-level localization. Achieving Cohen's κ=0.865 across independent human annotators addresses a longstanding methodological bottleneck, providing a foundation for fine-grained dataset creation and detector evaluation.

The novelty lies in making hallucination localization a reproducible annotation task, rather than proposing another list of hallucination types.

2. Unfair Comparison with Existing Taxonomies

We acknowledge that existing schemes were developed for different purposes and have different category structures. Our comparison examines their applicability to hallucination annotation and the consistency with which annotators can apply them, without claiming universal superiority.

Rather, our results demonstrate the proposed framework's suitability for reducing ambiguity and improving annotation consistency.

3. Downstream Utility

The immediate utility is more consistent and structured hallucination annotation. Ambiguity introduces noise into datasets, affecting the development and evaluation of detection and mitigation methods. By reducing ambiguity and achieving high inter-annotator agreement, our taxonomy addresses this foundational problem in span-level hallucination datasets such as RAGTruth and PsiloQA.

It provides a structured representation of hallucination types and locations, supporting subsequent work on detection, localization, mitigation and model debugging. While downstream performance improvements would be valuable to demonstrate, they are not the primary objective of this paper. We will distinguish demonstrated annotation benefits from potential downstream applications.

Additionally, our error distribution analysis (Section 5.1 and Appendix A.3) shows that over 70% of deceptive hallucinations involve subtle Object and State errors rather than total incoherence. This diagnostic granularity can inform targeted mitigation strategies, including RAG and knowledge-graph grounding.

4. Insufficient Validation of Taxonomy Quality

We agree that high inter-annotator agreement alone does not establish every dimension of scientific validity. However, our principal research question is whether independent annotators can consistently apply the taxonomy and its decision procedure to hallucinated claims. Inter-annotator agreement is therefore a central and directly relevant measure of the framework's effectiveness.

Our validation does not establish that the taxonomy captures every conceivable hallucination or is universally optimal across domains. We will clarify its scope and avoid treating agreement as a complete measure of construct validity.

5. Questionable Error Attribution

We acknowledge that complex hallucinations may involve overlapping errors and that assigning a single category can lose information. However, our procedure does not arbitrarily force overlapping errors into mutually exclusive categories. Its hierarchy follows semantic dependencies between claim components, grounded in CDT. It first establishes whether the subject or actor is valid before evaluating the associated state, object and action. Foundational errors take precedence because they affect the interpretation of dependent components. For example, an action attributed to an incorrectly identified actor cannot be meaningfully evaluated independently of that actor error.

This hierarchy establishes consistent primary attribution based on semantic dependency, supporting reproducible annotation. We acknowledge that this prioritizes consistent primary attribution over capturing every co-occurring error.

Thank you for acknowledging my concerns and providing the clarifications. I appreciate your detailed response. However, my overall assessment remains unchanged, and I will maintain my current rating.

Reviewer_4PTWOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper develops a taxonomy for hallucination localization, with the aim of reducing ambiguity in how one might annotate them. The paper then performs validation with human annotators and finds significantly higher agreement between annotators following the new taxonomy compared to previous versions.

The taxonomy itself was created using iterative coding with human annotators on a separate dataset (not the same as the validation above), based on the ideas of conceptual dependency theory (CDT) to keep the taxonomy language-independent, although this is only a higher-level inspiration and the taxonomy itself is not derived in any way from CDT.

Critical strengths
  1. Hallucination localization can indeed be a very ambiguous process, and having clearer taxonomies that can distinguish between different kinds of hallucinations and their span could indeed be useful. The results also support their claims. The contribution of the work is clear and impactful.
  2. The development of the taxonomy and its validation is well documented, including the open coding instructions, the intermediate taxonomies that were created leading up to the final taxonomy, as well as proper separation between the dataset used for open coding and the final data used for validation.
Critical weaknesses

My biggest concern with the taxonomy and its impact is how it treats ambiguity as something that needs to be necessarily reduced. As the paper itself establishes, a claim can sometimes have multiple points of hallucination, which can cause ambiguity. The taxonomy resolves this by establishing a clear hierarchy of questions (e.g., the first question asks whether the actor is a real, relevant entity; if not, it is classified as actor error, and no other piece is checked). I'm less convinced that this hierarchy means something beyond an arbitrary choice of ordering, and I wonder if we're losing information in an attempt to forcefully remove any ambiguity whatsoever.

Another smaller concern I have is the heavy dependence of the taxonomy on the dataset used for open coding, which is, of course, an inherent bias in the open coding technique itself. For instance, one of the five categories for the hallucination span taxonomy is temporal error, specifically referring to hallucinations of time/year/dates/etc., which is a big part of current hallucination datasets. This is because a lot of these datasets need to be created in an automated manner, and getting structured facts about dates (and also locations) is easier than most other kinds of facts. However, and this is a personal anecdote and might not be widely true, I don't believe most of my interactions with an LLM are ever at risk of containing a temporal error, since those are rarely the types of questions I ask an LLM. Because of this, I wonder how applicable such a taxonomy would remain in a different domain/distribution. I think it's still a good contribution, which is why this is a smaller concern for me, and not a critical weakness.

Questions for authors
  1. Why is ambiguity necessarily always bad? If not, how do you think it would work within your taxonomy, or what might you want to change in the future?
  2. How much do you think the taxonomy you have created would work in, to take an example, let's say the medical domain? Would it not end up possibly conflating distinct error types into one big category, while potentially other categories might remain empty?

Suggestion ->

  1. Given that the same test set was annotated by the five human annotators for various taxonomies (roughly 15, combining both the other baseline taxonomies and all intermediate taxonomies), clearly, they had to go through the same examples a significant number of times (please correct me if my understanding of this is wrong). It would be good to document more explicitly in the paper what the logistics of the process were. Were they annotating each example for all taxonomies at the same time, or one taxonomy after another? Was it all done in the same order? Could it have created saturation for the annotators or possibly older taxonomies influencing their choices in the new taxonomies? And so on. Not a question that needs to be answered/solved by the authors, but just something that should be properly documented for any future work.
2 comments

Thank you for the supportive review.

1. On Ambiguity and Hierarchical Ordering

We agree that ambiguity can be a natural feature of human language, but in evaluation benchmarks and training labels, ambiguity creates severe downstream noise. We would like to distinguish between ambiguity inherent in natural language (Phenomenon ambiguity) and ambiguity introduced during annotation (Protocol ambiguity).

The paper does not argue that ambiguity in natural language is inherently undesirable or that every complex claim can be reduced to a single underlying semantic phenomenon.

The concern is specifically about annotation ambiguity: when different annotators interpret the same hallucinated claim differently, the resulting annotations become inconsistent and introduce noise into the dataset.

As established in Section 1, when multiple annotators attribute the exact same underlying semantic error to different parts of a sentence (e.g., one labeling a relation error and another an object error), it results in inconsistent training labels and unreliable evaluation of hallucination detectors.

Phenomenon ambiguity is limited by design. Outputs are decomposed into atomic claims first, so multi-component failure within one claim is an edge case. For those cases the label means the first failing component under the tree. The ordering is not arbitrary. OBJECT presupposes a valid ACTOR, and ACTION presupposes a grounded ACTOR–OBJECT link. Where the order is a convention (e.g., TEMPORAL before OBJECT/ACTION), each precedence was adopted to resolve a specific confusion pair we observed (Sec 5.1).

The hierarchical decision procedure is not an arbitrary ordering; it follows a strict cognitive dependency structure. An error cannot propagate to an action or object relation if the primary entity (Actor) is fundamentally invalid or non-existent.

By prioritizing foundational semantic components first, the hierarchy ensures systematic, reproducible error attribution rather than arbitrary labeling. The objective is therefore not to eliminate the complexity of hallucinations but to provide a reproducible primary attribution procedure.

We agree this trades expressiveness for reliability. We will state it in Limitations.

2. Domain Adaptability (e.g., Medical Domain)

Because our taxonomy is grounded in Conceptual Dependency Theory (CDT), which models universal predicate-argument structures (actors, actions, states, and objects) present across all natural language utterances, it is inherently domain-agnostic.

In specialized domains like medicine, hallucinations still manifest as incorrect clinical actors (e.g., wrong drug names), incorrect states (e.g., incorrect dosages or disease classifications), wrong temporal timelines (e.g., treatment schedules), or incorrect relations (e.g., attributing a side effect to the wrong medication). We will clarify this domain-independence explicitly in the discussion.

However, domain independence does not imply that the frequency or distribution of error categories will be identical across domains. For example, medical claims may exhibit different distributions of actor, state, temporal, object and action errors compared with general-purpose conversational data. In medical text we would expect dosage and quantity errors to concentrate in STATE (quantity subtype, App. A.4.2).

The purpose of the categories is to provide a consistent semantic classification, not to enforce equal representation across categories. We agree that empirical validation on domain-specific datasets would provide further evidence of its generalizability. A finer-grained study there is future work.

3. Annotation Logistics and Potential Order Effects

We agree that annotator familiarity and exposure to different taxonomies are important considerations in comparative studies, but in our case, we explicitly took care of it. The five annotators involved in the final validation were independent of the taxonomy development process. They did not participate in developing or refining the proposed taxonomy.Furthermore, the final evaluation was conducted on a held-out test set, separate from the data used for taxonomy development and refinement.

We would also like to clarify that the comparison across taxonomies, was conducted during the development stage using the development set, which involved one human annotator and two LLM annotators. The five independent human annotators did not annotate the test examples under multiple taxonomies. They evaluated the held-out test set using only the final taxonomy and its decision procedure.

Consequently, the potential familiarity and order effects associated with applying multiple taxonomies to the same examples relate to the development-stage comparison, not the final test-set validation. We will clarify these annotation logistics in the manuscript to distinguish the two stages and make the evaluation procedure more transparent.

Thank you for the response. My concerns still remain.

  1. I'm confused by the statement that 'multi-component failure within one claim is an edge case'. If it is an edge case, what exactly do the authors believe is the cause of failure for other annotation schemes? Why do annotators disagree in other schemes? Can you give some examples of this 'protocol ambiguity'?

I think the phrase 'primary attribution procedure' is key here. It's not something mentioned in the paper, but it is mentioned in the response, and I agree that's what this attribution provides.

  1. A taxonomy is not really useful if it categorizes most things into a single category, even if it is 'correct'. So I do still think the dependence of the work here on the data used for open coding is quite strong.

  2. I misunderstood this. I thought the five annotators also annotated some of the previous taxonomies. The separation does make things clearer now. Although, I do feel now that the work is a bit weaker than my initial assessment.

Overall, I will stick with my initial scores. I think this is useful work, but I don't feel confident about its impact and value enough to suggest clear acceptance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.