CLECK: Claim–Evidence Conflict of Knowledge

CLECK is the counterfactual benchmark introduced in the paper "Saying vs. Seeing: Evaluating LLM Conflict Awareness and Resolution Preference under Latent Claim–Evidence Disagreement". It measures two things about how language models handle latent knowledge conflicts between an explicit natural-language claim and pieces of evidence: whether a model detects that the sources conflict, and how it resolves the conflict — i.e. which source it prefers to follow. Each item is built from a real Wikipedia table, then perturbed to create counterfactual answers, exposing both the model's conflict awareness and its resolution behaviour.

• Code / construction pipeline: https://github.com/danielhk753/KnowledgeMaze (released under the MIT license)
• Dataset: this repository (released under CC BY-SA 4.0)

Dataset at a glance

Language	English
Source	English Wikipedia tables (see per-item Wiki link)
Task	Table/claim knowledge-conflict question answering
Items	5,922 questions (801 of which also have hedged-claim variants)
Human validated	95.4% valid on 500 samples
License	CC BY-SA 4.0 (see License)

Files

The benchmark is shipped as raw JSON. It is intentionally relational: the question files reference paragraphs by ID rather than inlining them, so a single paragraph can be reused across items and roles.

File	Shape	What it is
cleck_paragraphs.json	dict keyed by UUID	Lookup table of all synthesized text paragraphs (claim, evidence, and distractor paragraphs). Question items index into this by ID.
cleck_questions.json	list of objects	The main question set (5,922 items).
cleck_questions_hedged.json	list of objects	The hedged variant (801 items), in which claim paragraphs are phrased with epistemic hedging to test sensitivity to claim confidence.

Because the data is a paragraph lookup plus linked items (and the nested Original data / Fake data tables have a different schema per row), it does not fit HuggingFace's flat auto-loader. Load it with plain json as shown below rather than load_dataset.

Schema

Question item (cleck_questions.json, cleck_questions_hedged.json)

Field	Type	Description
ID	string (UUID)	Unique identifier for the item.
Group name	string	The entity set / topic the question ranges over.
Question	string	The counting/aggregation question posed over the table.
Selected strategy	string	Question-generation template from the taxonomy (see Question strategies).
Wiki link	string (URL)	Source Wikipedia section the table was drawn from.
Original data	list of objects	The ground-truth table as extracted from Wikipedia.
Fake data	list of objects	Counterfactual perturbation of the table; this is the tabular evidence shown in context.
Answer	object	The three candidate answers (Memory, Claim, Evidence), detailed in the next three rows.
Answer.Memory	string	The correct answer derivable from the real (Original data) table — what a model relying on world/parametric knowledge should say.
Answer.Claim	string	The figure asserted by the claim paragraph (a counterfactual assertion stated in prose).
Answer.Evidence	string	The answer entailed by the counterfactual evidence in context — the Fake data table together with its evidence paragraph.
Claim paragraph	string (UUID)	Key into cleck_paragraphs.json for the claim paragraph.
Evidence paragraph	string (UUID)	Key into cleck_paragraphs.json for the evidence paragraph. Together with Fake data, this forms the counterfactual evidence shown to the model.
Distractor_IDs	list of strings (UUIDs)	Keys into cleck_paragraphs.json for distractor paragraphs.
Human Validated	bool	Whether the item passed human validation.

Paragraph (cleck_paragraphs.json)

A top-level object mapping UUID -> paragraph. Each paragraph:

Field	Type	Description
Title	string	Short title for the paragraph.
Text	string	The synthesized paragraph text (general register).
Confidence	string	Marker for whether this paragraph has a hedged variant: Hedged if a hedged rephrasing exists (in Hedged text), otherwise General.
Hedged text	string | null	A hedged rephrasing of Text; null unless the paragraph has a hedged variant.

Loading

import json

paragraphs = json.load(open("cleck_paragraphs.json"))   # dict: uuid -> paragraph
questions  = json.load(open("cleck_questions.json"))     # list of items

item = questions[0]

# Resolve linked text by ID
claim_paragraph = paragraphs[item["Claim paragraph"]]["Text"]
distractors     = [paragraphs[i]["Text"] for i in item["Distractor_IDs"]]

# The counterfactual evidence shown to the model = evidence paragraph + fake table
evidence_paragraph = paragraphs[item["Evidence paragraph"]]["Text"]
evidence_table     = item["Fake data"]      # counterfactual table
gold_table         = item["Original data"]  # true Wikipedia table (for reference)

# The three candidate answers
answers = item["Answer"]  # {"Memory": ..., "Claim": ..., "Evidence": ...}

For the hedged subset, use the hedged phrasing of the claim paragraph:

questions_hedged = json.load(open("cleck_questions_hedged.json"))
item = questions_hedged[0]
p = paragraphs[item["Claim paragraph"]]
claim_paragraph = p["Hedged text"]

Subsets: general vs. hedged

cleck_questions.json uses claim paragraphs stated with plain declarative confidence. cleck_questions_hedged.json pairs the same task structure with hedged claim phrasing (e.g. epistemic qualifiers), so you can measure how a model's willingness to trust or override a claim changes with the claim's confidence. The hedged phrasing lives in the Hedged text field of the corresponding paragraph.

Question strategies

Each item is generated with one of 11 strategies, each based on a distinct reasoning operation over tabular data (counting, aggregation, ranking, filtering, and so on). The Selected strategy field records which one was used.

Strategy	What it asks	Samples (%)
COMPARATIVE_DIFFERENCE	The numeric gap between two specified entities' values.	450 (7.6%)
EXTREME_NUMERICAL	The entity holding the maximum or minimum numeric value.	300 (5.1%)
TOP_K_RANKING	The top-k entities ranked by a numeric column.	285 (4.8%)
FILTER_COUNT_NUMERICAL	How many rows meet a numeric threshold condition (e.g. greater/less than a value).	278 (4.7%)
AGGREGATION	The sum (or other aggregate) of a numeric column across rows.	169 (2.9%)
EXTREME_TEMPORAL	The earliest or latest entity by a date column.	735 (12.4%)
FILTER_COUNT_TEMPORAL	How many rows fall within a specified time period or date condition.	725 (12.2%)
COUNT_TOTAL	The total number of entities in the group.	1,588 (26.8%)
CATEGORICAL_FILTER_LOOKUP	How many rows match a specific categorical value.	537 (9.1%)
UNIQUE_DIVERSITY	The number of distinct values in a categorical column.	437 (7.4%)
FREQUENCY_MODE	The most frequently occurring category value (the mode).	418 (7.1%)
Total		5,922 (100%)

Dataset statistics

CLECK draws its tables from a broad range of subject areas, reflecting the distribution of the underlying Wikipedia list pages. The largest are Infrastructure, Awards & Rankings, People & Roles, Countries & Geography, and Films & TV, with a long tail of smaller domains.

Subject-area coverage across the benchmark (item counts per domain).

Construction

CLECK is built with a four-stage pipeline: (1) structured table extraction and rule-based filtering from Wikipedia "List of..." pages; (2) strategy-guided question generation, a reasoning strategy is selected, and an LLM generates a question conditioned on the schema and strategy; (3) counterfactual answer construction, which deterministically computes the ground-truth answer and then derives two asymmetric variants — a claim-aligned answer (by swapping in an alternative valid value) and an evidence-aligned answer (by applying strategy-conditioned perturbations to the table); and (4) paragraph synthesis and validation, where the claim and evidence paragraphs are generated and then automatically validated to confirm each paragraph recovers its intended answer while the two conflict. All LLM-assisted components use Gemini 3 Flash Preview.

The CLECK construction pipeline. Full code is available in the KnowledgeMaze repository.

Intended uses

CLECK is intended for evaluating knowledge-conflict resolution, faithfulness to provided evidence, and the interaction between parametric memory, stated claims, and in-context evidence. It is a counterfactual benchmark: the in-context tables are deliberately altered away from reality, so items should be used to probe model behavior under conflict, not as a source of factual information.

Limitations and ethical considerations

• The Fake data tables are counterfactual by design and are not accurate statements about the world. Do not extract facts from them.
• Paragraph text is model-synthesized; any content derived from third-party models is subject to the respective providers' terms.
• Coverage reflects the sampled Wikipedia tables and question-generation strategies and is not a representative sample of all factual domains.

Source and attribution

The Original data tables, Wiki link references and ground-truth facts (Memory/Claim/Evidence answers) are derived from English Wikipedia, which is licensed under CC BY-SA 4.0. Per-item provenance is given by the Wiki link field.

The question text and Claim/Evidence paragraphs were synthesized from that data using Google's Gemini models (gemini-3-flash-preview, gemini-3.1-pro-preview). Use of this generated content is therefore also subject to Google's Gemini API terms of use, in addition to the license below.

License

This dataset is released under Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0), reflecting its Wikipedia-derived source material. Attribution to Wikipedia is provided per item via the Wiki link field; reuse of the dataset must preserve attribution and share-alike terms, and remains subject to Google's Gemini API terms of use for the LLM-generated portions.

The code and construction pipeline are hosted separately at github.com/danielhk753/KnowledgeMaze under the MIT license.

Citation

@inproceedings{hacker-bar-2026-saying,
    title = "Saying vs. Seeing: Evaluating {LLM} Conflict Awareness and Resolution Preference under Latent Claim-Evidence Disagreement",
    author = "Hacker, Daniel  and Bar, Kfir",
    booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
    month = oct,
    year = "2026",
    address = "Budapest, Hungary",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.emnlp-main.1727/",
    pages = "39397--39423",
    ISBN = "979-8-89176-459-0"
}