Narrative inclusion

F1 measures how well the answer cites articles associated with the requested narrative while omitting other narratives.

Narrative exclusion

F1 measures whether the answer covers other relevant narratives while omitting the specified one.

Include Missing Narratives

Recall measures the share of narratives missing from a partial answer that the model adds in the next turn.

Present diverse narratives

Baseline prompt recall and diverse prompt recall measure coverage of available narratives. The diverse prompt explicitly requests narrative diversity.

Preserve coverage under confirmatory queries

The absolute recall gap compares neutral queries with queries that favor a particular narrative. A smaller gap means more consistent coverage across framings.

Provide comparable access across languages

The query-language score gap compares scores by the language of the question. The target narrative-language score gap compares scores by the publication language of the requested narrative. Both use English as the reference.

Recall is the share of required narratives covered. F1 balances recall with precision, the share of cited narratives that match the task. Displayed ties share a rank.

Language-gated scoring. The scorer checks whether the answer is in the language of the question. Answers identified as different-language or mixed-language receive zero. Answers with an uncertain classification proceed to citation-based scoring. Cited articles can be in any language. The language violations column counts different-language and mixed-language answers.

The unfiltered scores reproduce Table 1 in the paper, including standard errors in parentheses. Filtered cells show point estimates. For a selected non-English query language, the query-language score gap compares its mean score with English within the same topic scope. Framing and target narrative-language gaps require cases in both comparison groups. Score gaps are reported in percentage points.

Evaluation documentation