Personalized sentence order improved understanding in a controlled study

Same information, different sequence

Can an explanation work better simply because its sentences appear in an order suited to its reader? In a preprint, Deo Munduku and Elsa Negre of Paris-Dauphine University investigate whether people’s earlier written interpretations can guide that ordering.

The distinction is important: the researchers did not personalize which facts readers received or rewrite the wording. They changed only sentence order. Their question was whether patterns extracted from previous writing could help organize a new explanation more effectively than alternative procedures.

Building a profile from earlier writing

The study began with four interpretation tasks involving an outdoor activity, a meal-kit service, a mobile-phone plan and a home-cleaning service. Of 135 people who started this phase, 122 completed all four tasks and subsequently took part in the explanation evaluation. Eighty worked in French and 42 in English.

Each situation presented six kinds of information: price, access or travel, core features, experience conditions, availability, and usage or contractual conditions. Participants wrote open-ended interpretations rather than explicitly ranking these categories.

An automated procedure estimated which information their responses mentioned, its relative prominence, and which pieces appeared together within text segments. The researchers call the last measure “co-mobilization.” It records joint appearance, not whether two facts have a causal relationship or whether someone regards one as outweighing another.

The four responses were combined into a fixed historical profile. This was a summary of observed writing, not a measurement of an immutable psychological trait. Profiles were completed before the evaluation and were not updated using its results.

The personalized ordering rule placed sentences according to estimated historical prominence, then brought the most strongly co-mobilized eligible pair together. These prominence estimates were textual proxies, not importance ratings supplied by participants.

Four ways to arrange six sentences

The evaluation used four new scenarios involving a film screening, exhibition, cooking workshop and concert. Every explanation contained six predefined sentences. Within a given scenario and language, their content, wording and visual presentation stayed constant across ordering procedures.

The researchers compared four approaches:

• OWN: the reader’s historical profile, applied through the fixed ordering rule.
• DONOR: another participant’s profile from the same language group, using that rule.
• POP: a profile aggregated from the other 121 participants, again using the rule.
• PROFILE-LLM: a Mistral language model given the reader’s profile, target context and sentences, returning only a sentence ordering.

Each participant encountered every procedure and every scenario once, with assignments and trial order counterbalanced. They did not repeatedly read different versions of the same scenario.

Reading was self-paced. The explanation then disappeared and could not be revisited. Participants reported whether they would accept the recommendation and answered two multiple-choice questions: one about a factual detail and another requiring information to be combined.

Higher scores, not greater acceptance

Across the study’s 976 comprehension responses, OWN achieved 86.5% correct answers. DONOR reached 66.0%, POP 63.9%, and PROFILE-LLM 73.0%. OWN’s advantage was therefore 20.5, 22.6 and 13.5 percentage points, respectively.

A mixed-effects analysis accounting for participant differences, question identity, language and trial position supported all three comparisons after correction for multiple testing.

OWN had the highest descriptive accuracy for both factual and information-combination questions. However, the analysis did not establish that the ordering effect differed between those question types. Participants’ ratings of understanding and sequence naturalness also showed overall differences between procedures.

Recommendation acceptance told a different story. The study detected no reliable overall effect of ordering on acceptance. Nor did it establish an overall mental-effort effect after statistical correction. Better comprehension scores should therefore not be equated with greater persuasion or proven decision quality.

What remains uncertain

These results come from a controlled study reported as a preprint, not an established general finding. Recruitment used convenience channels, evaluation covered only four leisure and cultural scenarios, and each profile drew on just four earlier responses.

Because questions followed removal of the explanation, scores reflect short-term retention as well as comprehension. The automated extraction procedure was not independently revalidated on this study’s writing.

The experiment also cannot identify which profile component caused the advantage. Historical prominence determined most ordering structure; the pairing step changed only 28 of 122 OWN orders. There was no direct comparison with personalization based on explicitly stated preferences.

Finally, PROFILE-LLM tested one model and prompting procedure, not language models generally. It even scored descriptively higher than OWN in the concert scenario. The supported conclusion is narrower: under these conditions, readers’ earlier writing provided a useful signal for arranging unchanged explanations.