FlowingFlowing
← Back to all development cases
Development case · blind model judgment pending independent human review; not a formal Benchmark conclusion.
AutocompleteNeither version usable133 / 162 · 466e301aa97ae184

Towards Human-centered Explainable AI: A Survey of User Studies for Model Explanations

Psychology and cognitive neuroscience · 2210.11584v5

FLOWING EVIDENCE BENCHMARK

How do we tell whether Evidence helps?

At the same writing position, with the same model and task, how does supplying retrieved paper passages change the first output? We compare matched versions and retain ties, unusable outputs and incomplete reviews.

SAME DRAFT · EVIDENCE ON OR OFF

01 · FIXED WRITING POSITION

MANUSCRIPT

Same writing position ▌

Matched autocomplete at the same draft position

02 · TWO MATCHED INPUTS

SHARED BY BOTH

Manuscript context, model, task and prompt

A · With Evidence

Retrieved paper passages supplied

B · Without Evidence

No retrieved passages supplied

LLM

Same model and version

A → first output

B → first output

Blind judge agent

First outputs are anonymized as X and Y

Continuation review: accuracy, fit to the writing task and usability
Returns: X preferred / tie / Y preferred / both unusable

Order check: X / Y → Y / X

HOW DOES BLIND JUDGING WORK?

① Anonymize both outputs
The judge sees the same draft and both first outputs without knowing which received Evidence.

② Compare and swap order
The judge applies task-specific criteria in X/Y and then Y/X order.

③ Review disagreements
A third pass resolves disagreements. Incomplete reviews remain in the denominator.

How are positions and cards selected?

The evaluation uses seven fields with forty fixed ordinary autocomplete positions in each, for 280 total. A position does not need to require a literature citation in its next sentence. Each Flowing request receives three to five retrieved cards: five when all pass the relevance gate, without padding with weak cards. The run supplied 1,355 cards: 244 positions received five, 27 received four, and nine received three.

ADOPTION · 150/280

Adoption coverage

After drafting, the model returns an evidenceUse array identifying cards that materially affected facts, correction, direction, scope, terminology, or expression. A nonempty array counts as adoption.

OUTCOMES · 70 / 7 / 30

Full outcomes after adoption

Paired outcomes are counted only when the model reports adoption and both arms produce text. Among 107 comparable positions, Flowing was preferred in 70, 7 tied, and no-Evidence was preferred in 30; one tie was unusable in both arms.

How are the two table percentages calculated?

ADOPTION COVERAGE

150adopted
÷
280all positions
=
54%

FLOWING PREFERENCE AFTER ADOPTION

70Flowing preferred
÷
107comparable adopted positions
=
65%

evidenceUse is the model’s structured report of its own generation process. It does not require a citation and is not rejected merely because a faithful source paraphrase is not verbatim. Actual manuscript citations continue to use separate strict source verification. The previous 140-position result remains retained for traceability.

These are development-stage model-reported adoption and blind-judge results pending independent human review. They are not formal Benchmark conclusions and do not, on their own, establish causality.

Manuscript writing position

Text excerpt · not a PDF page

Research manuscript · excerpt

Towards Human-centered Explainable AI: A Survey of User Studies for Model Explanations

VII Future Research Directions

…lidity of this aggregation with reliability measures such as the tau-equivalent reliability (also known as Cronbach’s α \alpha ). For example, if objective and subjective measures of a quantity, such as understanding are combined, it is necessary to verify that there is sufficient agreement. If multiple items (e.g., data samples or visualizations) are rated by several subjects, statistics such as Cohan’s κ \kappa as Fleiß’s κ \kappa for more than two raters [ 181 ] can be used to assess agreement beyond chance between these raters and serve as an indication for the reliability of the ratings.

… to each treatment group, their recruitment, consent and incentivization, and the exact treatment conditions they are subjected to. Furthermore, some descriptive statistics of the collected data can help readers assess the characteristics of the adequacy of the statistical tools used. Regarding the analysis, we found it important to mention how the underlying assumptions of the statistical tests used were checked and to mention the exact variant of the test used (e.g., stating “a two-way ANOVA with the independent variables X and Y” is used instead of just mentioning that ANOVA-test is used).

↓ Next paragraph was generated here

The highlighted text is near the writing position. This is a layout of the evaluation input, not a PDF screenshot.

A matched comparison at the same writing positionJudgment: Neither version usable

Both arms use the same model, task and manuscript context. Only the retrieved passages differ. The first recorded output from each arm is shown below.

Evidence

Same model and position · retrieved passages supplied · first output

Judged unusable

No continuation text was produced.

Model action：complete

Passages supplied to the Evidence version

A one-way ANOVA test for functional data with graphical interpretation

In order to apply the rank envelope test for an ANOVA design, it is necessary to apply the extension of the rank envelope test described in Mrkvička et al. (2016) . This extension is the rank envelope test applied on several observed test functions at once. This extension is in fact a multiple testi…

Read full passage excerpt

In order to apply the rank envelope test for an ANOVA design, it is necessary to apply the extension of the rank envelope test described in Mrkvička et al. (2016) . This extension is the rank envelope test applied on several observed test functions at once. This extension is in fact a multiple testing procedure which combines the output from several separate tests. The output of this combined rank envelope test is the set of envelopes with the same interpretation as above. Also the extreme length p p -value can be obtained. It is important to mention here that the graphical interpretion automatically identifies which of the tested functions are responsible for the potential rejection and also it identifies which parts of the functions are responsible for the rejection.

Model-reported card use

The model did not report that any card materially affected this completion.

No Evidence

Same model and position · no retrieved passages · first output

Judged unusable

No continuation text was produced.

Model action：complete

Result of this comparison

The blind model judge marked both outputs unusable; this is counted separately from a tie.

Read the original model judgment

Both outputs return empty text with action 'complete', failing to provide the requested autocomplete paragraph. The task explicitly requires continuing the manuscript with one coherent academic paragraph. The blank paragraph appears between two existing paragraphs discussing (1) reliability measures and inter-rater agreement, and (2) experimental design, participant details, and statistical analysis reporting. The sources provide factual context about ANOVA testing for functional data with graphical interpretation. A usable autocomplete would need to bridge these topics—likely transitioning from reliability/agreement measures to experimental design and statistical analysis details, or expanding on statistical reporting practices. Both outputs completely fail to generate any content, making them unusable for the autocomplete task.

The passages are input material; their presence does not mean an output used them or that they support every claim. Draft and source passages are excerpted; no simulated scores or PDF appear here.