Where Academic Discussion answers lose points: language, not ideas

LingoLeap's grader scores every Academic Discussion answer on the three dimensions of the published rubric, each out of 5. Ties are common because the scores move in half points: 63.7% of answers have two or three dimensions sharing the lowest score.

Embed this chart

Paste this HTML into your page. It is a static image with a credit link: no script, no iframe.

<figure style="margin:0;max-width:880px">
  <img src="https://lingoleap.ai/research/charts/academic-discussion-subscores.svg" alt="The three sub-scores" width="880" height="540" style="max-width:100%;height:auto" loading="lazy">
  <figcaption>Source: <a href="https://lingoleap.ai/research/academic-discussion-subscores">LingoLeap, Where TOEFL Academic Discussion Answers Lose Points: Data</a></figcaption>
</figure>

/research/charts/academic-discussion-subscores.svg

Language Use has the lowest average in every overall score band (at the top, only 0.02 below Clarity), and it is also the dimension that climbs furthest: from 3.14 in the 22 or below band to 4.24 in the 28-30 band, a gain of 1.10 points, against 0.99 for Relevance and 1.01 for Clarity.

11.1% of answers score a full 5.0 on Relevance. On Clarity and on Language Use, fewer than 30 of 3,683 answers (under 1%) do. Relevance is the only dimension on which this grader gives full marks with any regularity; on the other two, a 5.0 is almost never awarded to these answers.

Sample

Every non-legacy TOEFL Academic Discussion task in LingoLeap's sample library, and the first 20 published sample answers for each: 3,729 sample pages. After exclusions, 3,692 answers across 1,467 tasks (mean score 25.17, SD 2.42; median length 154 words). These are real learner answers that LingoLeap publishes as samples, not every answer ever submitted.

Period

Retrieved 26 Sep 2026. The endpoint does not expose when an answer was written, so submission dates are not recorded.

Source

LingoLeap's public sample endpoint (the same data the sample pages show), read at no more than two requests a second: 3,723 of 3,729 pages returned a record.

How scores were produced

LingoLeap's AI estimate on the 0-30 scale and three rubric sub-scores on 0-5 (Relevance and Contribution to Discussion; Clarity and Elaboration of Viewpoint; Language Use and Grammar), parsed from the grader's report. Not ETS scores, and not on the 1-6 band scale ETS reports for tests from 21 January 2026.

Exclusions

13 records with no answer text and 8 grader failures (overall 0 with every sub-score 0). Any group smaller than 30 answers is not reported.

Limitations

Observational: the pattern could come from the learners or from how LingoLeap's grader is calibrated, and this data cannot tell which. Published samples average 25.2 of 30, a stronger-than-typical group. LingoLeap AI scores, not ETS.

Privacy

Only aggregates are published. No answer text, names, or individual scores appear on this page or in the download.

Lowest sub-score

For each answer with all three sub-scores (3,683), the dimension with the lowest score is recorded; if two or three share it, the answer counts as a tie. "Lowest or tied" counts every answer where Language Use is at the minimum, alone or shared.

1. 1

   Not ETS

   These are LingoLeap AI sub-scores on the published rubric's three dimensions. ETS's human raters and e-rater may weight the dimensions differently.

2. 2

   Learners or grader?

   A grader that is stricter on language than on content would produce the same table. This study cannot separate the two.

3. 3

   Not every learner

   The answers are LingoLeap's published samples, which average 25.2 of 30. Weaker answers, which may fail on relevance too, are under-represented.

4. 4

   No submission dates

   The public sample data does not say when each answer was written, so the study has a retrieval date, not a date range.