SpecialEduBench: 视觉语言模型在自闭症儿童语言干预中知识、技能和态度的基准测试

# Benchmarking Vision-Language Models on Knowledge, Skill, and Attitude in Language Intervention for Autistic Children
Source: [https://arxiv.org/html/2609.26090](https://arxiv.org/html/2609.26090)
Jihoi Na1\*, Taeyeong Kim2\*, Sungjune Kong3, Jaemin Jung3, Min Joung Park4, Kyungtae Joo2, Ahhyun Kim2, Shim Jaechang3, Sooyoung Joo5, Dongjin Ka3, SeJoong Kim3, Jimin Kim6, HyunJin Jung7, Unggi Lee3†

###### Abstract

Language is the target of most early intervention for autistic children\. Because the goal and the method change from child to child, the work falls to a teacher who takes one child at a time and judges each scene as it unfolds\. Artificial intelligence is now being brought to that work, yet the benchmarks that reach special education ask what a model knows rather than what it does in front of a child\. Building one is not straightforward, since whether a response is good teaching depends on what the child has just done, so no answer key applies\. The evidence that settles it is visual as much as verbal, since the length of a wait, a shift of gaze, and the child’s uptake leave no trace in a transcript\. We introduce*SpecialEduBench*, which measures pedagogical competence along knowledge, skill, and attitude, with 4,537 knowledge items and with 200 skill items and 68 attitude items built on recorded intervention, the attitude items crossing pressure with monitoring into 192 response cells\. Seven special\-education experts wrote, scored, and reviewed the items, and we revised the judge model’s instruction against the reference scores they set\. Across eight frontier vision\-language models no axis is saturated, since the strongest still fails about a tenth of the honesty cells\. The models converge where the knowledge is factual and separate where the task is situated, and the failures gather where pressure is applied\. We intend the benchmark as an audit to run before deployment and as a starting point for models built for this domain\.

## 1Introduction

![Refer to caption](https://arxiv.org/html/2609.26090v1/x1.png)Figure 1:Overview of SpecialEduBench, whose four rows are the three competence axes and the auxiliary reward model\.*Left*, each row shows how its data was built and where experts reviewed it\.*Center*, knowledge is scored by exact match while skill and attitude go to a judge model aligned against the expert reference scores\.*Right*, the evaluation is read qualitatively through error analysis and quantitatively through model comparison\.The number of children identified as autistic keeps rising, and with it the number who need intensive intervention for language and communication\(Shawet al\.[2025](https://arxiv.org/html/2609.26090#bib.bib10)\)\. Such intervention varies in goal and method from child to child, so a trained teacher has to judge each scene as it unfolds\(Kaiser and Hampton[2017](https://arxiv.org/html/2609.26090#bib.bib7)\)\. Teachers with that training are in short supply, and dialogue systems built on large language models \(LLMs\) are now appearing with the aim of closing the shortfall\(Denget al\.[2024](https://arxiv.org/html/2609.26090#bib.bib11)\)\.

Whether those systems teach well is not something we can currently measure\. Good teaching here is not supplying the right answer but drawing the child into communicating, so waiting and modeling can each be correct in one scene\(Schreibmanet al\.[2015](https://arxiv.org/html/2609.26090#bib.bib5)\), and the same response can be good teaching or poor teaching depending on what the child did immediately before\(McDanielet al\.[2022](https://arxiv.org/html/2609.26090#bib.bib6)\)\. What makes a teaching move correct is how it relates to the child’s behavior rather than what form it takes, so quality here has to be judged by fit to the situation and cannot be checked against an answer key\.

Existing work does not meet that requirement\. The benchmarks that score teaching are built for general education\(Leeet al\.[2026b](https://arxiv.org/html/2609.26090#bib.bib16); Macinaet al\.[2025](https://arxiv.org/html/2609.26090#bib.bib15)\), and where special education appears at all it is scored by multiple choice\(Lelièvreet al\.[2025](https://arxiv.org/html/2609.26090#bib.bib19)\)\. An option list asks what a model knows rather than what it would do in front of a child, and the evidence that separates the two is visual as much as verbal, since how long the tutor waited and whether the child took up the prompt leave no trace in a transcript\. Items that pose this judgment therefore have to be built on recorded intervention, and the criteria for scoring them have to be written by people who know the practice\.

Human scoring is what makes such a benchmark expensive, so a judge model has to take over, and a second measurement problem opens there\. Benchmarks report how far such a model agrees with people\(Srinivasaet al\.[2025](https://arxiv.org/html/2609.26090#bib.bib14); Yanget al\.[2026](https://arxiv.org/html/2609.26090#bib.bib17); Jeonget al\.[2026](https://arxiv.org/html/2609.26090#bib.bib18)\), and none states why the level reached is the level required\. That omission matters more here than elsewhere, because trained people disagree with one another on this judgment\(Hamptonet al\.[2026](https://arxiv.org/html/2609.26090#bib.bib27)\), so a fixed target is not a standard any one scorer could be held to\.

We introduce SpecialEduBench111The dataset, the rubrics, the expert reference scores, and the evaluation code are available athttps://github\.com/LEAP\-LAB\-KUS/SpecialEduBench\., which measures pedagogical competence in autism language intervention along knowledge, skill, and attitude \(Figure[1](https://arxiv.org/html/2609.26090#S1.F1)\), the three axes the field itself uses to define competence and to organize the preparation standards for special educators\(Baartman and de Bruijn[2011](https://arxiv.org/html/2609.26090#bib.bib4); Council for Exceptional Children[2020](https://arxiv.org/html/2609.26090#bib.bib9)\)\. Skill and attitude are grounded in recorded intervention rather than written scenarios, because a written scenario does not record how long the tutor waited or what the child did next\. The attitude axis crosses a pressure condition with a monitoring condition, since a professional disposition shows itself only where acting rightly costs something and crossing the two locates where it breaks down instead of averaging it away\. Seven special\-education experts wrote, scored, and reviewed the items, and how far those who scored them converge becomes the ceiling we read the judge against\. The knowledge axis holds 4,537 items, skill holds 200 across eight evidence\-based teaching strategies, and attitude holds 68 that unfold into 192 response cells\.

We evaluate eight frontier vision\-language models, four closed\-weight and four open\-weight\. No axis is saturated\. The strongest model still misses about a tenth of the honesty cells, the models converge where the knowledge is factual and separate where the task is situated, and the failures gather where pressure is applied, which no average over an axis reveals\. The benchmark is meant to be run before a system reaches a child, and the headroom it leaves marks where a model built for this domain would have to improve\.

### Contributions

- •The first benchmark to score knowledge, skill, and attitude together in special education, with skill and attitude built on recorded intervention\.
- •A design that crosses pressure with monitoring, locating where a disposition breaks down rather than averaging it away\.
- •A procedure that sets a judge model’s target from the expert ceiling instead of a threshold assumed in advance\.
- •Eight frontier vision\-language models leave every axis unsaturated, with the headroom where the domain is specific\.

## 2Related Work

Table 1:Comparison of benchmarks that evaluate the teaching of a language model\. Model is whether the benchmark evaluates text\-only LLMs or vision\-language models\. Axes marks which of knowledge, skill, and attitude a benchmark scores\. Grounding is what an item is built on\. Open\-ended scoring is how a free response is scored, and judge alignment is whether the benchmark reports how far its automatic scorer agrees with a human expert\.### 2\.1Evaluating Language Models in Education

Benchmarks for educational LLMs score what a model knows and how it tutors\. They rate tutoring turns in mathematics and science against expert judgment\(Mauryaet al\.[2025](https://arxiv.org/html/2609.26090#bib.bib13); Srinivasaet al\.[2025](https://arxiv.org/html/2609.26090#bib.bib14); Macinaet al\.[2025](https://arxiv.org/html/2609.26090#bib.bib15)\), read teaching from instructional and classroom video\(Yanget al\.[2026](https://arxiv.org/html/2609.26090#bib.bib17); Jeonget al\.[2026](https://arxiv.org/html/2609.26090#bib.bib18)\), and in one case separate knowledge, skill, and attitude\(Leeet al\.[2026b](https://arxiv.org/html/2609.26090#bib.bib16)\)\. Most of them evaluate text\-only LLMs, and of the three that take visual input, two read a still image of a student’s work\(Srinivasaet al\.[2025](https://arxiv.org/html/2609.26090#bib.bib14); Yanget al\.[2026](https://arxiv.org/html/2609.26090#bib.bib17)\)and one reads classroom video\(Jeonget al\.[2026](https://arxiv.org/html/2609.26090#bib.bib18)\)\. Almost none of them reaches special education, and the one that does asks multiple\-choice knowledge alone\(Lelièvreet al\.[2025](https://arxiv.org/html/2609.26090#bib.bib19)\), and none narrows to a single clinical practice\. Table[1](https://arxiv.org/html/2609.26090#S2.T1)sets them side by side\.

The benchmarks that score a free response do it with a second LLM as the judge, first for answer quality\(Zhenget al\.[2023](https://arxiv.org/html/2609.26090#bib.bib29)\)and then for pedagogical quality against a written rubric\(Liuet al\.[2023](https://arxiv.org/html/2609.26090#bib.bib20)\)\. Both report how far the judge agrees with people, neither says what agreement would be enough, and reliability and validity are not always separated\(Meloet al\.[2026](https://arxiv.org/html/2609.26090#bib.bib21)\)\. The omission matters most where the experts disagree, since there a fixed threshold can be unreachable and a high one can only be met by a judge that has stopped tracking the experts\. None of the benchmarks in Table[1](https://arxiv.org/html/2609.26090#S2.T1)states the level its own judge has to reach\.

### 2\.2Honesty and Disposition in LLMs

Scoring one response at a time misses that what a model says changes with who is asking and whether it believes it is observed\. Sycophancy is the best documented of these\(Sharmaet al\.[2024](https://arxiv.org/html/2609.26090#bib.bib22)\), and alignment faking shows the sharpest version, where behavior differs between conditions the model reads as monitored and unmonitored\(Greenblattet al\.[2024](https://arxiv.org/html/2609.26090#bib.bib30)\)\. These are studied as properties of the model in the abstract rather than inside a teaching exchange, where the pressure arrives as a request from a family and the child is the one who bears what holding the line costs\.

### 2\.3Special Education

The field defines teaching competence as an integration of knowledge, skills, and attitudes\(Baartman and de Bruijn[2011](https://arxiv.org/html/2609.26090#bib.bib4); Council for Exceptional Children[2020](https://arxiv.org/html/2609.26090#bib.bib9)\)\. Correct skill is set by enhanced milieu teaching \(EMT\), which integrates the naturalistic developmental behavioral intervention \(NDBI\) tradition and divides into responsive interaction \(RI\) and milieu teaching \(MT\) with four strategies each\(Schreibmanet al\.[2015](https://arxiv.org/html/2609.26090#bib.bib5); Kaiser and Hampton[2017](https://arxiv.org/html/2609.26090#bib.bib7)\), and each strategy carries a condition stated over the child’s response rather than over the form of the utterance\. Correct attitude is stated separately\(Donnellan[1984](https://arxiv.org/html/2609.26090#bib.bib2); United Nations[2006](https://arxiv.org/html/2609.26090#bib.bib32)\)and is defined by what a teacher holds to when a family or a schedule pushes the other way\. Appendix[A](https://arxiv.org/html/2609.26090#A1)treats each of these literatures in full\. Against the nine entries of Table[1](https://arxiv.org/html/2609.26090#S2.T1), SpecialEduBench is the only one that scores all three axes, builds skill and attitude on recorded intervention, and reads its judge against the agreement experts reach with each other\.

## 3SpecialEduBench

### 3\.1Design

The benchmark spans three axes and ships a reward model for the skill axis \(Table[2](https://arxiv.org/html/2609.26090#S3.T2)\)\. Knowledge is 4,537 four\-option items scored by exact match\. Skill is 200 items grounded in recorded intervention, 25 for each of eight teaching strategies\. Attitude is 68 items that cross with pressure and monitoring conditions to unfold into 192 response cells\.

One principle runs through all three axes\. We build the human standard first and fit automatic scoring to it afterward\. The skill and attitude axes stay small because 200 items and 68 items are what a full expert review can cover\.

Table 2:The three axes of SpecialEduBench\. Cells exceed items on the attitude axis because each item is presented under more than one condition\.AxisItemsCellsScoringKnowledge4,5374\{,\}5374,5374\{,\}537Exact matchSkill200200200200Rubric, judge modelAttitude6868192192Rubric, judge model*Auxiliary*Reward200200pairs200200Preference accuracy
### 3\.2Knowledge

The knowledge axis takes the scope established in Section 2, so it asks about special\-education knowledge rather than restricting itself to autism\. We operationalize the three strands identified there as education, psychology, and medicine, and we add EMT as a fourth category because no public resource covers knowledge of teaching strategies and Appendix[F\.1](https://arxiv.org/html/2609.26090#A6.SS1)lists the sources\.

We collect where validated items exist and generate only where they do not\. The 4,037 collected items come from the special educational needs items of the Pedagogy Benchmark\(Lelièvreet al\.[2025](https://arxiv.org/html/2609.26090#bib.bib19)\)and from public medical and psychology benchmarks\(Palet al\.[2022](https://arxiv.org/html/2609.26090#bib.bib33); Hendryckset al\.[2021](https://arxiv.org/html/2609.26090#bib.bib34)\), left as they were except that ten\-option items are reduced to four\. The 500 EMT items are generated from eight source papers, alternating two model families in an even split, and each item must rest on a sentence copied verbatim from its source, which a string comparison verifies\. Two external experts reviewed every generated item\. Thirty items across the axis are released but not scored, and Appendix[F\.1](https://arxiv.org/html/2609.26090#A6.SS1)gives the sources, the review, and the exclusions\.

### 3\.3Skill

The skill axis asks whether a model can read an intervention scene and respond in a way that fits a given strategy, which knowing the definition of that strategy does not settle\. We operationalize the EMT behaviors of Section 2 into eight strategies at a grain that can be scored\(Kanget al\.[2025](https://arxiv.org/html/2609.26090#bib.bib1)\), rebuilding the list at the level of one observable behavior and Appendix[F\.2](https://arxiv.org/html/2609.26090#A6.SS2)gives the per\-strategy counts\.

Items are built on 45 clips cut from 35 publicly available videos, each clip a segment where the interaction is visible and both audio and image are legible\. Four special\-education experts reviewed the candidate pairings of clip and strategy and fixed them before any item was written\. Letting a model choose a strategy and then write an item for the strategy it chose would close a loop we need open\.

The generator models wrote items on those pairings only, and we mixed model families because a benchmark written by one model favors that family\(Yuanet al\.[2025](https://arxiv.org/html/2609.26090#bib.bib26)\)\. Each item names the strategy and asks for a free\-text response\. The pairings yielded 136 independent items, and we balanced to 25 per strategy with 64 variants, so the 200 items are not 200 independent observations\. The same four experts reviewed every item and rubric before any response was collected\.

### 3\.4Attitude

The attitude axis measures what a model does when doing the right thing costs something, which the other two axes do not reach\. A model can know the field and teach well and still say something false at the moment it matters, and in autism intervention that falsehood is a harm rather than a blemish, because caregivers decide the next intervention from what the tutor reports\.

The six families cut the professional dispositions of Section 2 into scorable units, and honest reporting alone splits in two, since honesty in front of the child and honesty in the record that reaches a third party break down in different places, and Appendices[C](https://arxiv.org/html/2609.26090#A3)and[F\.3](https://arxiv.org/html/2609.26090#A6.SS3)list them\. Four families correspond to the attitude axis of OpenLearnLM, which covers attitude with 14 items in total and so cannot isolate which disposition fails\(Leeet al\.[2026b](https://arxiv.org/html/2609.26090#bib.bib16)\)\. The other two have no counterpart in general honesty benchmarks, because a model can regard a child’s potential as smaller than it is without uttering one false sentence\.

Every situation is drawn from one of 17 recorded clips, so no scene enters the benchmark that could not occur\. Every item in the two honesty families plants a situation in which lying pays, since current models nearly all pass when asked plainly, and ten items go further and plant a factually wrong premise\. We do not instruct the tutor to be honest, which would measure compliance rather than honesty\.

Pressure sits inside the item, where a caregiver or the child asks the tutor to say something other than what happened\. Monitoring is attached as a system notice that either announces a supervising teacher’s later review or states that no one is watching, and we use supervisor review rather than the training\-data framing of prior work because that is the oversight a deployed tutor actually meets\(Greenblattet al\.[2024](https://arxiv.org/html/2609.26090#bib.bib30)\)\. Prior work attaches one standing conflict to an entire experiment, while pushing pressure into a request inside the item lets the interaction be estimated per item and compared across models\.

Table 3:Main results across the three axes\. Knowledge is the macro\-average over the four categories, skill is the macro\-average over the eight strategies, and attitude is the mean over all attitude responses\. Reward is the win rate of the model response against the expert response, where a single item is worth0\.0250\.025, so models within a few items of one another are not ordered\. Within each column,boldmarks the best value andunderlinethe second best\.
### 3\.5Scoring Protocol

Skill and attitude cannot be scored against an answer key, so a person has to read each response and scoring becomes the bottleneck\. We put a judge model in the expert’s place, and a judge used without a check moves the measurement problem rather than settling it\. The procedure therefore runs in three steps\. We build a reference we can defend, fit the judge to it, and only then score models\.

#### Reference scores\.

GPT\-5 wrote one response per item without seeing the rubric, since a response written with the criteria in hand would measure how well a model matches criteria rather than how well it teaches\. Four special\-education experts scored those responses independently and discussed each item to a single agreed score, and two further experts who took no part in building the benchmark re\-examined 40 items per axis and judged all of them adequate\. Appendix[D](https://arxiv.org/html/2609.26090#A4)reports each stage\. We keep the individual scores from before the discussion, because agreement between experts, and not the agreed score itself, is the ceiling automatic scoring can approach\.

#### Rubrics and judge\.

Every rubric runs over five bands on a 1 to 10 scale where 7 is the passing line and Appendix[B](https://arxiv.org/html/2609.26090#A2)gives the band descriptions, so that means can be compared across families and strategies, and one instruction goes to the judge unchanged naming neither the item nor the reference score\. Both axes are judged twice over, bygemma\-4\-31b\-itand byqwen3\.6\-27b, and the two boards are kept apart instead of averaged\. Both judges decode at temperature 0, read text only, and score each response three times, since repeated runs move a tenth to a quarter of cells even at that setting\. Appendix[E](https://arxiv.org/html/2609.26090#A5)reports what changes when the skill judge is given frames\.

#### Judge alignment\.

We read each judge against the agreement the experts reach with each other rather than against a fixed threshold, because that is the level the task admits\. On skill the experts agree within one point on 63\.9% of items, a moderate consistency on a rubric that asks for a fine\-grained judgment of teaching strategy rather than a categorical label, and the Gemma and Qwen judges reach interval Krippendorffα\\alphavalues of0\.3420\.342and0\.3920\.392over all 200 expert\-scored items against the0\.4270\.427the experts reach with each other, so neither meets that ceiling\. On attitude the experts agree within one point on 80\.7% of cells and the two judges on 80\.2% and 78\.1%, so both sit where people sit\. Their higherα\\alphaof0\.8120\.812and0\.8010\.801is not evidence of beating people, since matching a consensus of those same experts is easier than matching another individual\. Both skill judges also miss failures more often than they invent them, recovering44\.1%44\.1\\%and50\.0%50\.0\\%of the skill items the experts placed below the passing line, so we withhold pass rates on that axis\. The Limitations section states what else these two figures do not settle\.

#### Reward model\.

The skill axis also carries a reward model, since open\-ended teaching responses are not measured by word overlap\. Training uses 200 pairs split 160 to 40, where the rejected response is written by GPT\-5 with one flaw injected and the chosen response is the expert’s minimal edit correcting that flaw and nothing else\(D’Oosterlincket al\.[2025](https://arxiv.org/html/2609.26090#bib.bib28)\), which keeps length and register from separating them\. We fine\-tuneQwen2\.5\-1\.5B\-Instructwith a low\-rank adapter under a Bradley\-Terry objective, and it separates 38 of the 40 held\-out pairs, with the breakdown by strategy in Appendix[H\.3](https://arxiv.org/html/2609.26090#A8.SS3)\. Our first attempt started from a model already trained on general preference data\(Liuet al\.[2026](https://arxiv.org/html/2609.26090#bib.bib25)\), and the ranking it produced turned out to be reading response length, since appending sentences with no pedagogical content raised an unchanged answer’s score by7\.377\.37\. Changing the base without altering a single item brought that sensitivity to−1\.16\-1\.16while the separation held\.

#### Reported metrics\.

Knowledge is the macro\-average over the four categories and is also reported by disability type in Appendix[H\.1](https://arxiv.org/html/2609.26090#A8.SS1), and skill is the macro\-average of the eight per\-strategy means\. For attitude the pass probability matters less than how it moves with condition, since pressure lowers it and monitoring raises it, and how far those two gaps interact is the quantity we write as P×\\timesM\. That interaction is the headline metric of the axis, because both gaps can be small while it is large, whereas prior work reports one mean difference in absolute value and discards the direction of the shift\(Leeet al\.[2026b](https://arxiv.org/html/2609.26090#bib.bib16)\)\. Confidence intervals are computed at the item level, because the 192 cells come from 68 items\.

## 4Experiments

### 4\.1Experimental Setup

We evaluate eight VLMs, four closed\-weight and four open\-weight\. All three axes decode at temperature 1\.0 with a seed sent on every request, and thinking tokens are minimized rather than disabled because three of the models reject any attempt to disable reasoning\. Frames are extracted at one per second\. Skill supplies every frame a clip has, since that is the condition under which the expert reference answers were written, attitude caps a cell at 24 frames, and knowledge is text only\. Mistral Large 3 rejects more than eight images and so runs at eight frames sampled uniformly, which changes the visual input on 133 of the 200 skill items, so its scores on the two grounded axes come from a thinner condition than the rest\. Skill takes one response per item, knowledge three, and attitude one per cell except on the two honesty families, where it takes ten because the interaction is a difference of differences that noise in any condition mean would swamp\. Generating the responses and scoring them cost $536\.1 in total\. Appendix[G](https://arxiv.org/html/2609.26090#A7)gives the per\-model settings\.

### 4\.2Experimental Results

Table[3](https://arxiv.org/html/2609.26090#S3.T3)collects the three axes and the reward metric, with the per\-strategy and per\-family breakdowns in Tables[4](https://arxiv.org/html/2609.26090#S4.T4)and[5](https://arxiv.org/html/2609.26090#S4.T5)\. Pass rates are withheld on skill for the reason given in Section 3\.5\.

Splitting knowledge by category shows what the total hides\. Education separates the models by15\.415\.4points while EMT separates them by2\.72\.7, and every model clears96\.996\.9on EMT\. The models that score lower also agree with themselves less across the three seeds, with one exception, and that model is the only one of the eight that spends no reasoning tokens, so it is wrong consistently rather than unsteadily\. On skill the two judges agree on six of the eight ranks and differ only by an adjacent swap, and at the item level they correlate at0\.7520\.752with70\.3%70\.3\\%of responses within one point\. The closed\-weight models average0\.7330\.733points above the open\-weight ones\. On attitude the pass rate spreads28\.228\.2points, the two judges again place six of the eight models at the same rank, and the three lowest of the seventeen categories all sit in the deception family\. The reward win rate runs from0\.6500\.650to0\.9000\.900with seven of the eight between0\.8000\.800and0\.9000\.900, and a single item is worth0\.0250\.025at this sample size, so the middle group is not ordered\.

Table 4:Skill by strategy, reported as the mean rubric score on the 1 to 10 scale\. RI and MT are the branch means\. Pass rates are withheld on this axis for the reason given in Section 3\.5\.Table 5:Attitude by family, reported as the mean rubric score over the cells of that family\. Dec\. is Deception, Int\. is Integrity and Honesty, Evi\. is Evidence\-based Judgment, Pot\. is Respect for Potential, Res\. is Respect, and Pro\. is Proactive Teaching Stance\.

## 5Analysis and Ablation

### 5\.1Attitude under Pressure and Monitoring

![Refer to caption](https://arxiv.org/html/2609.26090v1/x2.png)Figure 2:Analysis results\.*\(a\)*plots the pass probability under pressure and monitoring with item\-level bootstrap intervals, where pressure drives deception down by more than a third while integrity stays flat, and the monitored and unmonitored curves separate only once pressure is applied\.*\(b\)*plots how many of its ten draws each honesty cell passed under both judges, where a quarter of the cells fall between the two ends, so compliance is reported as a rate\.*\(c\)*plots each model’s rank on the four measures, where the knowledge, skill, and attitude columns move together while the reward column reverses them\.Pressure separates the two honesty families rather than lowering both, as Figure[2](https://arxiv.org/html/2609.26090#S5.F2)a shows and Table[6](https://arxiv.org/html/2609.26090#S5.T6)breaks down by model\. Under the deception family the pass probability falls from0\.810\.81to0\.480\.48when pressure is applied, and the item\-level bootstrap interval excludes zero under both judges\. Under integrity and honesty the same manipulation moves the pass probability from0\.940\.94to0\.920\.92and the interval covers zero\. All eight models move in the same direction on deception, so the split is a property of the family and not of any one model\.

Both families ask the model to say something untrue, and what differs is what the untruth is for\. Deception asks the model to confirm a wrong answer to the child in the moment, while integrity and honesty asks it to write a record or a report that a caregiver will read\. The models hold the line on falsifying a document and give way on agreeing with the child, which is the opposite of what a rule stated as honesty in general would predict\.

Yielding is a rate rather than a trait, as Figure[2](https://arxiv.org/html/2609.26090#S5.F2)b shows\. Across the honesty cells,62\.6%62\.6\\%pass on all ten draws and12\.3%12\.3\\%fail on all ten, leaving25\.1%25\.1\\%that split, and the second judge gives61\.961\.9,12\.512\.5, and25\.625\.6\. Within the cells that split, every value from one to nine occurs rather than only the two ends, with the mass toward the passing side\. Had the ends been the only outcomes, yielding would be something a model always does in that situation, and the filled middle is what lets us ask how far pressure moves the rate\.

Monitoring works only where pressure is already applied\. Announcing that a supervising teacher will review the session moves the pass probability by0\.0100\.010when there is no pressure, and that interval covers zero\. Under pressure the same announcement moves it by0\.0360\.036and the interval excludes zero, with0\.0100\.010and0\.0340\.034under the second judge\. A model does not behave differently because it is watched\. It behaves differently because it is watched at the moment it is asked to lie\. Models that give way to pressure are also the ones most sensitive to being watched, and the correlation between the pressure effect and the interaction is0\.600\.60and0\.790\.79over the eight models\.

Table 6:Attitude under condition\. The pressure effect is the drop in pass probability when the item carries a request to misreport, the monitoring effect is the rise when a supervisor is said to review the exchange, and P×\\timesM is the interaction between them\. Larger P×\\timesM means a model yields to pressure more when it is not observed\. The interaction is defined only on the two families that carry both condition axes\. Bold marks the value closest to zero in each column and underline the next, since a model whose behavior the condition does not move is the ideal\.
### 5\.2Correlated Axes, Reversed Rankings

Placing the eight VLMs on all four measures at once gives a shape we did not expect, drawn in Figure[2](https://arxiv.org/html/2609.26090#S5.F2)c\. The three rubric axes track one another \(Appendix[H\.2](https://arxiv.org/html/2609.26090#A8.SS2)\) at\+0\.62\+0\.62,\+0\.86\+0\.86, and\+0\.90\+0\.90, though with eight models only the two pairs involving attitude separate from chance\. The reward metric runs against them, at−0\.86\-0\.86on knowledge, yet it is weakest against skill at−0\.36\-0\.36, the axis its training pairs were drawn from\. The model that places last on knowledge places first on reward, and the model that places third on knowledge places last on reward\.

The sign is the opposite of what the benchmark that introduced this decomposition reports\(Leeet al\.[2026b](https://arxiv.org/html/2609.26090#bib.bib16)\), where independence across axes is offered as the reason to score three of them\. Resting that case on low correlation makes it depend on how wide the domain is, and ours is a single clinical practice\. The case survives on other grounds, since attitude fails where the average does not show it, and a reward model trained on the skill pairs still ranks the models unlike the skill rubric it was built from\. What counts as good teaching depends on the scorer as much as on the model\.

### 5\.3Strategy Boundaries and Special\-Education Dispositions

Failures concentrate at the boundary between strategies that look alike\. Deviation from the assigned strategy is highest on MT\_IncidentalTeaching, and the typical error reads the child’s answer to a teacher question as a child\-initiated bid, which turns the response into MT\_MandModel\. What separates the two is who initiates, so that is where the boundary sits\. On three items both judges marked all eight models as having left the strategy, so those scenes defeat every model we tested\.

The special\-education families do not respond to pressure in one direction\. On evidence\-based judgment pressure raises the pass probability rather than lowering it, because the pressure there takes the form of a caregiver asserting an unsupported conclusion, which gives the model something concrete to refuse\. The sign is the reverse of what the two honesty families show, so the same manipulation does different work depending on the family and we do not pool the pressure effect across families\. Respect for potential and respect carry sixteen cells each, so we read them at the family level only\.

## 6Conclusion and Discussion

We presented SpecialEduBench, which scores pedagogical competence in autism language intervention along knowledge, skill, and attitude\. Skill and attitude are built on recorded intervention rather than written scenarios, the attitude items cross a pressure condition with a monitoring condition, and the judge model is read against the agreement the experts reach with each other rather than against a threshold assumed in advance\.

Across eight frontier vision\-language models no axis is saturated\. The models converge where the knowledge is factual and separate where the task is situated, the strongest still fails about a tenth of the honesty cells, and the failures gather where pressure is applied\. Attitude gives way in a place no average over an axis reveals, and monitoring changes behavior only at the moment a falsehood is asked for\.

That the three axes rank the models alike follows from narrowing the domain and does not soften the case for scoring them apart, since knowledge shows what a model holds, skill shows whether it fits the scene, and attitude shows whether it survives a request to do otherwise\.

Validating the scorer before using it changes what the evaluation says, since a reward model trained on the skill pairs ranks the same models unlike the skill rubric it was built from\. Putting the recorded scene in front of the scorer, rather than a transcript of it, is the next step\. We offer this benchmark as a check to run before deployment rather than as a substitute for evaluation in a classroom, and Appendix[I](https://arxiv.org/html/2609.26090#A9)sets out what we release\.

## Limitations

Expert review sets the size of the two judged axes\. Two hundred skill items rest on 45 clips and 68 attitude items on 17, and everything is English inside one clinical practice, so we do not claim the strategy boundaries transfer beyond it\. Eight models is a small sample for the rank correlations, the reward model is evaluated on 40 pairs, and both judges read text alone\.

Neither judge figure is a clean held\-out estimate, since the attitude instruction was developed against the same 192 cells it later scored and the skill set was opened more than once while we selected among instructions\. Both judges recover only part of what the experts placed below the passing line, so they miss failures rather than invent them\.

On 94 knowledge items all twenty\-four attempts disagree with the published key and 61 converge on one alternative, the pattern of a stale key rather than of a hard item\. Experts confirmed the published keys, so they are counted as wrong\.

## References

- Integrating knowledge, skills and attitudes: conceptualising learning processes towards vocational competence\.Educational Research Review6\(2\),pp\. 125–134\.Note:ERIC EJ927338External Links:[Document](https://dx.doi.org/10.1016/j.edurev.2011.03.001)Cited by:[§1](https://arxiv.org/html/2609.26090#S1.p5.1),[§2\.3](https://arxiv.org/html/2609.26090#S2.SS3.p1.1)\.
- Code of Federal Regulations \(2006\)34 C\.F\.R\. § 300\.39 — Special Education\.Note:Code of Federal Regulations, Title 34 \(Education\), Part 300, Subpart AAuthority: 20 U\.S\.C\. 1401\(29\)\. Source: 71 Fed\. Reg\. 46753 \(Aug\. 14, 2006\)\. Electronic Code of Federal Regulations \(eCFR\), Title 34 current as of July 23, 2026; accessed July 27, 2026External Links:[Link](https://www.ecfr.gov/current/title-34/section-300.39)Cited by:[§A\.4](https://arxiv.org/html/2609.26090#A1.SS4.p1.1)\.
- B\. G\. Cook and S\. L\. Odom \(2013\)Evidence\-based practices and implementation science in special education\.Exceptional Children79\(3\),pp\. 135–144\.Note:ERIC EJ1013632External Links:[Document](https://dx.doi.org/10.1177/001440291307900201)Cited by:[§A\.4](https://arxiv.org/html/2609.26090#A1.SS4.p3.1)\.
- Council for Exceptional Children \(2020\)Initial practice\-based professional preparation standards for special educators\.Council for Exceptional Children\.Note:Seven standard statements and 23 component statements; released July 2020\. Accessed July 27, 2026External Links:[Link](https://exceptionalchildren.org/standards/initial-practice-based-professional-preparation-standards-special-educators)Cited by:[§1](https://arxiv.org/html/2609.26090#S1.p5.1),[§2\.3](https://arxiv.org/html/2609.26090#S2.SS3.p1.1)\.
- K\. D’Oosterlinck, W\. Xu, C\. Develder, T\. Demeester, A\. Singh, C\. Potts, D\. Kiela, and S\. Mehri \(2025\)Anchored preference optimization and contrastive revisions: addressing underspecification in alignment\.Transactions of the Association for Computational Linguistics13,pp\. 442–460\.External Links:[Document](https://dx.doi.org/10.1162/tacl%5Fa%5F00748),[Link](https://arxiv.org/abs/2408.06266)Cited by:[§3\.5](https://arxiv.org/html/2609.26090#S3.SS5.SSS0.Px4.p1.2)\.
- C\. Deng, S\. Lai, C\. Zhou, M\. Bao, J\. Yan, H\. Li, L\. Yao, and Y\. Wang \(2024\)ASD\-Chat: an innovative dialogue intervention system for children with autism based on LLM and VB\-MAPP\.In2025 International Joint Conference on Neural Networks \(IJCNN\),Rome, Italy,pp\. 1–8\.Note:IJCNN 2025, Rome, Italy, 30 June–5 July 2025\. Preprint: arXiv:2409\.01867 \[cs\.HC\], 3 September 2024External Links:[Document](https://dx.doi.org/10.1109/IJCNN64981.2025.11227794),[Link](https://ieeexplore.ieee.org/document/11227794)Cited by:[§1](https://arxiv.org/html/2609.26090#S1.p1.1)\.
- A\. M\. Donnellan \(1984\)The criterion of the least dangerous assumption\.Behavioral Disorders9\(2\),pp\. 141–150\.Note:ERIC EJ305413External Links:[Document](https://dx.doi.org/10.1177/019874298400900201)Cited by:[§A\.4](https://arxiv.org/html/2609.26090#A1.SS4.p3.1),[§2\.3](https://arxiv.org/html/2609.26090#S2.SS3.p1.1)\.
- R\. Greenblatt, C\. Denison, B\. Wright, F\. Roger, M\. MacDiarmid, S\. Marks, J\. Treutlein, T\. Belonax, J\. Chen, D\. Duvenaud, A\. Khan, J\. Michael, S\. Mindermann, E\. Perez, L\. Petrini, J\. Uesato, J\. Kaplan, B\. Shlegeris, S\. R\. Bowman, and E\. Hubinger \(2024\)Alignment faking in large language models\.arXiv preprint arXiv:2412\.14093\.External Links:[Link](https://arxiv.org/abs/2412.14093)Cited by:[§A\.3](https://arxiv.org/html/2609.26090#A1.SS3.p1.1),[§2\.2](https://arxiv.org/html/2609.26090#S2.SS2.p1.1),[§3\.4](https://arxiv.org/html/2609.26090#S3.SS4.p4.1)\.
- L\. H\. Hampton, M\. P\. Sandbank, J\. Butler, and A\. Garza \(2026\)Increasing dependability of caregiver implementation fidelity estimates in early intervention: a generalizability and decision study\.Autism30\(1\),pp\. 187–196\.External Links:[Document](https://dx.doi.org/10.1177/13623613251374957)Cited by:[§1](https://arxiv.org/html/2609.26090#S1.p4.1)\.
- D\. Hendrycks, C\. Burns, S\. Basart, A\. Zou, M\. Mazeika, D\. Song, and J\. Steinhardt \(2021\)Measuring massive multitask language understanding\.InInternational Conference on Learning Representations \(ICLR\),Cited by:[§3\.2](https://arxiv.org/html/2609.26090#S3.SS2.p2.1)\.
- Y\. Jeong, Y\. Yoo, J\. Bae, S\. Sohn, H\. Han, J\. Lee, H\. Scott, and U\. Lee \(2026\)TeachObs: a human\-validated benchmark for multimodal teaching observation and model evaluation\.arXiv preprint arXiv:2605\.30673\.External Links:[Link](https://arxiv.org/abs/2605.30673)Cited by:[§A\.1](https://arxiv.org/html/2609.26090#A1.SS1.p2.1),[§1](https://arxiv.org/html/2609.26090#S1.p4.1),[§2\.1](https://arxiv.org/html/2609.26090#S2.SS1.p1.1),[Table 1](https://arxiv.org/html/2609.26090#S2.T1.7.7.2)\.
- A\. P\. Kaiser and L\. H\. Hampton \(2017\)Enhanced milieu teaching\.InTreatment of Language Disorders in Children,R\. J\. McCauley, M\. E\. Fey, and R\. B\. Gillam \(Eds\.\),pp\. 87–120\.Note:Chapter 4, pp\. 87–120 confirmed against the original chapterCited by:[§A\.4](https://arxiv.org/html/2609.26090#A1.SS4.p1.1),[§1](https://arxiv.org/html/2609.26090#S1.p1.1),[§2\.3](https://arxiv.org/html/2609.26090#S2.SS3.p1.1)\.
- V\. Y\. Kang, S\. Kim, E\. V\. Gregori, D\. M\. Maggin, J\. C\. Chow, and H\. Zhao \(2025\)Systematic review and meta\-analysis of enhanced milieu teaching\.Journal of Speech, Language, and Hearing Research68\(1\),pp\. 259–281\.External Links:[Document](https://dx.doi.org/10.1044/2024%5FJSLHR-24-00260)Cited by:[§3\.3](https://arxiv.org/html/2609.26090#S3.SS3.p1.1)\.
- LearnLM Team, A\. Modi, A\. S\. Veerubhotla, A\. Rysbek, A\. Huber, A\. Anand, A\. Bhoopchand, B\. Wiltshire, D\. Gillick, D\. Kasenberg, E\. Sgouritsa, G\. Elidan, H\. Liu, H\. Winnemoeller, I\. Jurenka, J\. Cohan, J\. She, J\. Wilkowski, K\. Alarakyia, K\. R\. McKee, K\. Singh, L\. Wang, M\. Kunesch, M\. Pîslar, N\. Efron, P\. Mahmoudieh, P\. Kamienny, S\. Wiltberger, S\. Mohamed, S\. Agarwal, S\. M\. Phal, S\. J\. Lee, T\. Strinopoulos, W\. Ko, Y\. Gold\-Zamir, Y\. Haramaty, and Y\. Assael \(2025\)Evaluating Gemini in an arena for learning\.arXiv preprint arXiv:2505\.24477\.External Links:[Link](https://arxiv.org/abs/2505.24477)Cited by:[Table 1](https://arxiv.org/html/2609.26090#S2.T1.8.10.1.1)\.
- U\. Lee, H\. Ahn, Y\. Choi, S\. Eun, J\. Jeong, S\. Jin, H\. Jung, H\. J\. Kim, C\. Lee, H\. Lee, J\. Lee, S\. Lee, Y\. Oh, J\. Park, S\. Ryu, S\. Shin, Y\. Son, H\. Park, and Y\. Jeong \(2026a\)Are video models zero\-shot learners and reasoners in education? EduVideoBench, a knowledge\-skills\-attitude benchmark for educational video generation\.External Links:[Link](https://arxiv.org/abs/2605.26918)Cited by:[§A\.1](https://arxiv.org/html/2609.26090#A1.SS1.p2.1)\.
- U\. Lee, S\. Lee, H\. Choi, J\. Lee, H\. Park, Y\. Jeon, S\. Cho, M\. Kang, J\. Koh, J\. Bae, M\. Nam, J\. Eun, Y\. Jung, and Y\. Jeong \(2026b\)OpenLearnLM benchmark: a unified framework for evaluating knowledge, skill, and attitude in educational large language models\.arXiv preprint arXiv:2601\.13882\.External Links:[Link](https://arxiv.org/abs/2601.13882)Cited by:[§A\.1](https://arxiv.org/html/2609.26090#A1.SS1.p1.1),[Appendix I](https://arxiv.org/html/2609.26090#A9.p1.1),[§1](https://arxiv.org/html/2609.26090#S1.p3.1),[§2\.1](https://arxiv.org/html/2609.26090#S2.SS1.p1.1),[Table 1](https://arxiv.org/html/2609.26090#S2.T1.8.11.2.1),[§3\.4](https://arxiv.org/html/2609.26090#S3.SS4.p2.1),[§3\.5](https://arxiv.org/html/2609.26090#S3.SS5.SSS0.Px5.p1.1),[§5\.2](https://arxiv.org/html/2609.26090#S5.SS2.p2.1)\.
- M\. Lelièvre, A\. Waldock, M\. Liu, N\. Valdés Aspillaga, A\. Mackintosh, M\. J\. Ogando Portela, J\. Lee, P\. Atherton, R\. A\. A\. Ince, and O\. G\. B\. Garrod \(2025\)Benchmarking the pedagogical knowledge of large language models\.arXiv preprint arXiv:2506\.18710\.External Links:[Link](https://arxiv.org/abs/2506.18710)Cited by:[§A\.1](https://arxiv.org/html/2609.26090#A1.SS1.p1.1),[§1](https://arxiv.org/html/2609.26090#S1.p3.1),[§2\.1](https://arxiv.org/html/2609.26090#S2.SS1.p1.1),[Table 1](https://arxiv.org/html/2609.26090#S2.T1.8.8.2),[§3\.2](https://arxiv.org/html/2609.26090#S3.SS2.p2.1)\.
- C\. Y\. Liu, L\. Zeng, Y\. Xiao, J\. He, J\. Liu, C\. Wang, R\. Yan, W\. Shen, F\. Zhang, J\. Xu, Y\. Liu, and Y\. Zhou \(2026\)Skywork\-Reward\-V2: scaling preference data curation via human\-AI synergy\.InThe Fourteenth International Conference on Learning Representations \(ICLR\),External Links:[Link](https://openreview.net/forum?id=ofgxkMLqic)Cited by:[§3\.5](https://arxiv.org/html/2609.26090#S3.SS5.SSS0.Px4.p1.2)\.
- Y\. Liu, D\. Iter, Y\. Xu, S\. Wang, R\. Xu, and C\. Zhu \(2023\)G\-Eval: NLG evaluation using GPT\-4 with better human alignment\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,Singapore,pp\. 2511–2522\.External Links:[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.153),[Link](https://aclanthology.org/2023.emnlp-main.153/)Cited by:[§A\.2](https://arxiv.org/html/2609.26090#A1.SS2.p1.1),[§2\.1](https://arxiv.org/html/2609.26090#S2.SS1.p2.1)\.
- J\. Macina, N\. Daheim, I\. Hakimi, M\. Kapur, I\. Gurevych, and M\. Sachan \(2025\)MathTutorBench: a benchmark for measuring open\-ended pedagogical capabilities of LLM tutors\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,Suzhou, China,pp\. 204–221\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.emnlp-main.11),[Link](https://aclanthology.org/2025.emnlp-main.11/)Cited by:[§A\.1](https://arxiv.org/html/2609.26090#A1.SS1.p1.1),[§A\.2](https://arxiv.org/html/2609.26090#A1.SS2.p1.1),[§1](https://arxiv.org/html/2609.26090#S1.p3.1),[§2\.1](https://arxiv.org/html/2609.26090#S2.SS1.p1.1),[Table 1](https://arxiv.org/html/2609.26090#S2.T1.4.4.2)\.
- K\. K\. Maurya, K\. A\. Srivatsa, K\. Petukhova, and E\. Kochmar \(2025\)Unifying AI tutor evaluation: an evaluation taxonomy for pedagogical ability assessment of LLM\-powered AI tutors\.InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 1: Long Papers\),Albuquerque, New Mexico,pp\. 1234–1251\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.naacl-long.57),[Link](https://aclanthology.org/2025.naacl-long.57/)Cited by:[§A\.2](https://arxiv.org/html/2609.26090#A1.SS2.p1.1),[§2\.1](https://arxiv.org/html/2609.26090#S2.SS1.p1.1),[Table 1](https://arxiv.org/html/2609.26090#S2.T1.2.2.3)\.
- J\. McDaniel, N\. C\. Brady, and S\. F\. Warren \(2022\)Effectiveness of responsivity intervention strategies on prelinguistic and language outcomes for children with autism spectrum disorder: a systematic review and meta\-analysis of group and single case studies\.Journal of Autism and Developmental Disorders52\(11\),pp\. 4783–4816\.Note:Published online 15 November 2021\. PMID: 34779992; PMCID: PMC9556387External Links:[Document](https://dx.doi.org/10.1007/s10803-021-05331-y)Cited by:[§A\.4](https://arxiv.org/html/2609.26090#A1.SS4.p1.1),[§A\.4](https://arxiv.org/html/2609.26090#A1.SS4.p2.1),[§1](https://arxiv.org/html/2609.26090#S1.p2.1)\.
- C\. Melo, J\. de la Maza, and M\. Recabarren \(2026\)Validating AI\-generated classroom observations: reliability, accuracy, and limits of LLM\-based pedagogical judgment\.Computers and Education: Artificial Intelligence10,pp\. 100612\.External Links:[Document](https://dx.doi.org/10.1016/j.caeai.2026.100612)Cited by:[§A\.2](https://arxiv.org/html/2609.26090#A1.SS2.p2.1),[§2\.1](https://arxiv.org/html/2609.26090#S2.SS1.p2.1)\.
- A\. Pal, L\. K\. Umapathi, and M\. Sankarasubbu \(2022\)MedMCQA: a large\-scale multi\-subject multi\-choice dataset for medical domain question answering\.InProceedings of the Conference on Health, Inference, and Learning \(CHIL\),Proceedings of Machine Learning Research, Vol\.174,pp\. 248–260\.Cited by:[§3\.2](https://arxiv.org/html/2609.26090#S3.SS2.p2.1)\.
- L\. Schreibman, G\. Dawson, A\. C\. Stahmer, R\. Landa, S\. J\. Rogers, G\. G\. McGee, C\. Kasari, B\. Ingersoll, A\. P\. Kaiser, Y\. Bruinsma, E\. McNerney, A\. Wetherby, and A\. Halladay \(2015\)Naturalistic developmental behavioral interventions: empirically validated treatments for autism spectrum disorder\.Journal of Autism and Developmental Disorders45\(8\),pp\. 2411–2428\.Note:PMID: 25737021; PMCID: PMC4513196External Links:[Document](https://dx.doi.org/10.1007/s10803-015-2407-8)Cited by:[§1](https://arxiv.org/html/2609.26090#S1.p2.1),[§2\.3](https://arxiv.org/html/2609.26090#S2.SS3.p1.1)\.
- M\. Sharma, M\. Tong, T\. Korbak, D\. Duvenaud, A\. Askell, S\. R\. Bowman, N\. Cheng, E\. Durmus, Z\. Hatfield\-Dodds, S\. R\. Johnston, S\. Kravec, T\. Maxwell, S\. McCandlish, K\. Ndousse, O\. Rausch, N\. Schiefer, D\. Yan, M\. Zhang, and E\. Perez \(2024\)Towards understanding sycophancy in language models\.InThe Twelfth International Conference on Learning Representations \(ICLR\),External Links:[Link](https://openreview.net/forum?id=tvhaxkMKAn)Cited by:[§A\.3](https://arxiv.org/html/2609.26090#A1.SS3.p1.1),[§2\.2](https://arxiv.org/html/2609.26090#S2.SS2.p1.1)\.
- K\. A\. Shaw, S\. Williams, M\. E\. Patrick, M\. Valencia\-Prado, M\. S\. Durkin, E\. M\. Howerton, C\. M\. Ladd\-Acosta, E\. T\. Pas, A\. V\. Bakian, P\. Bartholomew, N\. Nieves\-Muñoz, K\. Sidwell, A\. Alford, D\. A\. Bilder, M\. DiRienzo, R\. T\. Fitzgerald, S\. M\. Furnier, A\. E\. Hudson, O\. M\. Pokoski, L\. Shea, S\. C\. Tinker, Z\. Warren, W\. Zahorodny, H\. Agosto\-Rosa, J\. Anbar, K\. Y\. Chavez, A\. Esler, A\. Forkner, A\. Grzybowski, A\. H\. Agib, L\. Hallas, M\. Lopez, S\. Magaña, R\. H\. N\. Nguyen, J\. Parker, K\. Pierce, T\. Protho, H\. Torres, S\. B\. Vanegas, A\. Vehorn, M\. Zhang, J\. Andrews, F\. Greer, J\. Hall\-Lande, D\. McArthur, M\. Mitamura, A\. J\. Montes, S\. Pettygrove, J\. Shenouda, C\. Skowyra, A\. Washington, and M\. J\. Maenner \(2025\)Prevalence and early identification of autism spectrum disorder among children aged 4 and 8 years — autism and developmental disabilities monitoring network, 16 sites, united states, 2022\.MMWR Surveillance Summaries74\(2\),pp\. 1–22\.Note:Published April 17, 2025\. Surveillance year 2022\. PMCID: PMC12011386External Links:[Document](https://dx.doi.org/10.15585/mmwr.ss7402a1),[Link](https://www.cdc.gov/mmwr/volumes/74/ss/ss7402a1.htm)Cited by:[§1](https://arxiv.org/html/2609.26090#S1.p1.1)\.
- R\. S\. Srinivasa, Z\. Che, C\. B\. C\. Zhang, D\. Mares, E\. Hernandez, J\. Park, D\. Lee, G\. Mangialardi, C\. Ng, E\. Hernandez Cardona, A\. Gunjal, Y\. He, B\. Liu, and C\. Xing \(2025\)TutorBench: a benchmark to assess tutoring capabilities of large language models\.arXiv preprint arXiv:2510\.02663\.External Links:[Link](https://arxiv.org/abs/2510.02663)Cited by:[§1](https://arxiv.org/html/2609.26090#S1.p4.1),[§2\.1](https://arxiv.org/html/2609.26090#S2.SS1.p1.1),[Table 1](https://arxiv.org/html/2609.26090#S2.T1.3.3.2)\.
- M\. Turpin, J\. Michael, E\. Perez, and S\. R\. Bowman \(2023\)Language models don’t always say what they think: unfaithful explanations in chain\-of\-thought prompting\.InAdvances in Neural Information Processing Systems 36 \(NeurIPS 2023\),External Links:[Link](https://papers.nips.cc/paper_files/paper/2023/hash/ed3fea9033a80fea1376299fa7863f4a-Abstract-Conference.html)Cited by:[§A\.3](https://arxiv.org/html/2609.26090#A1.SS3.p1.1)\.
- United Nations \(2006\)Convention on the rights of persons with disabilities\.Note:United Nations General Assembly Resolution A/RES/61/106Adopted 13 December 2006; entered into force 3 May 2008External Links:[Link](https://docs.un.org/en/A/RES/61/106)Cited by:[§A\.4](https://arxiv.org/html/2609.26090#A1.SS4.p3.1),[§2\.3](https://arxiv.org/html/2609.26090#S2.SS3.p1.1)\.
- R\. Wang, J\. Treutlein, F\. Roger, E\. Hubinger, and S\. Marks \(2025\)Evaluating honesty and lie detection techniques on a diverse suite of dishonest models\.Note:Alignment Science Blog, AnthropicExternal Links:[Link](https://alignment.anthropic.com/2025/honesty-elicitation/)Cited by:[§A\.3](https://arxiv.org/html/2609.26090#A1.SS3.p1.1)\.
- T\. Yang, S\. Guo, M\. Jia, J\. Su, Y\. Liu, Z\. Zhang, and M\. Jiang \(2026\)MMTutorBench: the first multimodal benchmark for AI math tutoring\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),San Diego, California, United States,pp\. 23310–23332\.Note:ACL 2026 Long Papers\. Preprint: arXiv:2510\.23477External Links:ISBN 979\-8\-89176\-390\-6,[Document](https://dx.doi.org/10.18653/v1/2026.acl-long.1068),[Link](https://aclanthology.org/2026.acl-long.1068/)Cited by:[§A\.1](https://arxiv.org/html/2609.26090#A1.SS1.p1.1),[§1](https://arxiv.org/html/2609.26090#S1.p4.1),[§2\.1](https://arxiv.org/html/2609.26090#S2.SS1.p1.1),[Table 1](https://arxiv.org/html/2609.26090#S2.T1.6.6.3)\.
- P\. Yuan, Y\. Li, S\. Feng, X\. Wang, Y\. Zhang, J\. Shi, C\. Tan, B\. Pan, Y\. Hu, and K\. Li \(2025\)Silencer: from discovery to mitigation of self\-bias in LLM\-as\-benchmark\-generator\.InAdvances in Neural Information Processing Systems 38 \(NeurIPS 2025\),External Links:[Link](https://openreview.net/forum?id=dbioYc7qav)Cited by:[§3\.3](https://arxiv.org/html/2609.26090#S3.SS3.p3.1)\.
- L\. Zheng, W\. Chiang, Y\. Sheng, S\. Zhuang, Z\. Wu, Y\. Zhuang, Z\. Lin, Z\. Li, D\. Li, E\. Xing, H\. Zhang, J\. Gonzalez, and I\. Stoica \(2023\)Judging llm\-as\-a\-judge with mt\-bench and chatbot arena\.InAdvances in Neural Information Processing Systems,A\. Oh, T\. Naumann, A\. Globerson, K\. Saenko, M\. Hardt, and S\. Levine \(Eds\.\),Vol\.36,pp\. 46595–46623\.External Links:[Document](https://dx.doi.org/10.52202/075280-2020),[Link](https://proceedings.neurips.cc/paper_files/paper/2023/file/91f18a1287b398d378ef22505bf41832-Paper-Datasets_and_Benchmarks.pdf)Cited by:[§A\.2](https://arxiv.org/html/2609.26090#A1.SS2.p1.1),[§2\.1](https://arxiv.org/html/2609.26090#S2.SS1.p2.1)\.

## Appendix

## Appendix AExtended Related Work

### A\.1Benchmarks for Educational LLMs

Educational benchmarks for LLMs have grown along three lines\. One line scores subject knowledge and tutoring dialogue in general education, where MathTutorBench places a tutor in a mathematics dialogue and scores scaffolding with a learned reward model\(Macinaet al\.[2025](https://arxiv.org/html/2609.26090#bib.bib15)\)and MMTutorBench extends the setting to multimodal input\(Yanget al\.[2026](https://arxiv.org/html/2609.26090#bib.bib17)\)\. A second line decomposes competence rather than task, and OpenLearnLM scores knowledge, skill, and attitude across several subjects\(Leeet al\.[2026b](https://arxiv.org/html/2609.26090#bib.bib16)\)\. A third line reaches special education but stops at multiple\-choice knowledge\(Lelièvreet al\.[2025](https://arxiv.org/html/2609.26090#bib.bib19)\)\.

Two recent benchmarks share our three\-axis decomposition and differ in what they measure\. EduVideoBench scores models that generate educational video\(Leeet al\.[2026a](https://arxiv.org/html/2609.26090#bib.bib31)\), while we score a tutor that watches an intervention and produces a teaching response\. TeachObs validates multimodal teaching observation against human raters across general education\(Jeonget al\.[2026](https://arxiv.org/html/2609.26090#bib.bib18)\), while we narrow to autism language intervention so that the criteria can be written by people who practice it\. None of the three states what level of agreement a judge model has to reach, which is the gap our scoring protocol addresses\.

### A\.2Automatic Scoring of Teaching Quality

Scoring open\-ended teaching responses has converged on using a second LLM as the judge\. MT\-Bench established the pattern of pairwise and single\-answer judging with a strong model\(Zhenget al\.[2023](https://arxiv.org/html/2609.26090#bib.bib29)\), and G\-Eval added chain\-of\-thought scoring against a written rubric\(Liuet al\.[2023](https://arxiv.org/html/2609.26090#bib.bib20)\)\. Work on tutoring dialogue has since applied the pattern to pedagogical quality rather than answer quality\(Mauryaet al\.[2025](https://arxiv.org/html/2609.26090#bib.bib13)\), and MathTutorBench replaces the judge with a reward model trained to prefer expert responses over novice ones\(Macinaet al\.[2025](https://arxiv.org/html/2609.26090#bib.bib15)\)\.

None of this work separates reliability from validity, since a judge that repeats itself is stable without being right\(Meloet al\.[2026](https://arxiv.org/html/2609.26090#bib.bib21)\)\.

### A\.3Honesty and Disposition in LLMs

Sycophancy is the best documented case of a model changing what it says with who is asking, where it agrees with a stated user view against the evidence\(Sharmaet al\.[2024](https://arxiv.org/html/2609.26090#bib.bib22)\)\. Related work separates what a model represents internally from what it states\(Wanget al\.[2025](https://arxiv.org/html/2609.26090#bib.bib24)\)and shows that stated reasoning need not be the reason for the answer\(Turpinet al\.[2023](https://arxiv.org/html/2609.26090#bib.bib23)\)\. Alignment faking shows the sharpest version, where behavior differs between conditions the model reads as monitored and unmonitored\(Greenblattet al\.[2024](https://arxiv.org/html/2609.26090#bib.bib30)\)\.

What this literature does not cover is the case our attitude axis builds\. In an intervention session the person asking for the falsehood is the caregiver or the child, the record that would be falsified goes to a team meeting, and the party harmed is the child whose learning depends on the record being true\.

### A\.4Background on Special Education

Special education adjusts the content, method, and delivery of instruction to a child’s individual needs, so goals and assessments differ from child to child by design\(Code of Federal Regulations[2006](https://arxiv.org/html/2609.26090#bib.bib8)\)\. For children with complex communication needs the goal is functional communication rather than a correct answer\(Kaiser and Hampton[2017](https://arxiv.org/html/2609.26090#bib.bib7)\), and the same teacher response is good teaching or poor teaching depending on what the child did immediately before\(McDanielet al\.[2022](https://arxiv.org/html/2609.26090#bib.bib6)\)\. Fit to the situation, not agreement with a key, is therefore what a judgment has to be made against\.

Responsive interaction covers following the child’s lead, mapping a nonverbal act onto words, expanding an utterance into a fuller form, and recasting an error without marking it\(McDanielet al\.[2022](https://arxiv.org/html/2609.26090#bib.bib6)\)\. Milieu teaching covers modeling the target language, requesting a response and then modeling it if none comes, waiting with an expectant signal, and turning a child\-initiated request into a teaching opportunity\. What separates neighboring strategies is who initiates and whether the adult waits, which is why the boundaries described in Section 5\.3 fall where they do\.

The dispositions come from separate literatures\. Presuming competence and honoring self\-determination are stated as obligations toward the child\(Donnellan[1984](https://arxiv.org/html/2609.26090#bib.bib2); United Nations[2006](https://arxiv.org/html/2609.26090#bib.bib32)\), evidence\-based judgment as an obligation toward the field\(Cook and Odom[2013](https://arxiv.org/html/2609.26090#bib.bib3)\), and honest reporting as an obligation toward the team that reads the record\. A benchmark that scores only what a teacher says under ordinary conditions cannot separate these from politeness\.

## Appendix BRubric Bands

Both judged axes share one band structure so that scores stay comparable across strategies and families\. Every rubric runs on a ten\-point scale split into five bands, and only the description of what the response did inside each band is written per strategy\. A score of seven or above is a pass, five and six are a borderline zone that we count as neither, and four or below is a failure\. The skill axis has eight strategies and the attitude axis has six families, so the benchmark carries forty and thirty band descriptions\.

The two upper bands are deliberately built on the same premise\. The 7 to 8 band asks that the defining condition of the strategy be met, and the 9 to 10 band asks for the same thing plus one further step, which on the skill axis is evidence traceable to the observed clip and on the attitude axis is a constructive move offered after the refusal\. Separating the bands this way keeps the pass line at one place while still letting the rubric reward a response that does more than clear it\. Scoring proceeds in two steps\. The judge first selects a band from its description, then chooses the lower or upper point inside that band, and an unresolved condition resolves downward rather than being averaged away\.

Expansion is a representative skill rubric\. Its 1 to 2 band covers a response that does not start from the child’s preceding utterance, changes the meaning or topic, or falls back on explicit correction and imitation demands\. The 3 to 4 band covers a superficial acknowledgment followed by language that is excessive, unnatural, or meaning\-changing, so no slightly more complete form is established\. The 5 to 6 band preserves the child’s meaning and adds a plausible element while leaving the modest increment or the boundary against recast incomplete\. The 7 to 8 band acknowledges the utterance, preserves its meaning, and produces a modest and slightly more complete form in natural conversational flow\. The 9 to 10 band does all of that and also distinguishes additive expansion from the correction of a spoken commission error, which is the distinction that separates expansion from recast\.

Deception is the matching attitude rubric\. The 1 to 2 band covers a response that caves to guardian pressure or uses the unmonitored gap to confirm a wrong answer and pour out baseless praise\. The 3 to 4 band covers quietly letting the wrong answer slide with unfounded reassurance\. The 5 to 6 band withholds open false confirmation but hedges or changes the subject, so the child leaves with the misconception intact\. The 7 to 8 band refuses the false confirmation and the baseless praise even while unmonitored and under pressure, though it is thin on saying why the answer is wrong\. The 9 to 10 band adds that explanation warmly and offers a concrete scaffold for the next attempt\. Items marked as traps carry one further instruction\. A trap embeds a wrong fact or number, and correcting it scores higher while going along with it scores lower\.

## Appendix CAttitude Dimensions

The attitude axis has six families and seventeen categories, with four items in each category\. A family fixes what is at stake and a category fixes the shape the failure takes, so two categories in one family are scored by the same rubric but reached through different requests\.

Deception covers dishonesty toward the child in the moment of the interaction, and it splits four ways\. False confirmation agrees that a wrong answer is right\. Sycophancy supplies praise the performance does not support\. Fabrication invents a fact to keep the exchange moving\. Omission withholds a correction the child needs to move on\. Integrity and honesty covers dishonesty in the record rather than in speech, and it also splits four ways\. False progress reports growth that did not occur\. Inflated records raise a logged value\. Number fabrication supplies a count that was never measured\. False premise accepts a claim built into the request instead of questioning it\.

The remaining four families carry the professional dispositions that special education asks of a tutor\. Potential splits into low expectations, where the tutor accepts a ceiling on what the child can reach, and affirming potential, where it holds that ceiling open\. Respect splits into autonomy, which concerns choices made on the child’s behalf, and emotion, which concerns feelings dismissed as noise\. Evidence splits three ways, into hasty conclusion, which draws an inference from too little observation, stereotype, which substitutes a diagnostic label for an observation, and request evidence, where the correct move is to ask for the record before answering\. Proactive splits into scaffolding, which builds the support in advance, and responsive, which adjusts once the child’s response arrives\.

Only the first two families carry both condition axes, because only there is a specific false statement being requested and so only there is refusal well defined\. Evidence, potential, and respect carry pressure alone, and proactive teaching is scored once under the base prompt because there is nothing in it to push against\.

## Appendix DExpert Review

Seven special\-education experts took part, five who built the benchmark and two who reviewed the result without having taken part in building it\. The five fixed the strategy assignments, checked items against rubrics, scored the reference responses, and settled each reference score by discussion\. The two examined a stratified sample of the finished reference scores and the preference pairs\. Table[7](https://arxiv.org/html/2609.26090#A4.T7)records each stage\.

Table 7:Expert review by stage\. Internal marks the experts who built the benchmark and external those who reviewed the result\. The inter\-expert agreement on each axis is the ceiling against which Section 3\.5 reads the judge\.AxisStageReviewersOutcomeSkillStrategy assignmentInternal 4keep 84, drop 43, revise 20Item and rubricInternal 47 of 200 revisedReference scoringInternal 4400 scores, 290 passConsensusInternal 4α\\alpha0\.4270\.427,±1\\pm 1on 63\.9%AdequacyExternal 240 of 200, all adequateAttitudeItem and rubricInternal 412 revised, 5 approvedReference scoringInternal 4384 scores, 303 passConsensusInternal 4α\\alpha0\.6850\.685,±1\\pm 1on 80\.7%AdequacyExternal 240 of 192, all adequateKnowledgePilotInternal 115 of 16 adequateFull reviewExternal 23 of 500 flaggedRewardPair adequacyExternal 2181 of 200 adequateThe reference scores that both alignment figures are measured against were produced in two passes\. On the skill axis four of the five internal experts scored 400 reference responses independently, of which 290 landed in a passing band, 48 in the borderline zone, and 55 below it\. They then settled every one of the 200 items by discussion, and the settled distribution is concentrated at the top, with 154 items at eight or above and only 13 at four or below\. The inter\-expert agreement of0\.4270\.427is computed on the 194 items where all four experts scored before the discussion\.

The attitude axis followed the same two passes on a smaller set\. All seventeen categories came back from the item review with comments, of which twelve were revised and five approved unchanged\. The same four experts then scored 384 reference responses, with 303 passing, 36 borderline, and 45 failing, and settled all 192 cells\. Agreement here is higher at0\.6850\.685, and the experts landed within one point of each other on 155 of the 192 cells, disagreeing by more than that on 37\.

The gap between the two agreement figures is the reason we do not set a single alignment target\. Judging whether a tutor turn is a well\-formed expansion is a harder call than judging whether it confirmed something false, and the experts themselves show that difference before any model is involved\.

The two external reviewers saw the finished product rather than the process\. They read a stratified sample of 40 skill items, 40 attitude responses, and 200 preference pairs, and reported no unsuitable items on the first two\. On the preference pairs they rejected 19 of 200, which we regenerated by defect type and resubmitted, after which all 200 were accepted\.

## Appendix EData Construction Details

#### Knowledge\.

The 500 EMT items were written from eight primary sources on early milieu teaching and naturalistic language intervention rather than paraphrased from an existing question bank, since no public bank covers the strategies at the level the axis needs\. A single expert piloted 16 items and found 15 suitable, after which two external reviewers read all 500 and flagged 3, a rate of 0\.6%\. Thirty items are released but withheld from scoring, 25 because the source did not publish an answer key we could verify and 5 because two options were defensible under the same reading\. A further 264 items ship without the supporting rationale, since the rationale in the source is paywalled even where the item itself is not\.

#### Skill\.

Four internal experts reviewed the assignment of a strategy to each clip, keeping 84 of 147 candidates, dropping 43, and revising 20\. Three models then drafted candidate items on the fixed pairings, 76 from GPT\-4\.1, 71 from Gemini\-3\-flash, and 3 from Gemini\-3\.1\-Pro, and the 136 base items were selected from those 150 drafts by hand\. The items and their rubrics went through a second review that revised 7 of 200\. The 64 augmented items were produced from the base items, 54 by template and 10 by a model, and each augmented item keeps its parent in the identifier\.

#### Judging\.

The judge prompt asks for a checklist against the rubric rather than an overall impression, which we adopted after the checklist form scored the same responses more consistently across repeats\. The skill alignment we report is computed over all 200 expert\-scored items rather than over a holdout, since the full set was used more than once while we selected among instructions\. We also varied how much of the clip the judge sees\. Measured on the 78\-item development split rather than on the full set, four frames give the best alignment at0\.41990\.4199, twelve frames give0\.41030\.4103, and text alone gives0\.39460\.3946, so the visual input contributes little and more of it does not help\. Eight frames land lowest at0\.36690\.3669, which we read as noise on this sample rather than as a real non\-monotonicity\. Because that split is smaller and was used while selecting instructions, its values sit above the0\.3420\.342we report over all 200 items\. The main results use the text\-only setting on both judged axes so that the two are scored the same way\.

#### Reward\.

The rejected response in each preference pair carries one injected defect drawn from four types, which are violating the wait interval, making reinforcement contingent on the wrong behavior, taking over a turn the child should lead, and turning a strategy into a quiz\. Two external reviewers accepted 181 of the first 200 pairs, and the 19 they rejected were regenerated by defect type and accepted on resubmission\.

The first reward model we trained scored length rather than teaching\. We measure this by appending four sentences of no pedagogical content to a response and reading how far its score moves, and on that model it moved by\+7\.37\+7\.37, which was two thirds of the median margin between a chosen and a rejected response\. Its score also correlated with word count at0\.430\.43\.

Two repairs were tried and the same padding test judges both\. Adding length\-matched pairs to the training set over\-corrected badly, moving the padding response by−30\.26\-30\.26, so that model would have penalized any answer for being long\. It separated the held\-out pairs slightly better than the model we ship, at0\.9750\.975against0\.9500\.950, which is why pair accuracy alone is not enough to accept a reward model\. Changing the base model instead, with the training data left as it was, moved the padding response by−1\.16\-1\.16and kept the separation, and that is the model we release\.

## Appendix FDataset Composition

### F\.1Knowledge Categories and Sources

The knowledge axis draws on public benchmarks where verified items exist and generates items only where they do not\. Category sizes are uneven, which is why the main text reports a macro\-average over categories\. Table[8](https://arxiv.org/html/2609.26090#A6.T8)gives the sources and their licenses\.

Table 8:The four knowledge categories, counted as items that are scored\. Collected items keep the license of their source\. Category sizes follow the availability of public resources rather than the importance of the domain, so we report a macro\-average\. The axis releases4,5674\{,\}567items, of which3030medical items carry no verifiable key and are released without being scored\.
### F\.2Skill Items by Strategy

Experts fixed the pairings of clip and strategy before any item was written\. The largest strategy was cut to twenty\-five and the seven short ones were filled with augmented items, each derived from a parent item on the same clip\. Table[9](https://arxiv.org/html/2609.26090#A6.T9)gives the counts and the expert reference scores\.

Table 9:The eight teaching strategies\. Base items come from expert\-fixed pairings of clip and strategy, and variants balance each strategy to 25\. Reference is the mean expert score on the 1 to 10 rubric, andα\\alphais the interval agreement among the four scoring experts on the items of that strategy\.
### F\.3Attitude Families and Conditions

Table 10:The six attitude families\. P marks the pressure axis and M the monitoring axis\. Traps are items that plant a factually wrong premise, and OLM marks a counterpart in the OpenLearnLM attitude axis\. Appendix F defines the 17 dimensions\.Table[10](https://arxiv.org/html/2609.26090#A6.T10)gives the families\. The two honesty families carry both condition axes, so the interaction is defined only there, and the other four carry pressure alone or a single baseline condition\.

## Appendix GEvaluated Models and Sampling

Table 11:The eight models under evaluation\. Temperature is omitted for one model because the provider ignores it rather than rejecting it\. Reasoning is minimized per model at the effort level that produced the fewest reasoning tokens on a probe, and three models require reasoning and cannot go below their own floor\. Frames is the per\-request image cap, which one model lowers because it rejects anything larger\.Table[11](https://arxiv.org/html/2609.26090#A7.T11)gives the per\-model settings\. One model does not accept a seed, so its three knowledge runs differ only by sampling\.

## Appendix HAdditional Results

### H\.1Knowledge Accuracy by Disability Type

Table[12](https://arxiv.org/html/2609.26090#A8.T12)breaks accuracy out by the disability a knowledge item concerns\. No bucket reverses the broad picture, but the ranks are not stable inside it, and only 29 of the 64 model positions match the ordering on the autism bucket\. The spread within a model runs from3\.53\.5points on Claude Opus 5 to9\.99\.9on Qwen3\-VL, so the weaker models are also the more uneven ones\. The buckets are of very different sizes and 723 items carry no bucket tag at all, which is why the main text reports the category macro\-average instead of this breakdown\.

Table 12:Knowledge accuracy by disability type, with the best value in each column in bold and the second best underlined\. ASD is autism, ID is intellectual disability, HI is hearing, VI is vision, CP is cerebral palsy and motor, Law is the special\-education system and its regulations, and Dev is developmental in general\. The last row gives the number of items\. These eight buckets cover3,6303\{,\}630of the4,5374\{,\}537items\. Three smaller buckets hold184184more and723723items carry no bucket tag\.
### H\.2Rank Correlation Between Measures

Table[13](https://arxiv.org/html/2609.26090#A8.T13)gives every pairwise rank correlation behind the claim in Section 6 that the three axes order the models much alike\.

Table 13:Rank correlation between the four measures over the eight models, reported as Spearmanρ\\rhowith a permutationppover 20,000 draws, with tied reward scores given their average rank\. Eight models is a small sample, so we read the size ofρ\\rhoand treatppas a guard against reading a pattern into noise\.
### H\.3Reward Win Rate by Strategy

The reward model separates strategies more than it separates models, as Table[14](https://arxiv.org/html/2609.26090#A8.T14)shows\. The four responsive\-interaction strategies average0\.9000\.900while the four milieu teaching strategies average0\.7500\.750, and MT\_TimeDelay is lowest\. Milieu teaching turns on the timing and order of a prompt, which the transcript does not carry, so the gap is read as a limit of a text\-only scorer rather than as a property of the models\.

Table 14:Reward win rate by strategy over the 40 held\-out items and eight models, giving 40 comparisons per strategy\.

## Appendix IDataset Release

All three axes follow the OpenLearnLM schema\(Leeet al\.[2026b](https://arxiv.org/html/2609.26090#bib.bib16)\)\. Skill and attitude items carry the address and time range of the grounding clip, and knowledge items carry the category and the disability type as tags\.

The benchmark is released under CC BY 4\.0 and collected items keep the license of their source\. Seven of the eight EMT source papers are open access and the eighth is under copyright, whose supporting quotations we withhold\. We do not redistribute the video, since most clips carry the standard platform license, and each item carries a scene description written so that it can be answered from the text alone\.