Benchmark Almanac

AD4Edu-Bench

SmolVLM2-2.2B-Instruct 0.86 leads since October 2026
Not in model cards

Models write audio descriptions for university lectures. A rubric checks accessibility rules, and the authors compare the rankings with blind and low-vision listeners' preferences.

Released
October 2026
Kind
Video captioning
Level
University
Scored
Graded by a model
Answers
Open-ended
Measures
Capability
Subject
Education
Language
English
In and out
Video, Audio → Text
Interaction
Single turn
Items from
Exams and contests
Refresh
One fixed set
Run by
Anyone
Last model added
October 2025
Scores refreshed
November 20th, 2026
7
Models scored in its paper

SmolVLM2-2.2B-Instruct leads with 0.86 among models released by October 2025

Best score per model from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. The headline is mechanical AD4Edu-Eval compliance, the first score the paper uses to rank compliance with its accessibility standard. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.

The newest model with a score here came out in October 2025, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 1 SmolVLM2-2.2B-Instruct SFT, text-only 0.86
  2. 2 Gemma 3 4B SFT, text-only 0.85
  3. 3 InternVL3.5-8B SFT, text-only 0.85
  4. 4 Qwen3-VL-8B-Instruct SFT, text-only 0.84
  5. 5 Qwen3-VL-4B-Instruct SFT, text-only 0.84
  6. 6 Qwen3-VL-2B-Instruct SFT, text-only 0.84
  7. 7 InternVL3.5-4B SFT, text-only 0.81

Source: The AD4Edu-Bench paper

#ModelScoreSettingDateSource
1 SmolVLM2-2.2B-Instruct 0.86 SFT, text-only Feb 2025 Official leaderboard
2 Gemma 3 4B 0.85 SFT, text-only Mar 2025 Official leaderboard
3 InternVL3.5-8B 0.85 SFT, text-only Aug 2025 Official leaderboard
4 Qwen3-VL-8B-Instruct 0.84 SFT, text-only Oct 2025 Official leaderboard
5 Qwen3-VL-4B-Instruct 0.84 SFT, text-only Oct 2025 Official leaderboard
6 Qwen3-VL-2B-Instruct 0.84 SFT, text-only Oct 2025 Official leaderboard
7 InternVL3.5-4B 0.81 SFT, text-only Aug 2025 Official leaderboard

Rankings by part

We read 8 parts from the benchmark’s paper, each ranking models on one part of the test, such as AD4Edu-Eval mechanical, AD4Edu-Eval local and chrF. Qwen3-VL-8B-Instruct leads 3 of them, and 3 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.

AD4Edu-Eval mechanical

7 models we track have a score here.

  1. 1 SmolVLM2-2.2B-Instruct 0.86
  2. 2 Gemma 3 4B 0.85
  3. 3 InternVL3.5-8B 0.85
  4. 4 Qwen3-VL-8B-Instruct 0.84
  5. 5 Qwen3-VL-4B-Instruct 0.84
  6. 6 Qwen3-VL-2B-Instruct 0.84
  7. 7 InternVL3.5-4B 0.81

Scores at each effort level

Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.

Lowest effort to highest, light to dark
SmolVLM2-2.2B-Instruct
DPO, multim… SFT, multim… DPO, text-o… SFT, text-o…
0.86
Gemma 3 4B
DPO, text-o… SFT, text-o…
0.85
InternVL3.5-8B
DPO, text-o… SFT, text-o…
0.85
Qwen3-VL-8B-Instruct
DPO, text-o… SFT, text-o…
0.84
Qwen3-VL-4B-Instruct
DPO, text-o… SFT, text-o…
0.84
Qwen3-VL-2B-Instruct
DPO, multim… DPO, text-o… SFT, multim… SFT, text-o…
0.84
InternVL3.5-4B
DPO, text-o…
0.81

Source: The AD4Edu-Bench paper

ModelEffortHarnessScoreCost per taskOutput tokens
SmolVLM2-2.2B-Instruct SFT, text-only 0.86
Gemma 3 4B SFT, text-only 0.85
InternVL3.5-8B SFT, text-only 0.85
Qwen3-VL-8B-Instruct SFT, text-only 0.84
Qwen3-VL-4B-Instruct SFT, text-only 0.84
Qwen3-VL-2B-Instruct SFT, text-only 0.84
SmolVLM2-2.2B-Instruct DPO, text-only 0.83
InternVL3.5-4B SFT, text-only 0.81
InternVL3.5-8B DPO, text-only 0.81
InternVL3.5-4B DPO, text-only 0.81
SmolVLM2-2.2B-Instruct SFT, multimodal 0.79
Qwen3-VL-2B-Instruct SFT, multimodal 0.78
Qwen3-VL-4B-Instruct DPO, text-only 0.78
Qwen3-VL-2B-Instruct DPO, text-only 0.76
Gemma 3 4B DPO, text-only 0.75
Qwen3-VL-8B-Instruct DPO, text-only 0.74
SmolVLM2-2.2B-Instruct DPO, multimodal 0.72
Qwen3-VL-2B-Instruct DPO, multimodal 0.63

Number one over time

Lab claims count from the day the card came out, independent runs from the model’s release.

Hugging Face
Benchmark paper Not tested since Oct 2026

Source: The AD4Edu-Bench paper

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Plot with other benchmarks
In the benchmark’s paper
0.70 0.80 0.90 2025 2026 0.70 0.80 0.90 2025 2026

Source: Official AD4Edu-Bench leaderboard

No lab model card reports it yet

We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.

More in Audio and video

All 695 in Audio and video
  • Video-MMMU In use 300 expert-level lecture videos with questions testing perception, comprehension and adaptation of knowledge. Accuracy. Qwen 3.8 Max 88.7%
  • LVBench In use Hours-long video understanding with multiple-choice questions on reasoning, grounding and summarization. Gemini 4 Argon 91.7%
  • MMVU In use 3,000 expert-annotated questions on videos across 27 disciplines. Multiple-choice and open-ended accuracy. Qwen 3.8 Max 82.4%
  • MLVU In use Multi-task long video understanding (minutes to hours) with multiple-choice and generation tasks. Accuracy. Qwen 3.8 Max 90.8%
  • FLEURS In use Models transcribe read speech in 102 languages built from FLoRes-101 sentences. Model cards usually report word error rate, so lower scores are better. Phi-4-multimodal 4.00%
  • VoiceBench In use Spoken instructions with speaker, environment and content variation. Voice assistant responses scored by GPT judge or rules. Qwen3.5-Omni-Plus 93.1%

Data audit

7 of 9 checks pass, 2 warnings. Look first at the Task access check. We don’t know whether the tasks are public. Every benchmark

Catalog record 1 warning 1 pass
  • Warning Task access We don’t know whether the tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 1 warning 2 pass 2 notes
  • Pass Independent runs Scores from the benchmark’s paper (7 entries).
  • Pass Models scored 7 models we track.
  • Warning Newest score The newest came 13 months ago, in October 2025, for Qwen3-VL-2B-Instruct.
  • Note Frontier models None of its models are on the capabilities index.
  • Note Headroom Scored as a plain number, with no fixed ceiling.
Leaderboard import 3 pass
  • Pass Importer We read the board on arxiv.org: 18 rows for models we track at the last read.
  • Pass Last read November 20th, 2026.
  • Pass Read worked The last read worked.
Consistency 1 pass 1 note
  • Note Lab claims No model has both a lab number and an independent run.
  • Pass Known issues No published audit found problems.

How we check