🔬 ChemArgus

From Answer Accuracy to Rubric-Grounded Reasoning Diagnosis
in Multimodal Expert-Level Chemistry

---

ChemArgus is a competition-level, multimodal, diagnostic benchmark for expert chemistry reasoning. Problems are drawn from chemistry competition material, textbooks, and research papers — and scored by fine-grained rubrics with partial credit, instead of answer endpoints.

Chemistry evaluation has moved past multiple-choice probes toward open-ended expert problems, yet current frontier benchmarks still score answer endpoints alone. Such scoring conflates memorized facts with stepwise derivation, treats visual and textual inputs as interchangeable, and confounds long-horizon planning with stepwise execution. ChemArgus replaces the answer endpoint with three measurement properties:

• Fine-grained scoring — every subquestion is graded point by point with partial credit, so a score lands at the level of the step.
• Multimodal capability — every problem with a real image is paired with a controlled text description, pricing visual understanding against text.
• Diagnostic attribution — paired metrics share one scoring unit and differ only in context construction, isolating integration, propagation, or local reasoning as the cause of a score gap.

---

📊 Dataset Partitions

	Share	Problems	Subquestions	Marks	Public images	Rubrics
🧪 dev · Public Dev Set	25%	48	210	636	268	✅ included
🏁 test · Held-out Test Set	55%	105	445	1387	268	❌ stripped
🔒 private · Private Set	20%	38	165	559	—	🔐 local only
Total	100%	191	820	2582	536

🎲 Seeded stratified split — the distribution of level (exam series), subfield (knowledge category), and modality (with / without model-visible images) stays consistent across all three partitions and matches the full set (max deviation ≈ 3 percentage points).

📁 Repository Layout

dev/
  dataset.jsonl              one JSON object per dev problem (with rubrics)
  instance_ids.json          the 48 dev problem ids
  tasks/<problem_id>/
    problem.json             full problem: prompt + rubric + image metadata
    images/*.png
test/
  dataset.jsonl              one JSON object per test problem (rubrics stripped)
  instance_ids.json          the 105 test problem ids
  tasks/<problem_id>/
    problem.json             problem prompt; rubric fields absent
    images/*.png

🧬 Data Format

Each problem is a JSON object:

{
  "problem_id": "Exam-01-T10",
  "header": "Light",
  "header_images": [
    {"file": "Exam-01-T10-header-1.png", "description": "..."}
  ],
  "subquestions": [
    {
      "sub_id": "Exam-01-T10-1",
      "tags": ["2.2", "2.4"],
      "text_open": "10-1 Light can excite certain chemical bonds. ...",
      "text_images": [
        {"file": "Exam-01-T10-1-1-prompt.png", "description": "..."}
      ],
      "full_mark": 8,
      "rubric": ["[C]: ... (4 points)", "D: ... (4 points)"],
      "rubric_images": [
        {"file": "Exam-01-T10-1-1-rubric.png", "description": "..."}
      ]
    }
  ]
}

Field	Meaning
problem_id	Exam-<nn>-T<task> — nn is the 01–23 exam index
sub_id	{problem_id}-{n} (sequential within the problem)
header	Shared background of the whole problem
text_open	Subquestion text
full_mark	Integer maximum score of the subquestion
rubric	Scoring points with partial-credit annotations — absent in test/
*_images[].file	Image file name (see naming scheme below)
*_images[].description	Text description of the image content

Image files follow a unified naming scheme: {problem_id}-{sub_index}-{n}-prompt|rubric.png (subquestion images) and {problem_id}-header-{n}.png (problem-level images).

⚠️ Image visibility — header_images and text_images are model-visible; rubric_images are private scoring images, given only to the judge, never to the answering model. They do not appear in test/ at all.

🔀 Evaluation Protocols

Three protocols share one scoring unit and differ only in context construction:

1. full_problem — the model receives the complete problem in one call and answers everything in one response.
2. sequential_carry — subquestions are answered in order; each answer is carried into the next step, with no access to gold answers.
3. oracle_scaffolded — each subquestion is answered after being given the correct result of the previous one; isolates local reasoning from error propagation.

⚖️ Scoring

The core metric is the Normalized Rubric Score: rubric-grounded points awarded divided by the problem's total marks. Every subquestion is graded by an LLM judge against the private rubric (partial credit per scoring point), with structure/SMILES equivalence judged by chemical meaning.

🏅 Results & Leaderboard

Snapshot: September 2026 · Normalized Rubric Score (%). Runs are sorted by full_problem within each input mode.

🥇 full_problem leaders — Multimodal: Claude Opus 5 53.8%† · Text: Claude Opus 5 61.5%† · Text + reader: GLM-5.2 52.0%

Model	Input	full_problem	sequential_carry	oracle_scaffolded
Claude Opus 5 †	multimodal	53.8	59.4	64.9
Gemini 3.8 Flash §	multimodal	48.5	58.5	60.3
Grok 4.6 §	multimodal	33.8	41.8	45.5
Qwen3.8-Max	multimodal	29.3	31.2	34.0
Kimi K3	multimodal	27.5	30.6	31.8
MiniMax M3	multimodal	27.3	27.5	27.9
GPT-5.6-Sol	multimodal	23.2	28.8	30.4
Claude Opus 5 †	text	61.5	61.7	61.7
Gemini 3.8 Flash §	text	46.5	59.0	59.7
Qwen3.8-Max	text	36.0	39.0	41.9
Kimi K3	text	34.5	36.0	37.5
Grok 4.6 §	text	33.4	42.0	46.0
GPT-5.6-Sol	text	29.8	35.6	36.5
MiniMax M3	text	26.2	28.7	29.1
DeepSeek V4 Pro	text	25.0	25.0	27.6
GLM-5.2	text	23.5	24.9	25.1
GLM-5.2	text + reader	52.0	52.2	52.4
Claude Opus 5 †	text + reader	46.2	53.9	54.7
DeepSeek V4 Pro	text + reader	36.7	36.8	37.0
Gemini 3.8 Flash §	text + reader	33.8	45.1	49.7
Kimi K3	text + reader	32.3	32.3	32.4
GPT-5.6-Sol	text + reader	31.1	33.6	34.0
Grok 4.6 §	text + reader	28.6	29.0	31.1
Qwen3.8-Max	text + reader	27.9	29.2	31.5
MiniMax M3	text + reader	27.2	27.3	27.3

† Claude Opus 5 runs rest on a small problem prefix (39–103 points) and are provisional until the full set completes. § Gemini 3.8 Flash and Grok 4.6 runs sit on a 12-paper, 1136-point subset. Percentages are computed on each run's own scored problem set, so denominators differ across runs.

Live leaderboard: https://www.cosmosmind.ai/leaderboard/chemargus

📖 Citation

@misc{chemargus2026,
  title  = {ChemArgus: From Answer Accuracy to Rubric-Grounded Reasoning
            Diagnosis in Multimodal Expert-Level Chemistry Benchmark},
  author = {Hanyu Guo and Mo Cui and Zihan Tan and Leixin Sun and Yichun Wang and
            Yinan Shen and Guojun Zhu and Jia Jiang and Yichen Fan and Zui Chen and
            Wenzhe Gao and Zhihao Yu and Qijun Wang and Yetian Zhu and Jiande Chen and
            Chuhang Pei and Yao Jiang and Erhao Chen and Pinhan Wang and Xiaorui Pan and
            Qianyi Zhu and Yang Liu and Mang Ye and Guancheng Wan},
  year   = {2026}
}