ChemRD-Hard-100

100 bilingual (English / Chinese) PhD-level chemistry items, selected by measured difficulty from a 474-item verified pool. Every item is self-contained: everything needed to answer it is in the item.

Leaderboard

18 arms, each one isolated process or request per item, both languages, no shared context between items. Reasoning is off wherever the endpoint allows it, and the last column reports the reasoning tokens actually observed, so the claim is measured rather than asserted.

model	runner	reasoning	overall	single	multi (exact)	short answer /10	EN	ZH	reasoning tokens
GPT-5.6-sol	codex	off	72.5%	95.8%	68.8%	6.41	71.1%	73.8%	0
Claude Opus 5	claude CLI	off	71.1%	97.9%	75.0%	5.93	73.4%	68.8%	0
Claude Sonnet 5 *	claude CLI	ON (default)	68.4%	93.8%	87.5%	5.32	71.8%	65.0%	0
GPT-5.5	codex	off	66.0%	93.8%	59.4%	5.67	66.5%	65.6%	27
GPT-5.6-terra	codex	off	62.9%	87.5%	62.5%	5.33	62.9%	63.0%	0
Kimi-K3	openrouter	off	61.4%	91.7%	68.8%	4.73	62.0%	60.8%	0
Claude Sonnet 5	claude CLI	off	60.6%	93.8%	62.5%	4.67	63.2%	57.9%	0
Qwen3.8-Max	openrouter	MIN (cannot disable)	60.0%	97.9%	65.6%	4.33	60.3%	59.6%	333,572
Grok-4.5	openrouter	MIN (cannot disable)	59.5%	97.9%	65.6%	4.24	59.4%	59.5%	106,932
Gemini-3.1-Pro	openrouter	MIN (cannot disable)	59.4%	95.8%	65.6%	4.32	59.8%	59.0%	158,865
MiniMax-M3 *	REST	ON (adaptive)	55.8%	85.4%	65.6%	4.10	60.0%	51.4%	943,877
Gemini-3.5-Flash	openrouter	off	54.4%	93.8%	62.5%	3.63	55.3%	53.4%	0
GPT-5.6-luna	codex	off	53.5%	79.2%	40.6%	4.67	56.9%	50.2%	0
MiniMax-M2.7 *	REST	ON (adaptive)	49.4%	87.5%	53.1%	3.27	54.7%	44.0%	0
DeepSeek-V4-Flash	openrouter	off*	45.6%	79.2%	40.6%	3.34	47.0%	44.1%	8,789
Nemotron-3-Ultra-550B	openrouter	off	44.4%	83.3%	21.9%	3.48	44.3%	44.5%	0
Hunyuan-3	openrouter	off	44.1%	70.8%	50.0%	3.19	45.0%	43.3%	0
GLM-5.2	openrouter	off	43.3%	68.8%	31.2%	3.63	42.8%	43.8%	0

* scored against the corpus as it stood on 2026-08-12, before the last self-containment repairs.

Machine-readable in leaderboard.json, per-item scores in per_item_scores.jsonl, raw model outputs in runs/.

Reasoning could not be switched off everywhere

Four endpoints reject reasoning={"enabled": false} outright with HTTP 400, "Reasoning is mandatory for this endpoint and cannot be disabled": Gemini 3.1 Pro, Gemini 3.5 Flash, Qwen3.8-Max and Grok 4.5. Those arms ran at the lowest setting each endpoint accepts, effort: minimal, and the residual is in the table: Qwen3.8-Max spent 333,572 reasoning tokens over 200 runs, Gemini 3.1 Pro 158,865, Grok 4.5 106,932. Gemini 3.5 Flash accepted minimal and returned zero, so it belongs with the genuinely reasoning-free arms.

Rows marked "MIN (cannot disable)" are therefore not comparable with rows marked "off". Grok 4.5 illustrates why the flag alone cannot be trusted: on a trivial probe it used 22 reasoning tokens, and on the real items it averaged 535 per run. A setting named "minimal" is a floor on the budget, not an off switch. DeepSeek-V4-Flash is marked off* because it accepts the disable flag and then ignores it on 5 of 200 runs, for 8,789 tokens.

Three things the table does not say on its own

Sampling noise is about ±5 points. GPT-5.6-terra was run twice over identical text and settings and the two runs disagreed on 10% of choice runs, scoring 77.5% and 75.0%. Differences smaller than that are not model differences. Details in variance_check.json.

Reasoning is worth about 8 points. Claude Sonnet 5 scores 68.4% with reasoning on and 60.6% with it off, on the same items. Every cross-line comparison in this table compares two things at once.

Three arms ran on an earlier version of the corpus (marked *), before the final self-containment repairs. GPT-5.6-terra shows what that is worth: 46.5% on the old text against 62.9% on the current text, a 41-point swing on the choice items alone, far outside the sampling noise. Those three numbers should be treated as a lower bound rather than as a measurement of the current set.

Audit

Every arm records whether reasoning actually stayed off:

• codex arms (sol, luna, 5.5, terra): 0 reasoning tokens and 0 shell calls across all runs, except GPT-5.5 which emitted 27 reasoning tokens on 1 of 200
• OpenRouter arms: Nemotron and GLM-5.2 reported 0 reasoning tokens; DeepSeek-V4-Flash reported 8789 reasoning tokens across 5 of its 200 runs, so it does not fully honour reasoning.enabled=false
• Claude has no "none" setting, so --effort low was measured rather than assumed: on one item the default took 13.7 s and 1201 output tokens against 3.6 s and no reported token count at low, with the reply falling from 449 to 142 characters

Where a provider returns the chain of thought separately, or wrapped in <think> tags inside the content, it is stripped before scoring. Scoring a model's thinking as its answer is not scoring its answer.

Files

file	contents
hard100_solver.jsonl	what a model is shown. 11 fields, every answer-bearing field removed and verified absent
hard100_key.jsonl	gold answers, rubrics and explanations, keyed by item_id
hard100_full.jsonl	complete records including provenance and every measurement
leaderboard.json	the table above, machine readable
per_item_scores.jsonl	every arm's score on every item
variance_check.json	the repeat-run noise measurement
runs/*.jsonl	raw model outputs, one row per item-language, including the unparsed response

How difficulty was decided

Not by a label. The parent corpus carried D2/D3/D4 tags whose measured panel accuracy was 0.996, 0.988 and 0.989 across the three tiers, i.e. no discrimination at all. Difficulty here is the mean score of four models (Claude Sonnet 5, MiniMax-M3, MiniMax-M2.7, GPT-5.6-terra), each run on every item in both languages.

Ranking on four models rather than one is deliberate and was tested. Selecting the items a single model fails gives a set that measures that model's idiosyncrasies: MiniMax-M3 misses only 29 of 585 choice item-language pairs, 52% of those are answered correctly by all three other models, and 17% are hard by any shared standard. That model's own score on such a set is also fixed by the selection rule rather than measured.

Discrimination: the four selection models disagree on 51 of 100 items; only 1 is failed by all four.

What a low score does not mean here

A low score can come from a broken item rather than a hard one, so four checks ran before difficulty was considered.

1. Answer leakage. Every choice item was tested by showing a strong model the option texts alone, with the question and all data deleted. The corpus began at 0.925 on single choice against a 0.248 chance baseline: most of any score was recoverable without doing chemistry. After repair, the mean over this set is 0.030. Items above 0.25 are excluded. Each repair was confirmed by repeated independent probes, because a single miss is 75% likely by chance with four options.

2. Unanswerability. Every item was also solved with its full context. An item whose gold cannot be reached from its own data is excluded, since failing it measures nothing.

3. Second correct answers. An independent chemistry read of all rewritten distractors found items where a distractor had become true or defensible. No probe detects this: a second true option passes the leakage probe, the validity solve and every mechanical guard.

4. Self-containment. All 100 items were read line by line for references that survived extraction from the source paper: phrases such as "the standard conditions" where no conditions are stated, labels implying siblings that never appear, abbreviations never expanded. 49 such defects across 42 items were repaired by rewording only, never by supplying the missing data, and a separate model checked every repair for invented facts.

Short-answer items were checked the same way, by answering them from the question alone with no data and grading against their own key points. Items where the data added nothing were repaired or dropped.

Composition

items	100
single choice	24
multiple select	16
short answer	60
languages	English and Chinese, every item in both
source papers	62 distinct, at most 8 items from any one

Domains: asymmetric_catalysis_and_stereochemistry (29), photoredox_and_metallaphotoredox (19), radical_chemistry_and_radical_relay (17), transition_metal_catalysis_and_speciation_Ni_Cu (14), reaction_development_controls_and_troubleshooting (7), computational_mechanism_kinetics_and_photophysics (6), route_design_process_and_green_constraints (4), flow_photochemistry_process_and_reproducibility (4)

Scoring

• single choice: exact match
• multiple select: exact set match. Report per-option precision and recall too; exact match alone hides how close a near miss was
• short answer: each item carries atomised key_points with integer weights summing to 10. Grade against those, not against the gold prose, and do not reward fluency

from datasets import load_dataset
items = load_dataset("EscheWang/ChemRD-Hard-100", "solver")["train"]
key   = load_dataset("EscheWang/ChemRD-Hard-100", "key")["train"]

Provenance and limitations

Every item carries source_pointers with a DOI and the location within the paper. Local filesystem paths have been stripped. provenance_quote holds short verbatim excerpts from the source papers (mean 309 characters); if this dataset is ever made public, those excerpts need a copyright review first.

Repairs ran under guards that refuse any edit changing a number, the gold set, or the key-point weights, and both languages were checked to assert the same chemistry on every number, unit, negation, condition and stereodescriptor.

Known limitations. 60 of the 100 items are short answer, because difficulty genuinely concentrates there; an unconstrained selection would have been 80% short answer, and a floor on the choice types raised the set mean from 50.9% to 54.9%. 21 items from the parent pool remain solvable from their options alone after seven repair rounds and are excluded from this set rather than shipped as fixed.