README.md

---
license: cc-by-nc-4.0
task_categories:
  - question-answering
language:
  - en
tags:
  - reasoning
  - mathematics
  - physics
  - chemistry
  - jee
  - failure-analysis
  - contamination-aware
size_categories:
  - n<1K
gated: auto
---

# NULLSPACE — 267 JEE questions frontier models get wrong

Most STEM benchmarks are saturated. Frontier models score 96–98% on ordinary JEE questions,
so an aggregate number tells you almost nothing about which model reasons better.

This set contains **only the questions they fail**. Every one was answered incorrectly by at
least one frontier model, and every answer key was independently reproduced by a *different*
model solving the question blind — so a failure here is a real failure, not a broken key.

**267 questions public. 77 more are held out and unpublished**, so the set can still measure
a model that has seen this page.

## ⚠️ Read this before you score anything

**This benchmark cannot fairly score the models that built it.**

Every key here was confirmed by some model re-solving the question. A model that already
solved a question cannot be tested on it. Every row therefore carries a `do_not_score` list,
and **honest evaluation means excluding those rows per model**:

| Model | Scorable rows |
|---|---|
| `gemini-3.6-flash` | 183 of 267 |
| `gemini-3.5-flash` | 174 of 267 |
| `claude-opus-5` | 165 of 267 |
| `gemini-2.5-pro` | 161 of 267 |
| `gpt-5.1` | 5 of 267 |
| `gemini-3.1-pro-preview` | 3 of 267 |

This is not hypothetical. Scoring `claude-opus-5` across rows it had touched gave **58.3%**;
on rows it had never seen it scored **43.0%**. Ignore the field and you overstate by roughly
15 points.

`gpt-5.1` and `gemini-3.1-pro-preview` produced nearly every failure this set is made of, so
almost nothing is left to score them on. That is inherent to a failure-only benchmark — it
exists to evaluate models that came *after* it.

`do_not_score` covers **both** roles: a model that confirmed a key has already solved the
question, and a model that failed it is guaranteed wrong. Failures are gathered from every
run rather than only the one that admitted the question — an earlier build missed 126 of them
and advertised 56 rows as scorable for `gpt-5.1` that it had already failed.

```python
import json
rows = [json.loads(l) for l in open("benchmark.jsonl")]

MODEL = "claude-opus-5"
eligible = [r for r in rows if MODEL not in r["do_not_score"]]
print(f"{MODEL}: score on {len(eligible)} of {len(rows)} rows")
```

## Results — three frontier flagships, same 267 questions

Every model below ran the identical set under identical conditions: same prompts, figures
attached, high reasoning effort, 267 questions each.

| Model | Org | Accuracy | Clean subset |
|---|---|---|---|
| `claude-opus-5` | Anthropic | **62.2%** | 67.3% (n=165) |
| `gpt-5.1` | OpenAI | **49.8%** | n=5 — too few to report |
| `gemini-3.1-pro-preview` | Google | **23.6%** | n=3 — too few to report |

**43 of 267 questions — 16% — were failed by all three simultaneously.** Only 23 were
solved by all three. On ordinary questions from the same source, these models score 96–98%.

### The ranking is not a capability ranking

Read the second column before the first. The ordering above tracks **how much each model
helped build the set**, not how capable it is:

| Model | Rows it failed or helped check |
|---|---|
| `gemini-3.1-pro-preview` | 264 of 267 |
| `gpt-5.1` | 262 of 267 |
| `claude-opus-5` | 102 of 267 |

Gemini swept the entire source bank, so nearly every question here is one it already got
wrong — its 23.6% is close to arithmetically forced, not a statement about the model.
Claude leads largely because it had the least involvement.

**The one uncontaminated figure in the table is Claude's 67.3% on the 165 rows it never
touched** — a vendor with no hand in selecting those questions still failing a third of
them. That, and the 16% all three failed, are the numbers worth quoting. The full leaderboard
is context.

Any model that did **not** help build this benchmark starts with all 267 rows eligible, and
its accuracy is directly comparable to the first column.

## How a question qualified

Both conditions were required:

1. a frontier model answered it wrong on a **scored** measurement — truncations and API
   errors are lost measurements, never failures; and
2. an independent model re-solved it **without sight of the answer key or of the failed
   model's answer**, with figures attached, and reached that same key.

Condition 2 is the whole point. A model disagreeing with a key means either the model was
wrong or the key was wrong, and the two are indistinguishable from the outside.

**Across 704 checked questions, 514 turned out to be wrong keys rather than model failures.**
Among questions where two vendors gave the *same* wrong answer, 81% were bad keys — measured
on three independent pools. Every one of those was excluded. Without this step the set would
be mostly broken questions rather than hard ones.

## What's in it

| | |
|---|---|
| Questions | 267 public (77 held out) |
| Chemistry / Physics / Mathematics | 122 / 82 / 63 |
| JEE Advanced / JEE Main | 208 / 59 |
| With figures | 101 (251 images included) |
| Failed by two vendors independently | 130 |
| Answer types | multiple-choice, single-choice, numerical, integer, match-the-following |

## Fields

| Field | Meaning |
|---|---|
| `item_id` | stable public identifier, `STEM-NNNN` |
| `exam`, `subject`, `answer_type` | where it came from and what kind of answer is expected |
| `question`, `options`, `column_a`/`column_b` | the question content; HTML with MathML |
| `answer`, `answer_key` | the key, plain string and structured object |
| **`do_not_score`** | **models that must be excluded when scoring this row** |
| `models_failed` | which models got this question wrong |
| `verdict` | `CONFIRMED_MODEL_FAILURE`, `DISPUTED_BUT_KEY_BACKED`, `DISPUTED` |
| `thumbnail` | first figure image (typed `Image`), null for text-only questions |

## Grading

Answers are compared programmatically:

- **Single choice / match-the-following** — one letter, exact match
- **Multi-select** — exact set match; partial credit is not correct
- **Numerical** — tolerance `max(0.011, 1% of |truth|)`, since keys are quoted to 2dp
- **Integer** — exact

Figure questions require the image. A model answering a diagram question without the diagram
is answering a different question.

## Confidence tiers

Rows are not equally strong.

| Tier | Meaning |
|---|---|
| `CONFIRMED_MODEL_FAILURE` | solvers reached the key, no dissent |
| `DISPUTED_BUT_KEY_BACKED` | solvers split, but at least one reached the key |
| `DISPUTED` | solvers split between the key and the failed answer — **weakest rows** |

For a high-confidence subset: `both_vendors_failed == true and verdict != "DISPUTED"`.

## Limitations

- **Not usable for `gpt-5.1` or `gemini-3.1-pro-preview`** — see above.
- **Wrong keys are the dominant failure mode of the source material and are not all gone.**
  Excluded ones were caught by cross-vendor checking, but the base rate is high enough that
  residual bad keys should be assumed, not ruled out.
- **The evidence is mostly model-on-model.** Model agreement cannot rule out a shared blind
  spot; a measured case exists where three models from one vendor agreed and a fourth from
  another reached a different, correct answer.
- **Single-solver rows are weaker than they look.** Adding a second vendor to 65 verdicts
  settled by one vendor changed 20 of them — about one in three.
- **Subject mix is uneven.** Models fail Mathematics least, so it is the smallest slice.
- **Solutions are not included.** Answer keys are; worked explanations are not.

## Held-out split

77 further questions exist and are deliberately unpublished. A benchmark meant to test future
models loses its value once all of it is public — published questions can be trained on, and
the first release becomes the last honest measurement. Contact the maintainers for held-out
evaluation.

## Citation

```bibtex
@misc{nullspace2026,
  title        = {NULLSPACE: A Failure-Only JEE Benchmark with Contamination Tracking},
  author       = {Nalandadata},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Nalandadata/NULLSPACE}},
  note         = {267 public questions; 77 held out}
}
```