atomic-canyon/NuclearBench

NuclearBench

1. Introduction

NuclearBench is a benchmark for evaluating how well language models answer technical questions from nuclear regulatory and engineering documents. It measures whether a model can produce answers that are both substantively complete and grounded in the provided source text.

In practice, language models in the nuclear domain operate through retrieval-augmented generation (RAG). A model's ingrained knowledge is unlikely to be sufficient for two reasons. First, much of the data that practitioners work with is export-controlled or proprietary and therefore absent from public training corpora. Second, nuclear facilities often operate under specific revisions of regulatory documents, which have been revised repeatedly through the relatively long life of the industry. A model that answers from its general training may cite requirements that do not apply to the plant in question.

NuclearBench isolates the answer-generation half of this pipeline. The model receives the complete source document rather than retrieving it, so the benchmark measures answer quality without confounding it with retrieval quality. Atomic Canyon previously released FermiBench, a separate benchmark for evaluating retrieval.

The benchmark was built by Atomic Canyon in collaboration with Idaho National Laboratory (INL) subject-matter experts. It contains 280 questions spanning 15 nuclear-domain topics, each paired with a criterion-based rubric and a full-length source document. Questions are drawn from NRC regulatory documents, operating-experience reports, and other nuclear technical materials.

NuclearBench evaluates two distinct dimensions of answer quality:

• Criteria coverage: how much of the required answer substance is present.
• Source faithfulness: whether the answer stays grounded in the provided document.

This separation matters because the two failure modes are different. An answer can be incomplete but faithful, or complete-looking but unfaithful. NuclearBench makes that distinction visible rather than collapsing it into a single score.

2. Why NuclearBench Matters

There is no widely adopted benchmark for evaluating language models on nuclear technical and regulatory work. Existing LLM benchmarks typically emphasize intrinsic knowledge, short-context question answering, or single-scalar grading. None of these capture what matters in the nuclear domain: whether a model can produce a useful, grounded answer from a long, messy technical document.

Nuclear documents are not clean reference material. They include inconsistent formatting, OCR artifacts, dense cross-references, and specialized terminology. Useful answers often require connecting facts across multiple sections of a single document. A benchmark that tests only short-passage extraction or general knowledge does not reflect this reality.

NuclearBench is designed to fill three gaps:

• Domain-specific: It is the first open-source benchmark purpose-built for the nuclear energy industry, with questions authored and validated by practicing nuclear engineers and subject-matter experts.
• Source-grounded: It evaluates whether a model can anchor responses to the provided documents instead of its memorized knowledge, replicating actual nuclear AI deployment requirements.
• Multi-dimensional: It reports coverage and faithfulness separately rather than collapsing them into a single score, so that materially different answer behaviors are visible rather than hidden.

3. Dataset Design

Each benchmark record consists of three components: a question, a full-length source document, and a criterion-based rubric.

Questions

NuclearBench contains 280 available questions (Core + Adversarial) and 125 holdout questions to detect contamination.

Core set (219 questions):

Questions from the benchmark pool covering the full range of topics and difficulty levels. Of the 219 questions, 71 were authored by subject-matter experts and 148 were synthetically generated and then validated by multiple subject-matter experts.

The benchmark tracks provenance for each question rather than treating human-authored and synthetic questions as interchangeable.

Example Question: “What four specific pathways are typically considered during post-closure safety assessments for near-surface disposal facilities to evaluate the migration of radionuclides to the surface environment?”

Adversarial set (61 questions):

Transformed versions of benchmark pool questions designed to test specific failure modes. Some contain planted errors in the question text, some have key identifiers stripped to force a broader search of the document, and some are paired with a mismatched document that the model should flag rather than force an answer from.

Example Question: “How did the isotopic ratio analysis of the boric acid residue show it came from a CEDM nozzle leak in the early 1990s?”

Why this is considered adversarial: This is a planted error issue. The LLM should identify that the residue came from a reactor vessel flange leak in 1988, not a CEDM nozzle leak in the early 1990s. Acceptable answers are either a refusal to answer as the document does not address this or an identification of the error and then answering the corrected question.

Holdout set (125 questions):

In addition to the 280 published questions, a holdout set of 125 questions is withheld from public release. The holdout set is used to detect benchmark contamination, where a model's training data includes benchmark questions or answers. This is standard practice for modern LLM benchmarks and allows us to verify that strong performance reflects genuine capability rather than memorization.

These questions were not used in the model performance findings below.

Source Documents

Questions are grounded in 247 unique source documents drawn from nuclear technical and regulatory materials, including:

• Licensee Event Reports (LERs)
• NUREGs and NUREG-0800 Standard Review Plan sections
• NRC Regulatory Guides
• NRC Inspection Reports
• Safety evaluations, policy statements, and license amendments

These documents are long, specialized, and often messy in the way real technical material is: OCR quality varies, formatting is inconsistent, and useful answers frequently require connecting facts across multiple sections.

Only publicly available source documents were used. This is a deliberate choice. Much of the nuclear domain is subject to export controls and access restrictions that would limit who can use or reproduce a benchmark. By building exclusively on public documents, NuclearBench can be freely distributed, independently verified, and used by any organization.

Criteria

Each question is paired with a rubric made up of individual criteria. A criterion is a single, concrete fact or requirement that a correct answer should include. For example, a criterion might state that the answer must mention a specific regulatory limit, identify a particular failure mode, or reference a defined inspection interval. Each criterion is scored as met or not met, and the collection of criteria for a question forms its rubric.

Example Question: “What four specific pathways are typically considered during post-closure safety assessments for near-surface disposal facilities to evaluate the migration of radionuclides to the surface environment?”

Criteria:

• Transport by water (dissolved or in suspension)
• Direct transfer as a result of human intrusion
• Release as a result of natural disruptive events
• Transport as gas

Criteria focus on substantive content requirements, not writing quality or style. The benchmark contains 1,013 total criteria across the 280 questions used for analysis.

Criteria were generated by an LLM from a canonical answer for each question. The canonical answers were either human-written by INL subject-matter experts or LLM-generated from the source document. Every criterion was then reviewed by two subject-matter experts before inclusion in the benchmark.

4. Evaluation Method

A model is given a question and the full source document, then generates an answer. That answer is evaluated along two independent dimensions by separate LLM judges.

CriteriaJudge: Is the answer correct?

CriteriaJudge evaluates whether the answer contains the right information. Each criterion is judged in isolation. The judge receives the question, the model's answer, and a single criterion. It decides whether the answer satisfies that criterion (met or not met). The judge does not have access to the source document.

The criteria score for a question is the fraction of its rubric criteria that the answer satisfies. The benchmark reports the average criteria score across all questions.

FineSurE: Is the answer faithful to the document?

FineSurE evaluates whether the answer stays grounded in the provided source document. It splits the model's answer into individual sentences, then classifies each sentence against the source document. Each sentence is categorized as either supported by the source ("no error") or as containing a specific type of factual error (entity error, predicate error, out-of-context error, and others).

The FineSurE score for a question is the fraction of answer sentences classified as having no error. The benchmark reports the average FineSurE score across all questions.

5. Coverage

NuclearBench covers a broad cross-section of the nuclear regulatory and technical landscape. Questions span 15 topics grouped into 6 higher-level categories. We observed no significant difference in model performance across these categories.

Category	Topic	Questions
Safety & Risk	Safety & Accident Analysis	37
	Security & Safeguards	16
	Probabilistic Risk Assessment	10
	Emergency Preparedness	5
Reactor Engineering	Reactor Systems & Design	38
	Instrumentation, Controls & Electrical	18
	Fire Protection	7
Fuel, Waste & Regulatory	Licensing, Regulation & QA	42
	Fuel Cycle & Performance	6
	Waste, Storage & Decommissioning	6
Operations & Maintenance	Operations & Plant Performance	38
Materials & Structures	Materials & Degradation	19
	Structural Integrity & Inspection	15
Radiation, Health & Environment	Radiation Protection & Health Physics	12
	Environmental Monitoring & Siting	11

6. Model Performance Findings

We evaluated 16 models from 6 providers on both the core and adversarial sets. All error bars in figures are 95% bootstrap confidence intervals (10,000 resamples, percentile method). Figures are sorted by model release date; tables are sorted by criteria score.

Core Set

Model	Criteria Score	FineSurE Score
Claude Fable 5.1	95.1%	91.3%
Claude Opus 5.5	95.0%	91.5%
Gemini 4 Argon	93.6%	89.3%
Claude Sonnet 5.5	92.8%	88.3%
GPT-6 Astra	90.6%	92.7%
GPT-6.1 Sol	90.1%	93.7%
Claude Haiku 5.5	87.2%	88.4%
GPT-6.1 Terra	86.8%	94.9%
Gemini 3.8 Flash	84.8%	87.3%
Mistral Large 4	84.1%	87.1%
GPT-6 Luna	81.3%	93.4%
DeepSeek V4 Pro	80.3%	93.3%
Nemotron 3.5 Super VL	78.6%	90.5%
Nemotron 3.5 Lightning	74.6%	77.5%
Gemma 4 31B	72.4%	90.0%
Gemini 3.5 Flash-Lite	70.9%	90.4%

Models are ordered by descending criteria score

Adversarial Set

Model	Criteria Score	FineSurE Score
Claude Fable 5.1	99.5%	87.4%
Claude Opus 5.5	97.8%	88.5%
Claude Sonnet 5.5	97.8%	85.2%
GPT-6 Astra	97.8%	91.2%
Gemini 3.8 Flash	97.4%	86.3%
Claude Haiku 5.5	97.3%	85.0%
GPT-6.1 Sol	96.2%	88.2%
DeepSeek V4 Pro	95.1%	87.7%
Gemini 4 Argon	94.5%	84.4%
GPT-6.1 Terra	92.2%	86.9%
GPT-6 Luna	90.2%	79.3%
Mistral Large 4	86.1%	78.2%
Nemotron 3.5 Super VL	84.3%	86.0%
Gemma 4 31B	80.6%	82.2%
Gemini 3.5 Flash-Lite	78.0%	75.8%
Nemotron 3.5 Lightning	66.0%	72.9%

Models are ordered by descending criteria score

The adversarial set tests whether a model can recognize that something is wrong and push back rather than comply. This is a qualitatively different skill from straightforward document extraction, and the results reflect that.

Faithfulness to source documents drops for every model on the adversarial set. The models that score best on the core set are not consistently the ones hit hardest.

Document Length (Criteria Score, Core Set only)

Model	Short	Medium	Long
Gemini 4 Argon	98.5%	94.4%	88.0%
Claude Opus 5.5	98.1%	95.9%	91.1%
Claude Fable 5.1	97.8%	95.3%	92.1%
Claude Sonnet 5.5	96.9%	94.1%	87.2%
Claude Haiku 5.5	94.9%	87.1%	79.4%
GPT-6 Astra	92.5%	93.9%	85.4%
GPT-6.1 Sol	92.5%	91.5%	86.3%
GPT-6.1 Terra	90.8%	84.7%	84.9%
Gemini 3.8 Flash	90.5%	85.9%	78.1%
Mistral Large 4	86.2%	85.7%	80.5%
Nemotron 3.5 Super VL	83.0%	80.1%	72.7%
DeepSeek V4 Pro	82.8%	84.3%	73.8%
GPT-6 Luna	81.4%	81.2%	81.4%
Nemotron 3.5 Lightning	78.9%	72.7%	72.2%
Gemini 3.5 Flash-Lite	78.5%	69.4%	64.8%
Gemma 4 31B	77.4%	74.7%	65.1%

Questions are split into three equal bins by document length in characters (Short 904–17,181, Medium 17,197–66,150, Long 68,227–432,564). Models are ordered by descending Short score.

We split the core questions into three equal bins by source document length and computed criteria scores per bin. Most models score lower on longer documents, but the size of the drop varies. The drop tends to be larger for models with lower overall scores.

7. Limitations

NuclearBench has several limitations that motivate future work:

Prompt Engineering

Prompt engineering has a significant impact on response quality, which we have not evaluated here. This benchmark uses simple, transparent prompts rather than complex, model-specific optimizations. We anticipate that real-world deployments will involve customizing prompts to specific applications, a process that this benchmark readily facilitates.

Real-World Context Challenges

The adversarial set explores an important source-context failure mode: a provided document that appears relevant because it shares key terms with the query, but does not actually contain the answer. Models may then force-fit an answer from mismatched context rather than recognize that the evidence is insufficient or wrong.

This represents only a narrow slice of context challenges that can occur in real generative pipelines, such as:

• Multiple retrieved documents where only one or none contain a valid answer
• Documents from the wrong revision or time period, where the facts are plausible but temporally inconsistent with the user’s question
• Multi-hop queries, where a correct answer requires combining evidence across multiple retrieved documents

8. Conclusion

Language models are entering nuclear engineering workflows. The question is no longer whether they will be used, but whether we can measure how well they perform on the kind of work that actually matters: answering technical questions from real documents, under conditions where unsupported claims carry real consequences. NuclearBench provides a public, reproducible baseline for that measurement, built on real nuclear documents and validated by practicing nuclear engineers.

9. Acknowledgements

NuclearBench was built in collaboration between Atomic Canyon and Idaho National Laboratory. INL subject-matter experts contributed to question authoring, source-document pairing, criteria validation, and benchmark review. Atomic Canyon conducted benchmark engineering, synthetic question generation, evaluation pipeline development, and release preparation.

We gratefully acknowledge the following INL subject-matter experts for their contributions (listed alphabetically):

Abdalla Abou-Jaoude · Samuel Bays · Denver Conger · Jason Hales · Jackson Harter · Ross Hays · Robert Kile · Travis King · Travis Lange · Jieun Lee · Linyu Lin · Chad Pope · Emerald Ryan · Harleen Kaur Sandhu · Mark Schanfein · Ashley Shields · Sonali Sinha Roy · Ryan Stewart · Gerhard Strydom · Vaibhav Yadav · Patience Yockey

10. License

NuclearBench data and code are released under the Apache-2.0 license.

Appendices

Appendix A: Model Eligibility

NuclearBench is a full-document benchmark. The model receives the complete source document as context, so the benchmark requires models that can reliably operate at modern long-context lengths. Models must support a context length of at least 110K tokens to run the benchmark in this form.

Appendix B: Personal and Sensitive Information

The benchmark contains no personal or sensitive information. All source documents are drawn from publicly available nuclear technical and regulatory materials.

Appendix C: Prompts

Question-Answering Prompt

The following prompt is given to the model being evaluated, with {context} and {question} substituted:

Use only the provided context to answer the question. If the context is about a different document, facility, or system than the question asks about, state this before answering. Include all relevant details found in the context such as specific values, dates, conditions, component identifiers, procedural requirements, causal mechanisms, and negative findings. Do not elaborate beyond what the context states. The answer must be self-contained. Restate dates, locations, and identifiers from the question that belong in the answer.

<context>
{context}
</context>

<question>
{question}
</question>

Answer:

CriteriaJudge Prompt

Each criterion is judged independently. The judge receives the question, the model's answer, and a single criterion. It does not receive the source document.

## Question (context only; not evidence)
{question}

## Answer
{answer}

## Criterion
{criterion}

Decide whether the answer satisfies the criterion.

Evaluation guidance:
- Use only the answer text as evidence. Do not give credit for information that appears only in the question.
- Do not infer unstated facts.
- First, identify the key facts in the criterion (specific values, names, standards, thresholds, conditions, or relationships).
- The criterion is satisfied if and only if every key fact is clearly present in the answer, whether stated verbatim or as an unambiguous paraphrase.
- Do not require the answer to reproduce the criterion's exact sentence structure, connecting phrases, or rationale — only the key facts matter.
- If a key fact appears in the answer in a different section, context, or sentence order, it still counts.
- If the answer is only partially correct (some key facts present, others missing), return false.

FineSurE Prompt

Each model answer is split into sentences. The judge classifies every sentence against the source document. A sentence is counted as faithful only when its category is "no error".

You will receive a transcript followed by a corresponding summary. Your task is to assess the factuality of each summary sentence across nine categories:
* no error: the statement aligns explicitly with the content of the transcript and is factually consistent with it.
* out-of-context error: the statement contains information not present in the transcript.
* entity error: the primary arguments (or their attributes) of the predicate are wrong.
* predicate error: the predicate in the summary statement is inconsistent with the transcript.
* circumstantial error: the additional information (like location or time) specifying the circumstance around a predicate is wrong.
* grammatical error: the grammar of the sentence is so wrong that it becomes meaningless.
* coreference error: a pronoun or reference with wrong or non-existing antecedent.
* linking error: error in how multiple statements are linked together in the discourse (for example temporal ordering or causal link).
* other error: the statement contains any factuality error which is not defined here.

Instruction:
First, compare each summary sentence with the transcript.
Second, provide a single sentence explaining which factuality error the sentence has.
Third, answer the classified error category for each sentence in the summary.

Transcript:
```
{source_text}
```

Summary with {num_sentences} sentences:
{sentences_str}

Appendix D: Canary String

To make it easier to filter this benchmark from training corpora and prevent models from directly retrieving ground-truth answers, we include the following canary string: nuclearbench:518afaba-c710-4b5d-9366-03785dbd3956

Appendix E: Usage

Loading the Data

from datasets import load_dataset

core = load_dataset("atomic-canyon/NuclearBench", split="test")
adversarial = load_dataset("atomic-canyon/NuclearBench", "adversarial", split="test")

Setup

Requires Python 3.10+ and uv. uv sync installs all dependencies, including the spaCy model en_core_web_md.

uv sync

Running the Benchmark

Core set (219 questions):

uv run python nuclearbench.py --model anthropic/claude-opus-5-5 --cache

Adversarial set (61 questions; the subset field is planted_error, vague, or wrong_doc):

uv run python nuclearbench.py --model anthropic/claude-opus-5-5 --dataset nuclearbench_adversarial.jsonl --output-dir results_adversarial --cache

Run uv run python nuclearbench.py --help for all options.

Models are called via LiteLLM. Each candidate model's settings (temperature, reasoning effort, provider pin) are in models.yaml; add an entry there to evaluate a new model. Set the API key environment variable for your provider (e.g. OPENAI_API_KEY, ANTHROPIC_API_KEY, OPENROUTER_API_KEY).

Everything a run produces goes under --output-dir: lighteval results and per-sample details, the completion cache, and litellm_calls.jsonl.

To check a data file against the schema:

uv run python validate_jsonl.py nuclearbench_core.jsonl