Sealqa Eval

---
name: sealqa-eval
description: Evaluates a model's ability to reason over noisy, conflicting, and ambiguous real-world search results. It probes complex skills like contradiction resolution, temporal tracking, false-premise detection, and multi-document needle-in-a-haystack retrieval. Use when the user wants to benchmark on SealQA, or asks about evaluating this task. Reports accuracy.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2506.01062
  bibtex_key: pham2025sealqa
  confidence: high
---

# sealqa-eval

> SealQA: Raising the Bar for Reasoning in Search-Augmented Language Models — Pham et al. (2025) (arXiv:2506.01062, 2025)

## What this evaluates

Evaluates a model's ability to reason over noisy, conflicting, and ambiguous real-world search results. It probes complex skills like contradiction resolution, temporal tracking, false-premise detection, and multi-document needle-in-a-haystack retrieval.

## Datasets

- **SealQA** — total ?; splits: test (-1)

## Metrics

- `accuracy` **(primary)** — range: percent
  - Percentage of questions answered correctly. Calculated as the number of correct predictions divided by the total number of questions.

## Input / output format

**Input**: A question, optionally accompanied by retrieved search results or a set of documents (one gold, k hard negatives).

**Output**: A natural language answer to the question.

## Scoring recipe

```python
correct = 0
for pred, gold in zip(predictions, gold_answers):
    if pred.strip().lower() == gold.strip().lower():
        correct += 1
accuracy = (correct / len(predictions)) * 100
```

## Common pitfalls

- Providing search results can actually decrease model accuracy due to noisy or conflicting information.
- Increasing test-time compute (e.g., higher reasoning effort) does not reliably improve performance and may harm it.
- Models struggle disproportionately with false-premise detection and cross-lingual reasoning compared to other categories.

## Evidence (verbatim from paper)

> Seal-0 and Seal-Hard present significant challenges for frontier LLMs: Table[1] shows the accuracy of various LLMs on Seal-0 and Seal-Hard without access to a search engine (w/o search). Models perform poorly without web access, with accuracies ranging from 0.0% to 5.4% on Seal-0 and 0.0% to 22.4% on Seal-Hard.

## Citation

```bibtex
@misc{pham2025sealqa,
  title={SealQA: Raising the Bar for Reasoning in Search-Augmented Language Models},
  author={Pham et al. (2025)},
  year={2025},
  note={arXiv:2506.01062}
}
```

- arXiv: 2506.01062