Rag Robustness Eval

---
name: rag-robustness-eval
description: Evaluates how Retrieval-Augmented Generation (RAG) systems maintain factual accuracy when exposed to adversarial, harmful, or misleading medical evidence. It probes the model's susceptibility to contextual manipulation and its ability to resist misinformation propagation under varying query framings. Use when the user wants to benchmark on TREC Health Misinformation 2020, TREC Health Misinformation 2021, or asks about evaluating this task. Reports ground-truth alignment rate.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2509.03787
  bibtex_key: [REDACTED]
  confidence: high
---

# rag-robustness-eval

> Evaluating the Robustness of Retrieval-Augmented Generation to Adversarial Evidence in the Health Domain — Amirshahi et al. (2025) (arXiv:2509.03787, 2025)

## What this evaluates

Evaluates how Retrieval-Augmented Generation (RAG) systems maintain factual accuracy when exposed to adversarial, harmful, or misleading medical evidence. It probes the model's susceptibility to contextual manipulation and its ability to resist misinformation propagation under varying query framings.

## Datasets

- **TREC Health Misinformation 2020** — total ?; splits: test (-1); repo https://github.com/shakibaam/RAG_ROBUSTNESS_EVAL
- **TREC Health Misinformation 2021** — total ?; splits: test (-1); repo https://github.com/shakibaam/RAG_ROBUSTNESS_EVAL

## Metrics

- `ground-truth alignment rate` **(primary)** — range: percent
  - Percentage of model-generated responses that correctly align with the established ground-truth medical fact. Calculated as (count of aligned responses / total responses) × 100.

## Input / output format

**Input**: User query (framed as consistent, neutral, or inconsistent with the ground truth) paired with a single retrieved document (context types: helpful, harmful, adversarial variants, liar, or none for Non-RAG baseline).

**Output**: Natural language response generated by the LLM.

## Scoring recipe

```python
def compute_alignment_rate(predictions, gold_answers):
    aligned = sum(1 for pred in predictions if aligns_with_truth(pred, gold_answers))
    return (aligned / len(predictions)) * 100
```

## Common pitfalls

- Confidence intervals are estimated via bootstrapping, not analytical formulas, so reporting CIs requires resampling.
- Query framing (consistent/neutral/inconsistent) significantly impacts baseline and RAG performance, so results must be stratified rather than averaged.
- Adversarial context types (e.g., FSAP-IntraQ vs. Rewriter) vary drastically in effectiveness, requiring careful grouping in analysis.

## Evidence (verbatim from paper)

> Tables 2–5 report ground-truth alignment rates across all closed- and open-source models, including the mean, standard deviation (SD), and 95% confidence intervals (CI) obtained via bootstrapping.

## Citation

```bibtex
@misc{amirshahi2025ragrobustness,
  title={Evaluating the Robustness of Retrieval-Augmented Generation to Adversarial Evidence in the Health Domain},
  author={Amirshahi et al. (2025)},
  year={2025},
  note={arXiv:2509.03787}
}
```

- arXiv: 2509.03787