---
license: apache-2.0
task_categories:
  - text-classification
  - question-answering
language:
  - en
tags:
  - web-agents
  - benchmark
  - moderation
  - investigation
  - llm-evaluation
  - mirage
  - wikipedia
  - reddit
  - e-commerce
pretty_name: MIRAGE — Do Web Agents Investigate Before They Decide?
size_categories:
  - n<1K
configs:
  - config_name: wikipedia
    data_files:
      - split: test
        path: data/wikipedia.jsonl
  - config_name: shopping
    data_files:
      - split: test
        path: data/shopping.jsonl
  - config_name: reddit
    data_files:
      - split: test
        path: data/reddit.jsonl
---

# MIRAGE — Do Web Agents Investigate Before They Decide?

> **Misleading Investigation Reveals Agent Gaps in Evidence.**

MIRAGE is a benchmark for **investigative competence** in autonomous web
agents — the ability to recognise when visible information is insufficient,
seek hidden context, and integrate discovered evidence into a final
decision. The benchmark spans three structurally distinct moderation and
policy enforcement domains, each engineered around a two-layer information
architecture: a **Surface Context** (visible, often misleading) and a
**Hidden Context** (decisive, accessible only through proactive
investigation).

| Config | Tasks | Environment | Verdict space |
|---|---:|---|---|
| `wikipedia` | 300 | self-hosted Kiwix mirror (port 8888) | APPROVE / FLAG / REVERT / ESCALATE |
| `shopping`  | 300 | self-hosted Magento Open Source (port 7780) | APPROVE / DENY / ESCALATE / REQUEST_EVIDENCE / FLAG_AND_MONITOR / BAN_ACCOUNT / CANCEL_ORDER |
| `reddit`    | 159 | self-hosted Postmill (port 9999) | 8 verdicts spanning approval, correction, ban, lock, escalation |
| **Total**   | **759** | | |

## Loading

```python
from datasets import load_dataset

wiki   = load_dataset("SyedNazmusSakib/mirage", "wikipedia", split="test")
shop   = load_dataset("SyedNazmusSakib/mirage", "shopping",  split="test")
reddit = load_dataset("SyedNazmusSakib/mirage", "reddit",    split="test")

print(wiki[0]["surface_context"]["instruction"])
```

## Per-config details

### `wikipedia` — Wikipedia Forensics (300)

Editorial-review tasks set on a self-hosted Kiwix mirror. Each task
simulates a Wikipedia review queue ticket about a flagged edit; the agent
investigates the article and (where applicable) cross-reference pages
before issuing one of four verdicts. Combines two task generations,
distinguished by `domain_specific.schema_version`:

- `wiki_v2` (200) — multi-hop tasks with 2–4 required URLs across
  cross-reference and editor-history pages.
- `wiki_v1` (100) — single-page editorial-review tasks (vandalism,
  subtle disinformation, COI).

Categories: `false_alarm` (72), `conflict_of_interest` (62),
`cross_reference_required` (52), `coordinated_editing` (43),
`vandalism_simple` (34), `hard_disinformation` (20),
`subtle_disinformation` (17). Difficulty: 156 medium / 113 hard / 31 easy.

### `shopping` — Shopping Admin (300)

Fraud and refund-adjudication tasks on a self-hosted Magento Open Source
admin panel. Investigation graphs list 5–10 required URLs spanning the
flagged order, customer profile, full order history, and related accounts.
Combines two task generations:

- `v3` (150) — returns/refunds lifecycle: `wardrobing`,
  `legitimate_defense`, `b2b_reseller`, `reshipping_fraud`,
  `account_takeover`.
- `v4` (150) — order-level transaction signals: `gift_card_fraud`,
  `cross_account_linking`, `promo_abuse`, `false_item_not_received`,
  `friendly_fraud_chargeback`.

Each category contains 30 tasks, evenly distributed.

### `reddit` — Reddit Moderation (159)

Content-moderation tasks on a self-hosted Postmill mirror with synthetic
users, posts, comments, and crosspost networks. The agent investigates
user profiles, cross-subreddit submission history, and temporal patterns
before issuing one of eight moderation verdicts.

Categories: `cross_subreddit_spam` (58), `user_history_context` (26),
`fact_checking_source_verification` (26), `coordinated_brigading` (25),
`fact_checking_multimodal` (24). Difficulty: 38 easy / 76 medium / 45 hard.

## Unified task schema

Every row across all three configs follows the same shape:

```json
{
  "task_id": "WIKI_V2_COI_001",
  "domain": "wikipedia",
  "category": "conflict_of_interest",
  "difficulty": "hard",
  "surface_context": {
    "instruction": "[REVIEW QUEUE — Case #...] ...",
    "start_url": "http://localhost:8888/..."
  },
  "hidden_context": {"summary": "Decisive evidence (gold)"},
  "investigation_graph": {
    "required_nodes": [
      {"url": "...", "evidence": "what to extract from this page"}
    ]
  },
  "ground_truth": {
    "decision": "REVERT",
    "acceptable_actions": ["REVERT", "FLAG"],
    "harmful_actions": [],
    "reasoning": "Gold reasoning trace"
  },
  "domain_specific": { "...": "preserved source-format fields" }
}
```

## Evaluation framework

Each agent trajectory is decomposed into three sequential stages:

- **Investigation Accuracy** (`Acc_inv`) — fraction of ground-truth
  evidence URLs the agent actually visited (fuzzy URL matcher).
- **Reasoning Accuracy** (`Acc_rsn`) — LLM-as-judge score against gold
  facts (DeepSeek-V3.2 at temp 0; r=0.92 with human annotators).
- **Decision Accuracy** (`Acc_dec`) — correctness of the final verdict
  with semantic-equivalence partial credit.

Plus secondary metrics: Investigative Efficiency (`E_inv`), Evidence
Quality (`Q_ev`), Investigative Hallucination Rate. Two run conditions:
**Raw** (open-ended) and **Hint** (evidence-anchored prompts).

See the
[GitHub repository](https://github.com/SyedNazmusSakib/mirage)
for the full code, paper, and reproduction harness.

## Citation

```bibtex
@inproceedings{mirage2026,
  title  = {MIRAGE: Do Web Agents Investigate Before They Decide?},
  author = {Anonymous},
  booktitle = {Submitted to NeurIPS},
  year   = {2026}
}
```

## License

Code and unified task annotations are released under Apache-2.0. The
underlying environments inherit their original licenses: Wikipedia
content under CC-BY-SA-3.0, Magento Open Source under OSL-3.0, Postmill
under AGPL-3.0. All user/order/post data referenced in tasks is
synthetic; surface tickets and editor identities are anonymized
fictional accounts.