README.md

---
license: cc-by-4.0
language:
- en
pretty_name: PHREEQC-MCQ (verified items and model predictions)
size_categories:
- 1K<n<10K
task_categories:
- question-answering
- multiple-choice
tags:
- chemistry
- geochemistry
- phreeqc
- llm-agents
- tool-use
- scientific-simulation
- benchmark
configs:
- config_name: items
  default: true
  data_files:
  - split: test
    path: items.jsonl
- config_name: predictions
  data_files:
  - split: test
    path: predictions.jsonl
---

# PHREEQC-MCQ: verified items and model predictions

Four-option multiple-choice items that require running [PHREEQC](https://www.usgs.gov/software/phreeqc-version-3), a geochemistry simulator from the USGS, together with every per-item model prediction from [Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy](https://arxiv.org/abs/2610.26836).

The items come from [PHREEQC-MCQ-200](https://arxiv.org/abs/2607.00436). This release contains the **158 items** whose answer keys were reproduced by re-executing PHREEQC 3.8.6; all main-text results of the paper use these items.

The agent harness, the frozen PHREEQC database, and full agent trajectories are on GitHub: [grenadeComing/same-simulation-different-evidence](https://github.com/grenadeComing/same-simulation-different-evidence).

## Load

```python
from datasets import load_dataset

items = load_dataset("grenadeComing/same-simulation-different-evidence", "items", split="test")
predictions = load_dataset("grenadeComing/same-simulation-different-evidence", "predictions", split="test")
```

## Fields

**`items`** (158 rows)

| Field | Meaning |
|---|---|
| `id` | Item id (numbering of the original 200-item pool) |
| `question` | The prompt given to models, options included |
| `answer` | Correct option letter |

**`predictions`** (8,851 rows: one per run and item)

| Field | Meaning |
|---|---|
| `run` | Run directory under `result/` in the GitHub repository |
| `model` | Model |
| `setting` | `agent`, `no_tools`, or `paired_replay` |
| `condition` | `Raw100k`, `TOC`, `Grep`, `Hybrid`, `TOC+Report`, `TOC+Round`, `Direct`, or `CoT` |
| `repeat` | Repeat index (paired replay only) |
| `item_id` | Item id, matching `items.id` |
| `prediction` | Predicted option letter; null when the agent gave no answer |
| `answer` | Correct option letter |
| `correct` | Whether the prediction is correct; missing answers count as incorrect |

Conditions: **Raw100k** returns the simulator report inline, truncated to 100k characters; **TOC** returns an index of report sections that the agent reads by line range; **Grep** gives the agent a regular-expression search tool; **Hybrid** returns the report inline when it fits and the TOC otherwise; **TOC+Report** and **TOC+Round** add an instruction to answer from the report or to round to the options' precision; **Direct** and **CoT** use no tools.

Rows are stored per run, and several conditions have more than one run. The TOC results on the 33 long-output items come from the `custom/pool200_tocfix_uncapped_capaffected45_*` runs, which replace the TOC reruns on those items; the run map in the GitHub README lists every run.

## Citation

```bibtex
@article{zhang2026samesimulation,
  title   = {Same Simulation, Different Evidence: Where Retrieved Tool Output Costs {LLM} Agents Accuracy},
  author  = {Zhang, Ke and Chundur, Sahchit and Owusu, Jerry Peprah and Abdolhosseini Qomi, Mohammad Javad and Raissi, Maziar},
  journal = {arXiv preprint arXiv:2610.26836},
  year    = {2026}
}

@article{zhang2026phreeqcmcq,
  title   = {{PHREEQC-MCQ-200}: A Diagnostic Benchmark for Tool-Augmented Scientific Simulator Agents},
  author  = {Zhang, Ke and Chundur, Sahchit and Abdolhosseini Qomi, Mohammad Javad and Raissi, Maziar},
  journal = {arXiv preprint arXiv:2607.00436},
  year    = {2026}
}
```

## License

The items are released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). Model predictions are outputs of the evaluated models; their use is also subject to each provider's terms.