README.md

---
license: cc-by-4.0
language:
- en
- es
task_categories:
- question-answering
- text-retrieval
tags:
- retrieval-augmented-generation
- rag
- attribution
- interpretability
- context-attribution
- inference-time-scaling
- information-retrieval
pretty_name: 'ASCP: Causal Context Attribution and Probe Benchmark'
size_categories:
- 10K<n<100K
configs:
- config_name: attribution
  data_files: data/attribution.jsonl
- config_name: utilization
  data_files: data/utilization.jsonl
- config_name: probe_benchmark
  data_files: data/probe_benchmark.jsonl
---

# ASCP: Causal Context Attribution and Probe Benchmark

Released artifacts for **[The Laws of Context Allocation: Causal Measurement and
Closed-Loop Orchestration in Generative Search](https://arxiv.org/abs/2608.23252)**.

- 📄 Paper: https://arxiv.org/abs/2608.23252
- 💻 Code: https://github.com/PeiYangLiu/ascp

Retrieval-augmented generation is usually measured with relevance proxies —
BM25, query–document cosine, output overlap — that score how *related* a passage
looks, not whether the generator *used* it. This dataset ships the measurements
behind two claims of the paper:

1. Those proxies look near-perfect only because standard evaluation pools are
   padded with off-query distractors. Under same-query hard negatives that
   entail no answer, they collapse toward chance while causal leave-one-out
   probes are essentially unchanged.
2. Once attribution is measured causally, a fixed inference budget is better
   spent on several narrow contexts than on one wide context.

Each leave-one-out score costs `k+1` teacher-forced forward passes, so the
attribution records here represent tens of thousands of GPU forward passes that
are impractical to regenerate from scratch.

## Configurations

### `probe_benchmark` — 11,520 rows

The evaluation that produces the paper's headline finding. For every case the
necessary answer-bearing set is known by construction, so estimators can be
scored directly against ground truth as the *negative population* is swapped.

| field | meaning |
| --- | --- |
| `regime` | `distractor` (off-query padding), `samequery` (hard negatives), `duplicate`, `mixed` |
| `protocol` | `fixed` (target held constant) or `free` (padding may reshape the response) |
| `k` | context width, one of 2, 3, 5, 8, 12, 24 |
| `m` | size of the designed necessary set (always 2) |
| `loo`, `contextcite`, `sim_gen`, `sim_query`, `bm25_query`, `overlap_gen`, `random` | per-estimator recall / false-positive rate at five thresholds |
| `loo_necessary_scores`, `loo_null_scores` | raw probe scores for necessary and padding documents |
| `groups` | likelihood drop when dropping the necessary set, the padding, all, or half |

Balanced by design: 2,880 rows per regime, 5,760 per protocol, 5,760 per
generator, 5,760 per task.

### `attribution` — 54,100 rows

Per-round causal attribution over the full policy grid: 4 tasks × 2 generators ×
16 scheduling policies. `attribution_raw[i]` is the per-token log-likelihood drop
when `doc_ids[i]` is ablated from that round's context, i.e. the quantity in
Eq. (3) of the paper.

Tasks: ASQA, QAMPARI, ELI5, Recipes (Spanish, cross-cultural).
Policies: `vanilla`, `mmr`, `xquad`, `pm2`, `dpp`, `carriage`, `carriage_narrow`,
`rotate`, `ascp`, and their `-integrate` / width variants.

### `utilization` — 1,200 rows

The subset carrying full generations alongside the matrices the scheduler
consumes: the round-by-round `utilization` matrix, the document–facet matrix
`doc_facet_matrix`, facet importance `facet_importance`, gold answers, and
induced subtopics. This is the frame behind the judge meta-evaluation
(179 portfolios, 858 document-level judgments) in the paper.

## Usage

```python
from datasets import load_dataset

# the estimator benchmark
probe = load_dataset("PeiyangLiu/ascp-context-attribution", "probe_benchmark", split="train")

# does a proxy survive the swap from off-query to same-query padding?
off = probe.filter(lambda r: r["regime"] == "distractor" and r["protocol"] == "fixed")
hard = probe.filter(lambda r: r["regime"] == "samequery" and r["protocol"] == "fixed")

# causal attribution over every policy
attr = load_dataset("PeiyangLiu/ascp-context-attribution", "attribution", split="train")
```

Records join across configurations on `qid` (plus `task`, `generator`, `policy`).

## Reproducing the paper from these files

Averaging estimator AUC over context widths under the `fixed` protocol
reproduces the paper's Figure 1(c), the swap from off-query to same-query
padding:

| estimator | off-query | same-query | Δ |
| --- | --- | --- | --- |
| leave-one-out probe | 0.852 | 0.845 | −0.01 |
| output overlap | 0.952 | 0.831 | −0.14 |
| output–document cosine | 0.980 | 0.766 | −0.23 |
| query–document cosine | 0.999 | 0.564 | −0.52 |
| BM25 | 0.992 | 0.568 | −0.54 |

Every table and figure in the paper is regenerated by the analysis scripts in
[the GitHub repository](https://github.com/PeiYangLiu/ascp), which ships the
aggregated per-query rows and needs no GPU.

## Source data and licensing

Queries and retrieval pools are built on
[ALCE](https://github.com/princeton-nlp/ALCE) (ASQA, QAMPARI, ELI5). **The ALCE
passage text is deliberately not redistributed here**: records reference source
passages by their position in the evaluation pool (`doc_ids`) and, where useful,
by title. Reconstruct the pools from ALCE directly; `scripts/build_pool.py` in
the code repository does this.

The Recipes split follows the cross-cultural recipe setting of Hu et al. Please
observe the upstream licenses for both.

The annotations released here — attribution scores, utilization matrices, facet
matrices, estimator scores — are our own and are licensed CC BY 4.0.

## Limitations

- Probe validation covers 7–8B generators (Qwen2.5-7B, Llama-3.1-8B); the
  paper's scale checks at 14B and 32B are reported from the aggregated rows in
  the code repository rather than here.
- `utilization` covers Qwen2.5-7B on ASQA and QAMPARI only, because it backs the
  judge meta-evaluation rather than the main grid.
- Attribution is a leave-one-out likelihood drop on an already-produced
  response. It measures reliance of *that* generation, not counterfactual
  answer correctness.

## Citation

```bibtex
@article{liu2026laws,
  title   = {The Laws of Context Allocation: Causal Measurement and Closed-Loop
             Orchestration in Generative Search},
  author  = {Liu, Peiyang and Wang, Xi and Liang, Di and Ye, Wei},
  journal = {arXiv preprint arXiv:2608.23252},
  year    = {2026},
  url     = {https://arxiv.org/abs/2608.23252}
}
```