citead/geo-citation-benchmark

GEO Citation Benchmark

Benchmark, predictions and code for Predicting What Answer Engines Cite: Open Decision Models, Rerankers and a Distilled 4B Student on Ten Engines under Controlled Retrieval (Bajemon and Rochet, 2026). The model is citead/geo-citation-ranker-4b.

The task: given a query and five candidate sources, rank the sources by the share of citations each receives in the answer of an answer engine that was given the five sources in its prompt.

Reproduce the paper

Everything below runs offline in a few minutes and needs only numpy (plus matplotlib for the figure).

python code/reproduce_tables.py          # every scorer: rho, 95% CI, paired differences -> results/tables/all_scorers.csv
python code/analyses.py .                # cross-engine agreement, baselines, top-1, per-engine and per-query means, overlap -> results/analyses/extra_analyses.json
python paper/make_tables.py .            # the LaTeX tables of the paper -> paper/tables/, paper/numbers.json
python paper/make_figures.py .           # Figure 1 -> paper/figures/fig_scorers.pdf

Contents

Path	Content
benchmark/passages.jsonl	1145 evaluation passages: 229 GEO-Bench test queries times 5 sources (query, src_idx = position of the source in the engine prompt, text = first 320 words)
benchmark/citation_labels.jsonl	6345 rows, one per (engine, query, source): label_pawc, the position-adjusted word count share of citation in the engine's answer, and label_top
benchmark/folds_query.json	Query to fold (0 to 4), for the students trained on observed citations
benchmark/predictions/	Predictions of the 36 scorer configurations of the paper on every evaluation passage
benchmark/training_groups.jsonl	The 757 training groups (one per GEO-Bench test query outside the evaluation set, five sources each), without labels
benchmark/training_groups_mixed.jsonl	The 757 groups mixing three own sources and two sources of other queries, used in one ablation
code/label_jev.py	Labels the training groups with Jev (OpenRouter), as in the paper
code/train_student.py	Trains a Qwen3-Reranker student with LoRA and a pairwise loss, then predicts the benchmark
code/score_passages.py	Scores any (query, text) JSONL with a trained adapter
results/analyses/extra_analyses.json	Analysis outputs cited by the paper (the agreement between the two teachers is kept from the authors' run, since teacher labels are not redistributed)

Prediction files map "query||src_idx" to a score, except files ending in _by_group, which map "engine::query" to {src_idx: score}.

Engines

Ten configurations: gemini-3.1-flash-lite, gemma-4-31b-it, gpt-5.4, gpt-5.4-mini, gpt-5.5, llama-3.3-70b-instruct, mistral-medium-3.5-128b, nex-n2-pro, gpt-oss-120b, qwen3.5-122b-a10b. Each query is answered by one to ten of them (79 queries by one engine, 77 by all ten). Groups whose five labels are all equal are excluded, which leaves 1220 groups.

Retraining a student

OPENROUTER_API_KEY=... python code/label_jev.py benchmark/training_groups.jsonl jev_labels.jsonl   # about 0.10 USD
python code/train_student.py --labels jev_labels.jsonl --base Qwen/Qwen3-Reranker-4B --out runs/student_4b

Teacher labels are not redistributed. The commands above train a student on Jev labels alone; the students reported in the paper, including GEO Citation Ranker 4B, used the mean of Jev and an LLM judge.

License

Passages come from GEO-Bench (GEO-Optim/geo-bench, CC BY-SA 4.0). Passages, citation labels, folds and predictions are released under CC BY-SA 4.0. Code is released under the Apache 2.0 license.

Citation

@misc{bajemon2026citationranking,
  author = {Bajemon, Elisha and Rochet, Andr{\'e}-Louis},
  title  = {Predicting What Answer Engines Cite: Open Decision Models, Rerankers and a Distilled 4B Student on Ten Engines under Controlled Retrieval},
  year   = {2026},
  note   = {arXiv preprint}
}