ValerianFourel's picture
Add study-design overview, headline findings, and analysis cookbook
490667c verified
---
language:
- en
license: cc-by-4.0
pretty_name: GEODML EMNLP 2026 Reviewer Pack (Condensed)
size_categories:
- n<1K
tags:
- causal-inference
- double-ml
- dml
- llm
- reranker
- search
- geo
- generative-engine-optimization
- interpretability
- probing
- saliency
- ablation
- emnlp
- emnlp-2026
- reproducibility
task_categories:
- text-ranking
- text-classification
---
# GEODML — EMNLP 2026 reviewer pack
A **condensed** companion to the EMNLP 2026 submission **"Causal
Analysis of LLM Search Rerankers via Double/Debiased Machine
Learning."** Designed for fast verification, not full reproduction.
- **This pack** (5.6 MB) — every paper claim as a CSV + every figure
as PDF/PNG + a one-shot `verify.py` claim-checker.
- **Full reproducibility dataset** (1.8 GB) →
[`ValerianFourel/geodml-emnlp-2026`](https://huggingface.co/datasets/ValerianFourel/geodml-emnlp-2026)
— raw LLM rerank outputs, page features, DFS confounders, and the
full DML/probing/saliency/ablation code.
---
## What we did (1-paragraph)
We ask which content features of a web page *cause* an LLM reranker
to keep that URL in its top-K (admission) and to push it toward the
top — separating cause from spurious correlation with confounders
like domain authority. Six features are declared a priori as
treatments (stats density, question-headings, JSON-LD schema,
citation authority, topical competence, freshness) and tested with
Double/Debiased ML controlling for 28 confounders plus the other
five treatments mutually. 65K (keyword × url × condition) rows
across 4 prompt variants × 2 reranker backbones (Llama-3.3-70B,
Qwen-2.5-72B) × 2 SERP backends × 2 candidate-pool sizes feed 216
fitted DML models. Mechanism is verified by three orthogonal
interpretability probes (linear probes on hidden states, gradient
saliency, token ablation).
**Headline.** T5 topical competence is the dominant promoter; T3
JSON-LD schema is a small *demoter* (i.e. it does **not** help —
counter to common SEO advice); T1b stats density has high Qwen
saliency (1.93×) but zero DML coefficient — *attention is not effect*.
For deeper analysis (re-fitting DML on custom slices, inspecting raw
LLM rerank outputs, plotting layer-wise probes), grab the full pack:
[`ValerianFourel/geodml-emnlp-2026`](https://huggingface.co/datasets/ValerianFourel/geodml-emnlp-2026).
---
## Setup (60 seconds)
```bash
# 1. Pull the reviewer pack
huggingface-cli download ValerianFourel/geodml-emnlp-2026-reviewer \
--repo-type=dataset --local-dir ./geodml-reviewer
cd geodml-reviewer
# 2. Install
pip install pandas
# 3. Run the one-shot claim-checker
python verify.py
```
`verify.py` prints every headline number alongside its
paper-text expected value and a tolerance. The last line says either
```
All paper claims VERIFIED against the tables.
```
or it lists which claims fall outside tolerance.
---
## What is what
### `tables/` — condensed paper claims (CSV)
| File | Rows | What it is |
|---|---|---|
| `table2_dml_headline.csv` | 18 | **The paper's Table 2.** Spec B (mutually-controlled) DML coefficients for 6 treatments × 3 outcomes, pooled across all slices. Columns: outcome, treatment, coef, se, ci_lo, ci_hi, p_val, sig_stars, n, coef_promoter_dir. |
| `dml_all_specs_all_slices.csv` | 216 | Every fitted DML model — Spec A (marginal) and Spec B (partial), across 11 slices (POOLED + 10 sub-slices), 6 treatments, 3 outcomes. Columns: spec, slice, treatment, outcome, coef, se, p_val, n, bonferroni_sig. |
| `probing_peaks_per_variant.csv` | ~56 | Peak linear-probe ROC AUC per (variant, treatment, pooling). Columns: variant, treatment, pooling, peak_layer, peak_roc_auc, layer_0_roc_auc. |
| `admission_probe_headline.csv` | 8 | The admission pre-commitment story — L0, peak, and final-layer ROC AUC of the `Y1_admission_inctx` probe across 4 variants × 2 poolings. |
| `saliency_summary.csv` | 8 | Llama-3.3-70B vs Qwen-2.5-72B gradient-saliency ratios on 4 treatments. Numbers >1 mean the model attends more to treatment tokens than to surrounding tokens. |
| `ablation_summary.csv` | 24 | Mean / median / std Δrank when treatment tokens are ablated, per (treatment, prompt, frame). |
| `freshness_leakage_diagnostic.md` | — | Year-token leakage diagnostic for T6 freshness — included as a methodological caveat. |
### `figures/` — every PDF + PNG the paper references
Each figure is included as both PDF (for camera-ready) and PNG (for
quick HTML preview).
| File | What it is |
|---|---|
| `dml_estimation_framework.{pdf,png}` | The DML estimation framework diagram (data flow). |
| `fig01_admission_forest.{pdf,png}` | Forest plot of Spec B coefficients on Y1 admission. |
| `fig02_rank_delta_forest.{pdf,png}` | Forest plot on Y2 = rank_delta. |
| `fig03_post_rank_forest.{pdf,png}` | Forest plot on Y3 = rank_post. |
| `fig04_three_outcome_grid.{pdf,png}` | All three outcomes side-by-side. |
| `fig05_admission_vs_rank_scatter.{pdf,png}` | Y1 vs Y2 per treatment — admission and promotion magnitude correlate. |
| `fig06_marginal_vs_partial.{pdf,png}` | Spec A vs Spec B (does mutual control change the story?). |
| `fig07_variant_grid_admission.{pdf,png}` | Y1 across the 4 prompt variants. |
| `fig08_variant_grid_rank.{pdf,png}` | Y2 across the 4 prompt variants. |
| `fig09_compare_rag_vs_snippet.{pdf,png}` | RAG vs snippet-only prompt comparison. |
| `fig10_compare_ddg_vs_searxng.{pdf,png}` | DuckDuckGo vs SearXNG SERP-backend comparison. |
| `fig11_compare_llama_vs_qwen.{pdf,png}` | Llama vs Qwen reranker comparison. |
| `fig12_compare_pool20_vs_pool50.{pdf,png}` | Top-20 vs top-50 candidate-pool comparison. |
| `fig13_admission_detail.{pdf,png}` | Detail view of admission coefficients with CIs. |
| `fig14_robust_survivors.{pdf,png}` | The robust-winners frame (slice-stable treatments). |
| `fig_probing_layerwise.{pdf,png}` | Layer-wise probe ROC AUC per treatment. |
| `fig_probing_pooling.{pdf,png}` | Probe ROC AUC by token-pooling strategy. |
| `fig_admission_pooled.{pdf,png}` | Admission probe across layers, pooled across variants. |
| `fig_admission_variants.{pdf,png}` | Admission probe across layers, split by variant. |
| `fig_saliency.{pdf,png}` | Diverging-bar saliency-ratio chart (Llama vs Qwen). |
### `verify.py` — one-shot claim checker
Reads every CSV in `tables/` and asserts each headline number against
its paper-text value within tolerance. Exits 0 on PASS, 1 on FAIL.
Run after every pull to confirm the data matches what the paper says.
---
## Headline numbers — Table 2 (Spec B POOLED)
DML coefficients with `has_llms_txt` as a confounder (not a treatment),
across 6 canonical content treatments.
| Treatment | Y1 selected | Y2 Δrank | Y3 rank_post |
|---|---|---|---|
| T5 topical competence | **+0.037***\* | **−0.530***\* | **−0.299***\* |
| T2a Q-headings | **+0.016***\* | **+0.136***\* | **−0.041*** |
| T3 schema (JSON-LD) | **−0.014***\* | **−0.051*** | **+0.095***\* |
| T6 freshness | **−0.005***\* | **−0.061***\* | +0.005 |
| T4 citation authority | +0.001 | **−0.023*** | **−0.015*** |
| T1b stats density | −0.000 | −0.003 | −0.002 |
**Sign convention:** `+Y1`, `+Y2`, `−Y3` ⇒ URL favoured by the LLM.
**Stars:** `***` < .001, `**` < .01, `*` < .05.
---
## Admission pre-commitment probe
Linear probe for the binary admission outcome on the LLM's *internal
activations* at each layer. Averaged across the 4 prompt variants,
mean pooling over context tokens.
| Quantity | Value | Interpretation |
|---|---|---|
| Layer 0 ROC AUC | 0.671 | Embedding alone — weak admission signal. |
| Peak ROC AUC | 0.862 | At layer 60 (~75% network depth). |
| L0 → peak gain | +0.191 | Genuine compositional integration. |
---
## Saliency (gradient attribution)
Saliency ratio = mean |∇| on treatment tokens / mean |∇| on
surrounding-context tokens. `>1` means the model attends more to the
treatment than to context; `<1` means it ignores it.
| Treatment | Llama-3.3-70B | Qwen-2.5-72B | DML direction | Reading |
|---|---|---|---|---|
| T1b stats density | 0.89× | **1.93×** | null | Qwen looks at stats but DML says it doesn't move ranks. |
| T2a Q-headings | 1.05× | 0.90× | promoter | Comparable attention; works for both. |
| T3 schema | **0.19×** | **0.40×** | demoter | Both models ignore JSON-LD — DML demoter is consistent. |
---
## Citation
```bibtex
@inproceedings{fourel2026geodml,
title = {Causal Analysis of LLM Search Rerankers via Double/Debiased Machine Learning},
author = {Fourel, Valerian},
year = {2026},
booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing}
}
```
## License
CC BY 4.0 — free reuse with attribution.
## Contact
valerian.fourel@gmail.com