---
license: cc-by-4.0
task_categories:
  - text-classification
language:
  - en
tags:
  - bias
  - fairness
  - llm-evaluation
  - personalization
  - demographic-cues
pretty_name: Demographic Cues in LLM Personalization and Bias
size_categories:
  - 10M<n<100M
---

# Different Demographic Cues Yield Inconsistent Conclusions About LLM Personalization and Bias

Model responses and derived tables for the EMNLP 2026 paper. Analysis code:
**https://github.com/manueltonneau/llm-demographic-cues**

> Manuel Tonneau, Neil K. R. Sehgal, Niyati Malhotra, Sharif Kazemi, Victor
> Orozco-Olvera, Ana María Muñoz Boudet, Lakshmi Subramanian, Samuel P.
> Fraiberger, Sharath Chandra Guntuku, Valentin Hofmann. *Different Demographic
> Cues Yield Inconsistent Conclusions About LLM Personalization and Bias.*
> [arXiv:2601.18486](https://arxiv.org/abs/2601.18486). To appear at EMNLP 2026.

The paper asks whether the demographic cues commonly used to probe LLMs, that is
explicit identity statements, names, dialog history and dialect, are
interchangeable operationalizations of the same underlying identity-conditioned
behavior. They are not: cue choice changes both the magnitude and the direction
of the estimated effect.

## What is here

Three models (LLaMA-3.1 8B, OLMo-2 7B, GPT-5.2) answering three advice tasks
(medical, legal, salary) under eight demographic cues, plus the aggregates the
paper's tables and figures are computed from.

```
decoder_model_responses_cleaned/<task>_<cue>_constrained_<model>_seed_<0|1|2>.csv
decoder_model_responses_race_pred/<task>_<cue>_constrained_<model>_seed_0.csv
plot_data/                       # per-cue outcome ratios behind Figures 3 and 11
fk_cache/<task>_<cue>.parquet    # Flesch-Kincaid grade per prompt
results_all_models/<model>/<task>/master.parquet
```

with

- `<task>`  ∈ `medical_advice`, `legal_advice`, `salary_rec`
- `<cue>`   ∈ `neutral`, `explicit`, `dialect`, `name_specific_{rosenman,hayes_elder,tzioumis,an}`, `convo_prefix`, `convo_prefix_prism`
- `<model>` ∈ `llama3.1`, `olmo2`, `gpt52` (seeds 1–2 for `llama3.1` / `olmo2`)

`decoder_model_responses_cleaned/` holds the task answers; `*_race_pred/` holds
what the model inferred the user's race to be for the same prompts.

## Usage

```bash
git clone https://github.com/manueltonneau/llm-demographic-cues
huggingface-cli download manueltonneau/llm-demographic-cues --repo-type dataset --local-dir data
CUES_DATA_DIR=$PWD/data python llm-demographic-cues/scripts/analysis/race_pred_recall.py
```

`fk_cache/` and `asl_cache/` can stay here; the scripts fall back to
`CUES_DATA_DIR` when they are not present in the clone. `results_all_models/`
holds the master tables that `scripts/appendix/task2_regress.py` reads.

## What is deliberately not here

**Prompt text.** The prompts are not released. The dialog-history cues
(`convo_prefix`, `convo_prefix_prism`) prepend conversations taken verbatim from
two third-party human-subjects corpora: the Community Alignment Dataset
([Zhang et al., 2025](https://arxiv.org/abs/2507.09650)) and PRISM (Kirk et al.,
NeurIPS 2024), which are distributed under their own terms and are not ours to
redistribute. Obtain them upstream and
rebuild the prompts with the procedure in Appendix A of the paper.

**Annotator attributes.** `annotator_id`, `age`, `political`, `assigned_lang` and
`source_conversation_id` are dropped from every file. They originate with the
upstream corpora, and no analysis in the paper uses them. `convo_concat_id`, a
conversation-cluster key with no personal content, is kept because the master-table
build groups on it.

**The two master tables.** `scripts/analysis/replicate_*.py` read
`masters/medical_master.parquet` and `masters/legal_salary_master.parquet`, which
are not shipped here because they are derived. Rebuild them from the files in
this release:

```bash
CUES_DATA_DIR=/path/to/this/data python scripts/build/build_masters.py
```

`fk_cache/` and `asl_cache/` are included precisely so this works without the
prompt text: `build_masters.py` joins the responses here against those two
caches and the feature table committed in the code repo.

## Contents warning

The salary and legal tasks elicit model behavior that differs by inferred race
and gender; the dialect cue consists of African American English renderings of
neutral prompts, generated with an LLM. The data documents model bias and is
intended for research on measuring it.

## License

The files here are model outputs and derived aggregates produced by us, released
under **CC BY 4.0**. The analysis code is MIT-licensed and lives in the
[code repository](https://github.com/manueltonneau/llm-demographic-cues).

The prompts these outputs respond to were built in part from the Community
Alignment Dataset (CC BY 4.0) and PRISM, which carry their own terms. No text
from either corpus is redistributed here. If you rebuild the prompts, obtain
those datasets from their maintainers and honour their licenses and citation
requirements in addition to ours.

## Citation

```bibtex
@article{tonneau2026cues,
  title     = {Different Demographic Cues Yield Inconsistent Conclusions About {LLM} Personalization and Bias},
  author    = {Tonneau, Manuel and Sehgal, Neil K. R. and Malhotra, Niyati and
               Kazemi, Sharif and Orozco-Olvera, Victor and Mu{\~n}oz Boudet, Ana Mar{\'i}a and
               Subramanian, Lakshmi and Fraiberger, Samuel P. and
               Guntuku, Sharath Chandra and Hofmann, Valentin},
  journal   = {arXiv preprint arXiv:2601.18486},
  note      = {To appear at EMNLP 2026},
  year      = {2026}
}
```