| --- |
| license: cc-by-4.0 |
| language: |
| - en |
| pretty_name: Persona-Aware Cognitive-Bias Benchmark |
| size_categories: |
| - n<1K |
| task_categories: |
| - text-classification |
| - question-answering |
| tags: |
| - cognitive-bias |
| - llm-evaluation |
| - persona |
| - framing |
| - behavioral-economics |
| - decision-making |
| configs: |
| - config_name: bias_benchmark |
| default: true |
| data_files: |
| - split: test |
| path: bias_benchmark.jsonl |
| - config_name: personas |
| data_files: |
| - split: test |
| path: personas.jsonl |
| - config_name: financial_tasks |
| data_files: |
| - split: test |
| path: financial_tasks.jsonl |
| --- |
| |
| # Persona-Aware Cognitive-Bias Benchmark |
|
|
| Data resources from the LREC 2026 paper **"Persona-Aware Evaluation of Cognitive |
| Bias in LLMs: From Benchmark to Applied Decision-Making"** (Yoshikawa, Takayama, |
| Yamazaki). The suite couples a 12-category cognitive-bias benchmark with a |
| factorial persona space and an applied financial-framing task, to study how |
| persona conditioning modulates bias-consistent responding in LLMs. |
|
|
| > **Reconstruction notice.** The original construction data was lost. The 120 |
| > bias items here were **re-authored to the paper's specification** (same |
| > construct, format, invariant, and controls) and are **not** the identical |
| > strings used in the paper's experiments. Absolute Yes-rates a model produces |
| > may therefore differ from the paper's tables. The persona space and financial |
| > tasks are regenerated deterministically and match the paper exactly. |
|
|
| ## Subsets |
|
|
| ### `bias_benchmark` (120 items) |
| 12 cognitive-bias categories × 10 items. Each item is a short vignette ending in |
| a Yes/No question. **Invariant: the bias-consistent answer is always `"Yes"`.** |
| |
| | field | type | description | |
| |---|---|---| |
| | `id` | string | e.g. `availability_01` | |
| | `category` | string | one of 12 categories | |
| | `scenario` | string | vignette + Yes/No question | |
| | `answer` | string | gold label, always `"Yes"` (bias-consistent) | |
| | `surface_form` | string | `positive` / `negative` wording variant | |
| | `bias_consistent_reason` | string | why Yes is the bias-consistent response | |
|
|
| Categories: anchoring, framing, confirmation, availability, loss_aversion, |
| status_quo, sunk_cost, social_proof, hindsight, representativeness, |
| overconfidence, endowment. |
|
|
| ### `personas` (162 personas) |
| Factorial Gender(2) × Age(3) × Politics(3) × Income(3) × Education(3) = 162, each |
| with a marginal-product weight `w(p)` for weighted aggregation. |
|
|
| | field | type | description | |
| |---|---|---| |
| | `id` | string | `persona_000` … `persona_161` | |
| | `gender` / `age` / `politics` / `income` / `education` | — | attribute values | |
| | `weight` | float | `P(g)P(a)P(pol)P(inc)P(edu)`; the 162 weights sum to 1 | |
| | `persona_text` | string | natural-language preface ("a 40-year-old male with …") | |
|
|
| Country (8) and Occupation (8) are **uniform, non-stratified** randomization |
| fields (not part of the 162-key); their value lists ship in the repo's |
| `persona_fields.json`. |
|
|
| ### `financial_tasks` (100 matched pairs) |
| Matched gain/loss framing pairs. Fixed decision skeleton A=Certain, B=Risky; each |
| pair shares numeric parameters and flips only the wording. Amounts/probabilities |
| are filtered so EV(certain) ≈ EV(risky) (≤3% relative gap). |
| |
| | field | type | description | |
| |---|---|---| |
| | `id` | string | `fin_000` … `fin_099` | |
| | `wealth` / `prob` / `certain_amount` / `risky_amount` | num | parameters | |
| | `ev_certain` / `ev_risky` | float | expected values (matched) | |
| | `gain_frame` / `loss_frame` | string | rendered prompts | |
| |
| ## Usage |
| |
| ```python |
| from datasets import load_dataset |
|
|
| bias = load_dataset("katsumasay/cog-bias-bench", "bias_benchmark", split="test") |
| personas = load_dataset("katsumasay/cog-bias-bench", "personas", split="test") |
| fin = load_dataset("katsumasay/cog-bias-bench", "financial_tasks", split="test") |
| ``` |
| |
| **Metrics (from the paper):** Yes-rate (bias-consistent rate; weighted by `w(p)` |
| under persona aggregation), reversal rate, and confidence mean. Persona weights |
| are applied at **aggregation** time only; all 162 personas are sampled uniformly |
| when running a model. Output schemas: bias `{"answer":"Yes/No","reason":"..."}`; |
| financial `{"choice":"A/B","confidence":0.0-1.0,"rationale":"..."}`. |
| |
| ## Intended use & limitations |
| |
| - **Intended use:** evaluating cognitive-bias susceptibility and persona |
| sensitivity of LLMs. Research only. |
| - **Not** a measure of human bias, nor a fairness/social-bias dataset (cf. BBQ, |
| StereoSet) — it targets behavioral–cognitive biases. |
| - Items are English and necessarily stylized; absolute scores are model- and |
| prompt-dependent. Treat small per-category samples (n=10) as indicative. |
| |
| ## Ethical considerations |
| |
| Personas include sensitive attributes (political orientation, income, education). |
| They are coarse, synthetic constructs for controlled evaluation and must not be |
| used to profile, stereotype, or make decisions about real individuals. |
| |
| ## License & citation |
| |
| Data: **CC BY 4.0**. (Repository code: Apache-2.0.) |
| |
| ```bibtex |
| @inproceedings{yoshikawa2026persona, |
| title = {Persona-Aware Evaluation of Cognitive Bias in {LLM}s: From Benchmark to Applied Decision-Making}, |
| author = {Yoshikawa, Katsumasa and Takayama, Junya and Yamazaki, Takato}, |
| booktitle = {Proceedings of the 2026 Conference on Language Resources and Evaluation (LREC 2026)}, |
| year = {2026} |
| } |
| ``` |
| |
| Code & reproduction scripts: https://github.com/Wildkatze/cog_bias_bench |
| |
