BowenSun1998/Benchmarking_Risk_Attitudes_of_LLMs

Risk Attitude in LLMs: A Cross-Task Behavioral Dataset

Dataset accompanying the NeurIPS 2026 paper (Track on Evaluations and Datasets):

Benchmarking Risk Attitudes of LLMs Bowen Sun*, Rui Min*, Xianyao Li, Yuxi Wang, Yang Ye, Qi R. Wang, Jing Du (* equal contribution)

Code: https://github.com/Bowens1998/LLMs-Risk-Attitude_NeurIPS-2026 (MIT). It contains the task environments, the LLM runner, the human-study interface, and the pipeline that regenerates every table and figure of the paper from this dataset.

This dataset contains LLM response data and human baseline data from three structured decision tasks designed to measure risk attitude and the behavioral reliability of contextual beliefs and risk decisions.

---

Tasks

Abbreviation	Full Name	Description
CTD	Clinical Triage Decision	Model acts as an emergency triage clinician assigning ESI severity scores
DNC	Drone Navigation Control	Model navigates a drone through a hazardous grid environment
FIP	Financial Investment Portfolio	Model allocates assets across L/M/H-risk instruments under market conditions

Key Constructs

Symbol	Name	Description
$B_C$	Contextual Belief	Model's perceived danger level (0–100) given situational context
$R_D$	Risk Decision	Categorical action chosen in response to that belief (1–5, 1 = most cautious)

Two phases

• Reliability phase (*_reliability_N90/): three frozen scenarios per task, each replayed 30 times (90 trials per model per task). The environment is held fixed, so between-trial variation reflects only the model's own $B_C \rightarrow R_D$ mapping.
• Attitude estimation phase (*_main_N100/): 100 trials per model per task, used to fit the ordered logistic models.

---

Repository Structure

├── FrontierModel_main_N100/        — Attitude phase, frontier models (N=100 per task)
│   ├── CTD/                        — 6 files, one per model
│   ├── DNC/                        — same 6 models, DNC_ prefix
│   └── FIP/                        — same 6 models, FIP_ prefix
│
├── EarlierModel_main_N100/         — Attitude phase, earlier-generation models (N=100)
│   ├── CTD/   ├── DNC/   └── FIP/  — 6 files each
│
├── FrontierModel_reliability_N90/  — Reliability phase, flat records
│   ├── Reliability_CTD_merged.json —   892 records
│   ├── Reliability_DNC_merged.json —   849 records
│   ├── Reliability_FIP_merged.json —   900 records
│   ├── Reliability_ALL_merged.json — 2,641 records (DNC + CTD + FIP, in that order)
│   ├── base_instruct_gemma/             — per-run source files, gemma-4-E4B base and instruct
│   └── base_instruct_qwen/              — per-run source files, Qwen3 30B A3B base and instruct
│
├── EarlierModel_reliability_N90/   — Reliability phase, earlier-generation models
│   ├── EarlierModels_Reliability_CTD_merged.json —   540 records
│   ├── EarlierModels_Reliability_DNC_merged.json —   540 records
│   ├── EarlierModels_Reliability_FIP_merged.json —   540 records
│   └── EarlierModels_Reliability_ALL_merged.json — 1,620 records
│
├── RiskAttitude_HumanData/         — Raw task records from 100 human participants
│
├── README.md
└── croissant.json                  — machine-readable metadata (MLCommons Croissant)

---

Models

Frontier models (6): main analysis

Model	model_key
Claude Sonnet 4.5	ClaudeSonnet4.5
DeepSeek V3.2	DeepSeekV3.2
GPT 5.2	GPT5.2
Qwen3 Max	Qwen3Max
Gemini 3 Pro	Gemini3Pro
Grok 4	Grok4

Open-weight base/instruct ablation (4): reliability phase only

Matched base and post-trained checkpoints from two families, used to test whether post-training narrows within-condition $B_C$ dispersion.

Model	model_key	Variant
gemma-4-E4B	Gemma4E4b	base
gemma-4-E4B-it	Gemma4E4bIt	instruction-tuned
Qwen3 30B A3B Base	Qwen330bA3bBase	base
Qwen3 30B A3B Instruct 2507	Qwen330bA3bInstruct2507	instruction-tuned

These four appear only in FrontierModel_reliability_N90/. They are not part of the attitude-estimation phase and have no *_main_N100/ counterpart.

Earlier-generation models (6)

Model	model_key
Claude Sonnet 4	Sonnet4
DeepSeek V3	DeepSeekV3
GPT-4.1	GPT4.1
Gemini 2.5 Pro	Gemini2.5Pro
Qwen 3	Qwen3
Grok 3	Grok3

---

Data Formats

Reliability data (*_reliability_N90/*_merged.json): flat record format

Each file is a JSON array with one record per trial.

[
  {
    "experiment":        "CTD",
    "model":             "Grok 4",
    "model_key":         "Grok4",
    "level":             1,
    "trial_index":       0,
    "level_trial_index": 0,
    "contextual_belief": 95,
    "risk_decision":     2
  }
]

Field	Type	Description
experiment	string	Task: "CTD", "DNC", or "FIP"
model	string	Full model display name
model_key	string	Compact identifier
level	int	Frozen scenario danger level: 1 / 2 / 3
trial_index	int	Zero-based trial index within this model's cell for the task
level_trial_index	int	Zero-based trial index within the level
contextual_belief	number	$B_C$: reported danger estimate (0–100)
risk_decision	int	$R_D$: discrete risk action (1–5)

Reliability_ALL_merged.json is the exact concatenation of the DNC, CTD, and FIP files, in that order. To recover a single model–task cell, filter on model_key and experiment.

Observed trial counts

The nominal design is 30 repetitions × 3 conditions = 90 trials per model–task cell. Seven cells departed from nominal because trials were lost to API errors during collection.

Model	Task	C1 / C2 / C3	Cell total
Grok 4	DNC	30 / 29 / 30	89
gemma-4-E4B	CTD	24 / 28 / 30	82
gemma-4-E4B	DNC	9 / 19 / 19	47
Qwen3 30B A3B Base	DNC	30 / 23 / 30	83

Every other model–task cell contains 30 / 30 / 30 records. The gemma-4-E4B DNC cell is sparse enough that condition-level statistics from it should be read with care.

Per-run ablation sources (FrontierModel_reliability_N90/base_instruct_*/)

One directory per model per task, holding the run as produced by the collection pipeline:

• <TASK>_<model>_N<n>.json: trial records, identical in schema to the merged files
• <TASK>_<model>_N<n>_meta.json: provenance sidecar with the seed, model identifier, decoding settings, manipulation-check outcome, and raw trial payloads

Every record in these files is also present in the four Reliability_*_merged.json files. They are included for provenance; the merged files are the analysis inputs.

Main data (*_main_N100/): raw trial format

Each file holds the trial-level records for one model on one task (100 trials):

{
  "experiment": "CTD",
  "model": "Grok 4",
  "n_trials": 100,
  "trials": [ { "...": "raw task record of one trial" } ]
}

Frontier-model files contain the raw task records. $B_C$ and $R_D$ are computed from them as described below; experiments/scoring/rd_discretization.py in the code repository does this. Earlier-model files use legacy field names (session instead of experiment, plus controls, ui or params) and also carry a precomputed risk_decision.

$B_C$ extraction per task

Task	Source field
DNC	post-trial danger rating (context_belief; contextual_belief in earlier-model files)
FIP	report.contextual.risk
CTD	steps[-1].BC_updates[-1].ctx

$R_D$ per task

Task	Method
DNC	Strategy Index SI = logit(VCR / (VCR + HPR + ε)), with the ratio clipped to [ε, 1 − ε] and ε = 10⁻³, on drift steps (all steps if none); nearest of five frozen benchmark-v1 centers
FIP	high-risk share H / (L + M + H) of report.alloc; nearest of five frozen benchmark-v1 centers
CTD	finalized.ESI used directly (already 1–5)

VCR is the vertical correction ratio (share of UP/DOWN actions) and HPR the horizontal progression ratio (share of RIGHT actions). The frozen centers are given in experiments/config/frozen_ordinalization_map.json of the code repository.

---

Human Baseline Data

RiskAttitude_HumanData/ contains the complete raw task records for 100 unique participants who completed the same three tasks. Source-tab labels map to the paper tasks as follows: TPB = CTD, FIP = FIP, DSB = DNC. Participants are identified only by random study identifiers that cannot be linked back to individuals.

File	Contents
Human_TPB_N100_full.csv / .jsonl	Clinical triage records, 100 subjects
Human_FIP_N100_full.csv / .jsonl	Financial investment records, 100 subjects
Human_DSB_N100_full.csv / .jsonl	Drone navigation records, 100 subjects
Human_Cohort_First100Unique.csv	Cohort order and subject IDs
Human_Data_Integrity_QC.csv	Per-record parse, length, prefix, and SHA-256 checks
Human_Data_N100_QC.xlsx	Metadata-only review workbook (no raw long JSON cells)
NeurIPS2026_RiskAttitude_HumanData.xlsx	Workbook of all submissions; long payloads are truncated by Excel's cell limit, so use the CSV/JSONL files for complete records
manifest.json	Per-task counts, length summaries, and per-file hashes
README.md	Cohort rule, storage rationale, and integrity checks

Raw payloads are distributed as CSV and JSONL rather than XLSX: some Data (JSON) values exceed Excel's 32,767-character per-cell limit, so an XLSX export would silently truncate them. The CSV files preserve the exact source strings; the JSONL files additionally expose each payload as a parsed JSON object. See RiskAttitude_HumanData/README.md for the full cohort definition and QC results.

The human study was approved by an Institutional Review Board (protocol ET00049588). Participants were recruited via Prolific and gave informed consent.

---

Citation

@inproceedings{sun2026benchmarking,
  title     = {Benchmarking Risk Attitudes of {LLMs}},
  author    = {Sun, Bowen and Min, Rui and Li, Xianyao and Wang, Yuxi and Ye, Yang and Wang, Qi R. and Du, Jing},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Track on Evaluations and Datasets},
  year      = {2026}
}

---

License

Released under Creative Commons Attribution 4.0 International (CC-BY 4.0).

You are free to share and adapt the data for any purpose, provided you give appropriate credit to the original authors.