chengruiqu/risk-representation-llm-data

Risk Representation in LLM Activations

Full dataset from a cross-model study testing whether language models have internal representations of risk attitude separable from stimulus encoding.

Report: Interactive results page

Models Tested

• Qwen2.5-7B-Instruct (28 layers, 3584 hidden dim, rev a09a3545)
• Mistral-7B-Instruct-v0.3 (32 layers, 4096 hidden dim, rev c170c708)

Key Finding

The Representation Paradox: Neither model has orderly behavioral risk preferences (choices are position-biased, intransitive, and inconsistent). Yet both contain a detectable, multi-dimensional (~10D) internal representation of risk attitude that is separable from stimulus encoding and — in Qwen's case — causally connected to behavior without disrupting non-risk decisions.

Structure

stimuli/                          # Shared across models (position-randomized)
├── exp0_behavioral.jsonl         # 1,100 binary choices (EV-matched, transitivity, dominance, retest, order-swap)
└── exp1_indifference.jsonl       # 12,000 binary choices (600 lotteries × 20 sure-thing levels)

{qwen,mistral}/
├── model_lock.json               # Pinned model revision
├── responses/
│   ├── exp0_choices.jsonl        # Model's binary choices + logits (Exp 0)
│   ├── exp1_choices.jsonl        # Model's binary choices + logits (Exp 1)
│   └── exp5_steered.jsonl        # Steered choices at 11 coefficients
├── activations/
│   ├── exp0_layer_*.npy          # FP32 hidden states at last prompt token (Exp 0)
│   ├── exp1_layer_*.npy          # FP32 hidden states at last prompt token (Exp 1)
│   └── exp3_layer_*.npy          # Teacher-forced activations (Exp 3)
├── stimuli/
│   └── exp3_controlled.jsonl     # Controlled-answer sequences
└── analysis/
    ├── exp0_behavioral_summary.json
    ├── exp2_probing_results.json      # All layers, permutation tests, dimensionality, transfer, geometry
    ├── exp3_controlled_answer.json
    ├── exp5_steering_dose_response.json
    ├── lottery_results.json           # All 600 lotteries with CE estimates
    ├── lottery_results_filtered.json  # Valid lotteries only (≤2 monotonicity violations)
    └── risk_direction.npy             # Fitted risk direction vector

Experiments

Exp	Name	Stimuli	What it tests
0	Behavioral Baseline	1,100	Orderly preferences: EV-matched risk aversion, transitivity, dominance, test-retest, position bias
1	Indifference Mapping	12,000	Certainty equivalents and risk premiums across 600 lotteries
2	Activation Probing	—	Ridge prediction of RP from activations, with permutation tests, residualization, split-half reliability
3	Controlled-Answer	600	Teacher-forced activations to isolate context signal from response signal
4	Cross-Domain Transfer	—	Train on one lottery type, test on another (stakes, probability, outcome type, grid vs random)
5	Causal Steering	1,650	Inject risk direction at best layer, measure dose-response and specificity

Activation Files

Each .npy file is a float32 array of shape [n_stimuli, hidden_dim] containing the hidden state at the last prompt token before generation. Layers are 1-indexed (e.g., exp1_layer_21.npy = layer 21 of 28 for Qwen).

Reproducing

pip install numpy scipy scikit-learn
python -c "
import numpy as np, json
X = np.load('qwen/activations/exp1_layer_20.npy')  # best layer
with open('qwen/analysis/lottery_results_filtered.json') as f:
    lots = json.load(f)
print(f'Activations: {X.shape}, Valid lotteries: {len(lots)}')
"

Citation

If you use this dataset, please cite:

@misc{qu2026risk,
  title={The Representation Paradox: Risk Attitude in LLM Activations},
  author={Qu, Chengrui},
  year={2026},
  url={https://huggingface.co/datasets/chengruiqu/risk-representation-llm-data}
}