inferencelabs/TruthTensor

# TruthTensor: Measuring instruction-following under drift

Large language models are usually evaluated as if the world were static. Real deployments aren’t: **instructions persist while environments drift**; probabilities shift, narratives evolve, and agents must decide whether to update, resist, or overreact.

TruthTensor evaluates **instruction divergence**: how far a model shifts away from its prescribed decision procedure as the environment changes.

Paper: **TruthTensor: Evaluating LLMs Through Human Imitation on Prediction Markets Under Drift and Holistic Reasoning** ([arXiv:2601.13545](https://arxiv.org/abs/2601.13545)).

- `UserFinetuning.parquet` — user-defined finetuned agent decisions; Public Dataset Export: 2026-01-09 to 2026-01-10.
- `Experiment_InstructionLocked.parquet` — instruction-locked experiment execution logs.

## Citation

```bibtex
@misc{shahabi2026truthtensor,
title = {TruthTensor: Evaluating LLMs through Human Imitation on Prediction Market under Drift and Holistic Reasoning},
author = {Shirin Shahabi and Spencer Graham and Haruna Isah},
year = {2026},
eprint = {2601.13545},
archivePrefix= {arXiv},
primaryClass = {cs.AI},
url = {https://arxiv.org/abs/2601.13545}
}
```

## Contact

For the entire public dataset available on [TruthTensor.com](https://truthtensor.com), contact the Inference Labs team at [e-mail].