| --- |
| license: cc-by-4.0 |
| pretty_name: Per-Sentence Future-Behavior Probabilities |
| tags: |
| - interpretability |
| - reasoning-models |
| - steering |
| - probing |
| --- |
| |
| # Per-Sentence Future-Behavior Probabilities |
|
|
| Ground-truth future-behavior probability dynamics from the Behavior Distribution Analysis, together with the gathered activations, trained probes, and difference-in-means steering vectors. |
|
|
| This dataset is part of the data release for the paper [**Predicting Future Behaviors in Reasoning Models Enables Better Steering**](https://arxiv.org/abs/2606.11172). |
|
|
| You can [view these behavior distributions in the interactive viewer](https://behavior-distributions-demo.github.io). |
| Download an `*_outputs.json` file from the table below, and upload it to the viewer using the `Load full data...` button. |
|
|
| The **Browse** link opens the full folder for that model and dataset; the **Download** link downloads the ground-truth test-set `outputs.json` directly. |
|
|
| ## Data |
|
|
| | Model | Dataset | Browse | Download | |
| |---|---|---|---| |
| | DeepSeek-R1-Distill-Llama-8B | elephant_aita | [browse](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/tree/main/DeepSeek-R1-Distill-Llama-8B/elephant_aita) | [outputs.json](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/resolve/main/DeepSeek-R1-Distill-Llama-8B/elephant_aita/n100_nbase10_nsamp30_l8192_s-42_t1.0_outputs.json?download=true) | |
| | DeepSeek-R1-Distill-Llama-8B | myopic_reward | [browse](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/tree/main/DeepSeek-R1-Distill-Llama-8B/myopic_reward) | [outputs.json](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/resolve/main/DeepSeek-R1-Distill-Llama-8B/myopic_reward/n100_nbase10_nsamp30_l8192_s-42_t1.0_outputs.json?download=true) | |
| | DeepSeek-R1-Distill-Llama-8B | sep | [browse](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/tree/main/DeepSeek-R1-Distill-Llama-8B/sep) | [outputs.json](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/resolve/main/DeepSeek-R1-Distill-Llama-8B/sep/n100_nbase10_nsamp30_l8192_s-42_t1.0_outputs.json?download=true) | |
| | DeepSeek-R1-Distill-Llama-8B | sorrybench | [browse](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/tree/main/DeepSeek-R1-Distill-Llama-8B/sorrybench) | [outputs.json](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/resolve/main/DeepSeek-R1-Distill-Llama-8B/sorrybench/n100_nbase10_nsamp30_l8192_s-42_t1.0_outputs.json?download=true) | |
| | DeepSeek-R1-Distill-Llama-8B | survival_instinct | [browse](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/tree/main/DeepSeek-R1-Distill-Llama-8B/survival_instinct) | [outputs.json](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/resolve/main/DeepSeek-R1-Distill-Llama-8B/survival_instinct/n100_nbase10_nsamp30_l8192_s-42_t1.0_outputs.json?download=true) | |
| | DeepSeek-R1-Distill-Llama-8B | wealth_seeking | [browse](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/tree/main/DeepSeek-R1-Distill-Llama-8B/wealth_seeking) | [outputs.json](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/resolve/main/DeepSeek-R1-Distill-Llama-8B/wealth_seeking/n100_nbase10_nsamp30_l8192_s-42_t1.0_outputs.json?download=true) | |
| | QwQ-32B | elephant_aita | [browse](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/tree/main/QwQ-32B/elephant_aita) | [outputs.json](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/resolve/main/QwQ-32B/elephant_aita/n100_nbase10_nsamp10_l4096_s-42_t1.0_outputs.json?download=true) | |
| | QwQ-32B | myopic_reward | [browse](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/tree/main/QwQ-32B/myopic_reward) | [outputs.json](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/resolve/main/QwQ-32B/myopic_reward/n100_nbase10_nsamp30_l8192_s-42_t1.0_outputs.json?download=true) | |
| | QwQ-32B | sep | [browse](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/tree/main/QwQ-32B/sep) | [outputs.json](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/resolve/main/QwQ-32B/sep/n100_nbase10_nsamp10_l4096_s-42_t1.0_outputs.json?download=true) | |
| | QwQ-32B | sorrybench | [browse](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/tree/main/QwQ-32B/sorrybench) | [outputs.json](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/resolve/main/QwQ-32B/sorrybench/n100_nbase10_nsamp10_l4096_s-42_t1.0_outputs.json?download=true) | |
| | QwQ-32B | survival_instinct | [browse](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/tree/main/QwQ-32B/survival_instinct) | [outputs.json](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/resolve/main/QwQ-32B/survival_instinct/n100_nbase10_nsamp30_l8192_s-42_t1.0_outputs.json?download=true) | |
| | QwQ-32B | sycophancy | [browse](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/tree/main/QwQ-32B/sycophancy) | [outputs.json](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/resolve/main/QwQ-32B/sycophancy/n100_nbase10_nsamp30_l8192_s-42_t1.0_outputs.json?download=true) | |
| | QwQ-32B | wealth_seeking | [browse](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/tree/main/QwQ-32B/wealth_seeking) | [outputs.json](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/resolve/main/QwQ-32B/wealth_seeking/n100_nbase10_nsamp30_l8192_s-42_t1.0_outputs.json?download=true) | |
| | Qwen3-14B | elephant_aita | [browse](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/tree/main/Qwen3-14B/elephant_aita) | [outputs.json](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/resolve/main/Qwen3-14B/elephant_aita/n100_nbase10_nsamp10_l4096_s-42_t1.0_outputs.json?download=true) | |
| | Qwen3-14B | myopic_reward | [browse](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/tree/main/Qwen3-14B/myopic_reward) | [outputs.json](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/resolve/main/Qwen3-14B/myopic_reward/n100_nbase10_nsamp10_l4096_s-42_t1.0_outputs.json?download=true) | |
| | Qwen3-14B | sep | [browse](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/tree/main/Qwen3-14B/sep) | [outputs.json](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/resolve/main/Qwen3-14B/sep/n100_nbase10_nsamp10_l4096_s-42_t1.0_outputs.json?download=true) | |
| | Qwen3-14B | sorrybench | [browse](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/tree/main/Qwen3-14B/sorrybench) | [outputs.json](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/resolve/main/Qwen3-14B/sorrybench/n100_nbase10_nsamp10_l4096_s-42_t1.0_outputs.json?download=true) | |
| | Qwen3-14B | survival_instinct | [browse](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/tree/main/Qwen3-14B/survival_instinct) | [outputs.json](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/resolve/main/Qwen3-14B/survival_instinct/n100_nbase10_nsamp10_l4096_s-42_t1.0_outputs.json?download=true) | |
| | Qwen3-14B | wealth_seeking | [browse](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/tree/main/Qwen3-14B/wealth_seeking) | [outputs.json](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/resolve/main/Qwen3-14B/wealth_seeking/n100_nbase10_nsamp10_l4096_s-42_t1.0_outputs.json?download=true) | |
| | gpt-oss-20b | elephant_aita | [browse](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/tree/main/gpt-oss-20b/elephant_aita) | [outputs.json](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/resolve/main/gpt-oss-20b/elephant_aita/n100_nbase10_nsamp10_l4096_s-42_t1.0_outputs.json?download=true) | |
| | gpt-oss-20b | myopic_reward | [browse](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/tree/main/gpt-oss-20b/myopic_reward) | [outputs.json](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/resolve/main/gpt-oss-20b/myopic_reward/n100_nbase10_nsamp10_l4096_s-42_t1.0_outputs.json?download=true) | |
| | gpt-oss-20b | sep | [browse](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/tree/main/gpt-oss-20b/sep) | [outputs.json](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/resolve/main/gpt-oss-20b/sep/n100_nbase10_nsamp10_l4096_s-42_t1.0_outputs.json?download=true) | |
| | gpt-oss-20b | sorrybench | [browse](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/tree/main/gpt-oss-20b/sorrybench) | [outputs.json](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/resolve/main/gpt-oss-20b/sorrybench/n100_nbase10_nsamp10_l4096_s-42_t1.0_outputs.json?download=true) | |
| | gpt-oss-20b | survival_instinct | [browse](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/tree/main/gpt-oss-20b/survival_instinct) | [outputs.json](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/resolve/main/gpt-oss-20b/survival_instinct/n100_nbase10_nsamp10_l4096_s-42_t1.0_outputs.json?download=true) | |
| | gpt-oss-20b | wealth_seeking | [browse](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/tree/main/gpt-oss-20b/wealth_seeking) | [outputs.json](https://huggingface.co/datasets/future-probes/per_sentence_probabilities/resolve/main/gpt-oss-20b/wealth_seeking/n100_nbase10_nsamp10_l4096_s-42_t1.0_outputs.json?download=true) | |
|
|
|
|
| ## Paper |
|
|
| [Predicting Future Behaviors in Reasoning Models Enables Better Steering](https://arxiv.org/abs/2606.11172) |
|
|
| ## Citation |
|
|
| ```bibtex |
| @misc{kortukov2026predictingfuturebehaviorsreasoning, |
| title={Predicting Future Behaviors in Reasoning Models Enables Better Steering}, |
| author={Evgenii Kortukov and Piotr Komorowski and Florian Klein and Paula Engl and Gabriele Sarti and Seong Joon Oh and Sebastian Lapuschkin and Wojciech Samek}, |
| year={2026}, |
| eprint={2606.11172}, |
| archivePrefix={arXiv}, |
| primaryClass={cs.LG}, |
| url={https://arxiv.org/abs/2606.11172}, |
| } |
| ``` |
|
|
