Views
No views yet
Qwen/Qwen3.5-9B — Recovering Instruction Sets from
Language Model Activations (arXiv:2606.09563).Qwen/Qwen3.5-9B, which you obtain separately
under its own licence. No base weights are included or redistributed here.| Target model | Qwen/Qwen3.5-9B |
| Hook layer | 16 |
| Activation window | last 128 response tokens |
| Training | SFT followed by judge-guided GRPO. This is the main method in the paper. |
| File | prism-qwen3.5-9b-grpo.pt (266 MB) |
1git clone https://github.com/Offensive-AI-Lab/prism-eval
2cd prism-eval && uv sync --extra dev
3
4export PRISM_EVAL_CHECKPOINT_DIR=./checkpoints
5python scripts/download_weights.py --only prism-qwen3.5-9b-grpo
6
7prism-eval evaluate --config configs/main/qwen3.5-9b-grpo.yaml --offlineprism-qwen3.5-9b-grpo.pt sha256 0968b4581517dca88ee229a44dc575240ebfcb8f2b6e2b7d9674c8b3d29a0659scripts/download_weights.py checks this on download and treats a mismatch as a
hard error, so a truncated or substituted file fails loudly instead of producing
wrong numbers.1@article{gressel2026prism,
2 title = {PRISM: Recovering Instruction Sets from Language Model Activations},
3 author = {Gilad Gressel and Rahul Pankajakshan and Julia Diament and Efim Hudis and Krishnashree Achuthan and Yisroel Mirsky},
4 journal = {arXiv preprint arXiv:2606.09563},
5 year = {2026},
6 url = {https://arxiv.org/abs/2606.09563}
7}