PRISM without reinforcement learning for Qwen3.5-9B

This repository contains the supervised-only PRISM checkpoint used for the "PRISM w/o RL" result in PRISM: Recovering Instruction Sets from Language Model Activations. It reads residual-stream activations from a target model's response and decodes the instructions that produced it.

The checkpoint contains a learned projection and LoRA parameters. It does not contain Qwen3.5-9B weights.

Checkpoint

Target model	Hook layer	Activation window	Training
`Qwen/Qwen3.5-9B`	16	Last 128 response tokens	SFT

prism-qwen3.5-9b-sft.pt SHA-256: 347cc6c6674a839dd995633a057f9ddb6cb45e153f912671445eed66c6a56002

On the paper's 1,000-record evaluation suite, this checkpoint obtains 0.653 mean reward, 0.691 mean coverage, and 0.037 mean hallucination rate. These are the results reported in the paper.

Use

Use the checkpoint with prism-eval:

git clone https://github.com/Offensive-AI-Lab/prism-eval
cd prism-eval
uv sync
uv run python scripts/download_weights.py --only prism-qwen3.5-9b-sft
uv run prism-eval evaluate --config configs/main/qwen3.5-9b-sft.yaml --offline

The evaluation requires a separate judge endpoint; see the repository README. For the interactive demo and training code, use prism.

Limitations

PRISM can omit instructions or report instructions that were not present. The released evaluation is primarily English and uses single-response examples. Its outputs are interpretability evidence, not a safety guarantee.

License

The checkpoint is licensed under Apache-2.0. Qwen3.5-9B is downloaded separately and remains subject to its own license.

Citation

@inproceedings{gressel2026prism,
  title     = {PRISM: Recovering Instruction Sets from Language Model Activations},
  author    = {Gressel, Gilad and Pankajakshan, Rahul and Diament, Julia and
               Hudis, Efim and Achuthan, Krishnashree and Mirsky, Yisroel},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in
               Natural Language Processing},
  year      = {2026},
  url       = {https://arxiv.org/abs/2606.09563}
}