ZIP-RC

Official repository for "Zero-Overhead Introspection for Adaptive Test-Time Compute".

ZIP-RC equips an LLM with zero-overhead introspective predictions of a joint distribution over: (i) final reward (e.g., correctness) and (ii) remaining generation length.

It does this by reserving a contiguous slice of vocabulary logits for an auxiliary head and reading those logits in the same forward pass used for next-token prediction. During decoding, those reserved tokens must be masked so they are never sampled.

Links

• Project page: https://rohinmanvi.github.io/ZIP-RC/
• Paper (PDF): https://arxiv.org/pdf/2512.01457.pdf
• arXiv: https://arxiv.org/abs/2512.01457

Quickstart

Create the conda environment:

conda env create -f environment.yml
conda activate zip

Run a tiny end-to-end smoke test (downloads Hugging Face models/datasets; GPU required):

bash scripts/smoke_test.sh

For the full training pipeline + script flags, see docs/pipeline.md.

Citation

If you find ZIP-RC useful, please cite:

@misc{manvi2025zerooverheadintrospectionadaptivetesttime,
      title={Zero-Overhead Introspection for Adaptive Test-Time Compute},
      author={Rohin Manvi and Joey Hong and Tim Seyde and Maxime Labonne and Mathias Lechner and Sergey Levine},
      year={2025},
      eprint={2512.01457},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2512.01457},
}