Input-Conditioned Search for Efficient Test-Time Scaling of Recursive Reasoners
Paper • 2610.18547 • Published
Checkpoints for ICS (Input-Conditioned Search; code: https://github.com/icsearch/ICS) and the baselines.
Paper: https://arxiv.org/abs/2610.18547 · Project page: https://icsearch.github.io
<model>/<task>/seed<k>/ holds config.json and model.safetensors.
from huggingface_hub import snapshot_download
root = snapshot_download(repo_id="hcshi/ICS", allow_patterns="trm/maze/seed0/*")
With the code installed, python -m ics download --model trm --task maze --seed 0 writes the same folder to
ckpt/trm/maze/seed0.
@article{shi2026ics,
title = {Input-Conditioned Search for Efficient Test-Time Scaling of Recursive Reasoners},
author = {Haochen Shi and Tianwen Fu and Qianggang Ding and Yinhan Lu and Yizhan Li and Huan Zhang and Yajie Zhao and Xingdi Yuan and Hengyang Lu and Bang Liu},
journal = {arXiv preprint arXiv:2610.18547},
year = {2026}
}
MIT.