Instructions to use smSquirrel/AnyS1-4B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use smSquirrel/AnyS1-4B with PEFT:
from peft import PeftModel from transformers import AutoModel base_model = AutoModel.from_pretrained("Qwen/Qwen3.5-4B-Base") model = PeftModel.from_pretrained(base_model, "smSquirrel/AnyS1-4B") - Notebooks
- Google Colab
- Kaggle
AnyS1-4B
AnyS1-4B is the System-1 model from Thinking Fast, Anywhere: One System-1 for the Decision Points of Any Agent Workflow (arXiv:2611.23751). It is a 4B pointer model that scores every candidate option of an enumerable agent decision (an E node) in a single forward pass and returns calibrated probabilities. A confidence gate uses these probabilities to decide whether to answer locally or defer to the large System-2 model.
- Paper: https://arxiv.org/abs/2611.23751 Β· HF paper page
- Project page: https://middle-chunjie.github.io/AnyS1/
- Code: https://github.com/middle-chunjie/AnyS1
- Training data: Any-DP
Model details
| Base model | Qwen/Qwen3.5-4B-Base (revision 1001bb4d826a52d1f399e183466143f4da7b741b) |
| Adaptation | LoRA, rank 16, alpha 32, dropout 0.05, all linear layers |
| Scoring head | OptionInteractionHead (oc, head dim 256, permutation-equivariant across options), stored in head.pt |
| Objective | Cross-entropy with label smoothing Ξ΅ = 0.1 over options |
| Training data | Any-DP train split: 25,986 container-executed states from 1,072 public skills, each seen under 3 option orders |
| Precision | bf16 |
Files: adapter_model.safetensors and adapter_config.json (LoRA), head.pt (pointer head), tokenizer files, training_config.json, training_metrics.json.
Results (from the paper)
- Offline held-out Dev: 83.8% accuracy; SOPBench E nodes 93.5%.
- Online, with the confidence gate calibrated on 50 development episodes (
warm), across SOPBench, ALFWorld and WebShop with DeepSeek-V4.1-Flash or GPT-6 Sol as System-2: mean task-success change of +0.5 (Flash) and β0.2 (Sol) points against the compiled all-System-2 agent, while episode cost falls by 75β91% and model-call time by 63β92% relative to an uncompiled System-2 agent. - Median single-pass latency 21.8β31.7 ms on one H100, versus 1.51β1.78 s per System-2 call.
Usage
Inference needs the custom head and therefore the AnyS1 code:
git clone https://github.com/middle-chunjie/AnyS1.git && cd AnyS1 && pip install -e .
huggingface-cli download smSquirrel/AnyS1-4B --local-dir ckpt/AnyS1-4B
python -m anys1.model.serve_s1 --checkpoint ckpt/AnyS1-4B
The server exposes /decide, which takes a state, a question and a list of options and returns one probability per option.
Limitations
The model only scores enumerable decisions whose options are listed explicitly; open-ended generation stays with System-2. Calibration was checked on the benchmarks in the paper. A new workflow should either calibrate the deferral threshold on a few development episodes or use the conservative transferred threshold.
Citation
@misc{li2026anys1,
title = {Thinking Fast, Anywhere: One System-1 for the Decision Points of Any Agent Workflow},
author = {Rui Li and Liyang He and Zheng Zhang and Zhenya Huang and Qi Liu},
year = {2026},
eprint = {2611.23751},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
url = {https://arxiv.org/abs/2611.23751}
}
- Downloads last month
- -
Model tree for smSquirrel/AnyS1-4B
Base model
Qwen/Qwen3.5-4B-Base