HALT-CoT

HALT-CoT is an inference-time early stopping method for chain-of-thought reasoning. After each generated reasoning step, it computes the Shannon entropy of the model's answer distribution over candidate answers. When entropy stays below a threshold, generation stops and the current most likely answer is returned.

This repository packages the method from:

HALT-CoT: Model-Agnostic Early Stopping for Chain-of-Thought Reasoning via Answer Entropy
Yassir Laaouach

Official paper page · PDF mirror.

The release is a method implementation and Hugging Face Space, not trained model weights.

What Is Included

• Dependency-free core entropy and halting controller in halt_cot/core.py.
• Hugging Face Transformers backend in halt_cot/transformers_backend.py.
• CLI entrypoint: halt-cot.
• Gradio Space app in app.py.
• Publish helper for huggingface_hub.
• Focused tests for entropy, candidate sets, and halting behavior.
• Paper PDF in paper/halt-cot-paper.pdf.

Install

pip install -e ".[transformers]"

For the Gradio Space locally:

pip install -e ".[demo]"
python app.py

CLI Example

python -m halt_cot.cli \
  --model Qwen/Qwen2.5-0.5B-Instruct \
  --question "If a shop has 12 apples and sells 5, how many apples are left?" \
  --candidates 5 6 7 8 9 \
  --theta 0.6 \
  --consecutive 2

Python Example

from halt_cot import HaltCoTConfig
from halt_cot.transformers_backend import HaltCoTForCausalLM

runner = HaltCoTForCausalLM.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct")
config = HaltCoTConfig(theta=0.6, consecutive_low_entropy=2, max_steps=8)

result = runner.run(
    "If a shop has 12 apples and sells 5, how many apples are left?",
    candidates=["5", "6", "7", "8", "9"],
    config=config,
)

print(result.answer)
print(result.reasoning)

Method

At reasoning step i, HALT-CoT scores candidate answers a in A from the next-token logits after the current partial chain. It normalizes those candidate scores into p_i(a) and computes:

H_i = - sum_a p_i(a) log p_i(a)

The controller halts when H_i < theta, optionally requiring the condition for multiple consecutive steps. This implementation uses bits by default, which aligns with the entropy traces in the paper. If you tune thresholds in nats, set entropy_unit="nats".

Suggested starting thresholds from the paper:

• Math-style numeric tasks: theta in [0.5, 0.7].
• StrategyQA-style yes/no tasks: theta = 0.8.
• CommonsenseQA-style multiple choice: theta = 0.7.

For numeric tasks, use task-specific candidate answers where possible. The helper numeric_candidates_from_texts can build candidates from training answers plus an integer range.

Publish To Hugging Face

Log in once:

python -m huggingface_hub.commands.huggingface_cli login

Free Code/Method Repository

Hugging Face may require a PRO subscription for hosted Gradio Spaces on cpu-basic. To publish the method, code, docs, and tests for free, upload this folder as a normal Hugging Face model repository:

python scripts/publish_to_hf.py --repo-id <your-username>/halt-cot --repo-type model

For this account, the repo id is:

python scripts/publish_to_hf.py --repo-id yass4/halt-cot --repo-type model

Interactive Gradio Space

If your account can create Gradio Spaces, upload the same folder as a Space:

python scripts/publish_to_hf.py --repo-id <your-username>/halt-cot --repo-type space

The default Space model is controlled by HALT_COT_MODEL_ID. Set it in the Space variables to use a different causal LM. Use HALT_COT_DEVICE_MAP=auto on hardware that supports Accelerate device placement.

Citation

@misc{laaouach2026haltcot,
  title = {HALT-CoT: Model-Agnostic Early Stopping for Chain-of-Thought Reasoning via Answer Entropy},
  author = {Laaouach, Yassir},
  year = {2026}
}