longcot-24k-1.5b

McGill-NLP/longcot-24k-1.5b

<p align="center"> <img src="./method.png" width=100%/> </p>

TL;DR

• Markovian Thinking for RL in reasoning LLMs: replace the trivial MDP where state = prompt + all past thinking tokens (quadratic compute) with a bounded, fixed-size state, yielding linear compute in thinking tokens and constant memory by design.
• Delethink RL trains a model to “think” in fixed-size chunks with bounded state..
• This 1.5B model uses an effective thinking budget of about 24K tokens while only requiring an 8K active context at any time via chunked rollouts and short carryovers.
• Initialized from deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B, trained with the Delethink RL paradigm. See the paper for full details.

Links

• Repo: https://github.com/McGill-NLP/the-markovian-thinker
• Paper: https://arxiv.org/abs/2510.06557v1
• Collection: The Markovian Thinker

Model Summary

• Base model: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
• Objective: Reinforcement Learning using standard LongCoT, trained for 1000 steps.
• Thinking 24K budget; uses the entire context.
• Intended use: Math/logic reasoning with step-by-step derivations; final answer typically formatted inside LaTeX \boxed{}.
• Library compatibility: Works well with SGLang for chunked inference; also usable with Transformers for standard generation.

Intended Uses and Limitations

• Intended uses:
  • Long-form reasoning on math and related tasks.
  • Bounded-context rollouts with repeated chunking and short carryovers.
• Not intended for:
  • Safety-sensitive applications without human oversight.
  • Use cases requiring faithful, verifiable citations to external sources.
• Limitations:
  • May hallucinate, make arithmetic/algebraic mistakes, or produce inconsistent plans.
  • The chunked rollout procedure is needed to realize Delethink’s efficiency advantages.

Prompting

• Use the model’s chat template and request a step-by-step solution with a final boxed answer:
  • “Please reason step by step, and put your final answer within \boxed{}.”

Quickstart (SGLang, chunked Delethink rollout)

import asyncio
import sglang as sgl

def main():
    llm = sgl.Engine(
        model_path="McGill-NLP/longcot-24k-1.5b",
        dtype="bfloat16",
        attention_backend="flashinfer",
        mem_fraction_static=0.8,
        log_level="WARNING",
    )

    prompt = (
        r"There exist real numbers $x$ and $y$, both greater than 1, such that "
        r"$\log_x\left(y^x\right)=\log_y\left(x^{4y}\right)=10$. Find $xy$."
        "\n\nPlease reason step by step, and put your final answer within \\boxed{}."
    )
    tok = llm.tokenizer_manager.tokenizer
    query_ids = tok.apply_chat_template(
        [{"role": "user", "content": prompt}],
        tokenize=True,
        add_generation_prompt=True,
    )

    params = {"temperature": 0.6, "max_new_tokens": 24576}
    ids = llm.generate(input_ids=query_ids, sampling_params=params, return_logprob=True)
    print(tok.decode(ids, skip_special_tokens=False))

if __name__ == "__main__":
    main()

Suggested generation settings

• temperature: 0.6
• top_p: 1.0
• top_k: -1

Safety and Use

• This model can produce incorrect or misleading reasoning steps and answers. Always verify results.
• Do not deploy in high-stakes domains without human oversight.

Citation

@misc{Aghajohari2025:TheMarkovianThinker,
    title={The Markovian Thinker}, 
    author={Milad Aghajohari and Kamran Chitsaz and Amirhossein Kazemnejad and 
    Sarath Chandar and Alessandro Sordoni and Aaron Courville and Siva Reddy},
    year={2025},
    eprint={...},
    archivePrefix={arXiv},
    primaryClass={cs.LG},
    url={...}, 
}