Special Issue for Any New Related Work

[DevoAllen]

Hello guys,

It's our privilege to receive your attention to our work!

Although we have tried our best to survey the relevant work, we may still have overlooked some, and the field of reasoning economy is evolving rapidly.

Therefore, we welcome you to submit any papers you think are related to the reasoning economy in this issue.

We will first consider adding them to the GitHub repo and then updating them in the next revision of our paper.

Thanks!

[Jiayu320]

And there is one work about a routing method for Multi-Agent System (MAS) reduce the cost of multiple LLMs:
MasRouter: Learning to Route LLMs for Multi-Agent Systems

[ChengpengLi1003]

START: Self-taught Reasoner with Tools
https://arxiv.org/abs/2503.04625

The authors proposed Hint-infer, which is the first to integrate the use of Python tools into Long CoT, introduced Long TIR, and enabled LRM to achieve test time scaling through Hint-infer.

[ruipeterpan]

Thanks for maintaining this repo! Two related papers in this space:

• Reasoning Without Self-Doubt: More Efficient Chain-of-Thought Through Certainty Probing seems like a good fit for 4.2.2's Early Stopping
• SpecReason: Fast and Accurate Inference-Time Compute via Speculative Reasoning is our new paper -- the core idea is to exploit the semantic approximation tolerance nature of thinking tokens in LRM’s CoTs and opportunistically offload certain reasoning steps to a lightweight model to accelerate inference while maintaining accuracy. It probably best fits in 4.2 (output-side adaptive-decoding)?

[DevoAllen]

And there is one work about a routing method for Multi-Agent System (MAS) reduce the cost of multiple LLMs: MasRouter: Learning to Route LLMs for Multi-Agent Systems

Added! 😊

[DevoAllen]

Thanks for maintaining this repo! Two related papers in this space:

• Reasoning Without Self-Doubt: More Efficient Chain-of-Thought Through Certainty Probing seems like a good fit for 4.2.2's Early Stopping
• SpecReason: Fast and Accurate Inference-Time Compute via Speculative Reasoning is our new paper -- the core idea is to exploit the semantic approximation tolerance nature of thinking tokens in LRM’s CoTs and opportunistically offload certain reasoning steps to a lightweight model to accelerate inference while maintaining accuracy. It probably best fits in 4.2 (output-side adaptive-decoding)?

Added! 😊 Thanks for your discussion!

[DevoAllen]

START: Self-taught Reasoner with Tools https://arxiv.org/abs/2503.04625

The authors proposed Hint-infer, which is the first to integrate the use of Python tools into Long CoT, introduced Long TIR, and enabled LRM to achieve test time scaling through Hint-infer.

Added! 😊 Thanks for your discussion!

[woominsong]

Thanks for the great work!

Title: Accelerated Test-Time Scaling with Model-Free Speculative Sampling
arxiv: https://arxiv.org/abs/2506.04708

This paper introduces a model-free speculative decoding method to accelerate test-time scaling without affecting its outputs, by utilizing the N-gram overlaps across different reasoning trajectories. I believe this will be a useful addition to the list!

I also created a pull request #8 for easier integration.

[DevoAllen]

Thanks for the great work!

Title: Accelerated Test-Time Scaling with Model-Free Speculative Sampling arxiv: https://arxiv.org/abs/2506.04708

This paper introduces a model-free speculative decoding method to accelerate test-time scaling without affecting its outputs, by utilizing the N-gram overlaps across different reasoning trajectories. I believe this will be a useful addition to the list!

I also created a pull request #8 for easier integration.

Already updated! 😊 Thanks for your discussion!

[XinXU-USTC]

Thanks for the great work!

Title: TFPI: Thinking-Free Policy Initialization Makes Distilled Reasoning Models More Effective and Efficient Reasoners
arxiv: https://arxiv.org/abs/2509.26226

This paper discusses an alternative way to train token-efficient reasoning models without specialized reward shaping for RLVR.

[DevoAllen]

Thanks for the great work!

Title: TFPI: Thinking-Free Policy Initialization Makes Distilled Reasoning Models More Effective and Efficient Reasoners arxiv: https://arxiv.org/abs/2509.26226

This paper discusses an alternative way to train token-efficient reasoning models without specialized reward shaping for RLVR.

Thanks for the discussion! The updates have been pushed to GitHub, and our new paper will incorporate your feedback.