# Thinking Tokens for Language Modeling

arXiv: https://arxiv.org/abs/2405.08644 · HTML: https://mlsanity.com/paper/2405.08644/

## Key facts

- Cited by 24 papers in the map (data through 14 Oct 2026).
- Placed in Mathematical reasoning, one of 17 directions in LLM reasoning.
- Builds on 0 of its 18 references that are in the map.
- First posted on arXiv 14 May 2024.

## Placed in

- LLM reasoning / [Mathematical reasoning](https://mlsanity.com/lr/mathematical-reasoning/)

## Built on by (most cited first, 24 of 24)

- [RAGEN: Understanding Self-Evolution in LLM Agents via Multi-Turn Reinforcement Learning](https://mlsanity.com/paper/2504.20073/) (2025)
- [The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity](https://mlsanity.com/paper/2506.06941/) (2025)
- [Reasoning Beyond Language: A Comprehensive Survey on Latent Chain-of-Thought Reasoning](https://mlsanity.com/paper/2505.16782/) (2025)
- [Multi-Agent Verification: Scaling Test-Time Compute with Multiple Verifiers](https://mlsanity.com/paper/2502.20379/) (2025)
- [Implicit Reasoning in Large Language Models: A Comprehensive Survey](https://mlsanity.com/paper/2509.02350/) (2025)
- [LaDi-RL: Latent Diffusion Reasoning Prevents Entropy Collapse in Reinforcement Learning](https://mlsanity.com/paper/2602.01705/) (2026)
- [The Illusion of Superposition? A Principled Analysis of Latent Thinking in Language Models](https://mlsanity.com/paper/2604.06374/) (2026)
- [Large Language Models are Limited in Out-of-Context Knowledge Reasoning](https://mlsanity.com/paper/2406.07393/) (2024)
- [Meta-Reasoning Improves Tool Use in Large Language Models](https://mlsanity.com/paper/2411.04535/) (2024)
- [Latent Recurrent Transformer: Architecture Exploration, Training Strategies, and Scaling Behavior](https://mlsanity.com/paper/2605.26797/) (2026)
- [SPOT: Span-level Pause-of-Thought for Efficient and Interpretable Latent Reasoning in Large Language Models](https://mlsanity.com/paper/2603.06222/) (2026)
- [From Reasoning Traces to Reusable Modules: Understanding Compositional Generalization in Language Model Reasoning](https://mlsanity.com/paper/2606.18089/) (2026)
- [Beyond Memorization: Extending Reasoning Depth with Recurrence, Memory and Test-Time Compute Scaling](https://mlsanity.com/paper/2508.16745/) (2025)
- [Thoughtbubbles: an Unsupervised Method for Parallel Thinking in Latent Space](https://mlsanity.com/paper/2510.00219/) (2025)
- [Capabilities and Fundamental Limits of Latent Chain-of-Thought](https://mlsanity.com/paper/2602.01148/) (2026)
- [Maglev: Sliding Recurrent Memory](https://mlsanity.com/paper/2608.02870/) (2026)
- [Rethinking Thinking Tokens: Understanding Why They Underperform in Practice](https://mlsanity.com/paper/2411.11371/) (2024)
- [Expanding Computation Spaces of LLMs at Inference Time](https://mlsanity.com/paper/2509.24884/) (2025)
- [MUX: Continuous Reasoning via Multiplexed Tokens](https://mlsanity.com/paper/2607.18264/) (2026)
- [Thinking into the Future: Latent Lookahead Training for Transformers](https://mlsanity.com/paper/2603.20219/) (2026)
- [HRM-Agent: Training a recurrent reasoning model in dynamic environments using reinforcement learning](https://mlsanity.com/paper/2510.22832/) (2025)
- [Liquid Reasoning Transformers: A Sudoku-Based Prototype for Chess-Scale Algorithmic Tasks](https://mlsanity.com/paper/2512.12792/) (2025)
- [Masked diffusion LLMs can use EoS tokens for hidden reasoning](https://mlsanity.com/paper/2603.05197/) (2026)
- [What Matters for Latent Reasoning with Flow Matching](https://mlsanity.com/paper/2610.06666/) (2026)

## Abstract (arXiv)

How much is 56 times 37? Language models often make mistakes in these types of difficult calculations. This is usually explained by their inability to perform complex reasoning. Since language models rely on large training sets and great memorization capability, naturally they are not equipped to run complex calculations. However, one can argue that humans also cannot perform this calculation immediately and require a considerable amount of time to construct the solution. In order to enhance the generalization capability of language models, and as a parallel to human behavior, we propose to use special 'thinking tokens' which allow the model to perform much more calculations whenever a complex problem is encountered.