# Thinking tokens

Give a model extra tokens for computation, such as pauses, filler symbols, or learned planning tokens. These tokens need not spell out the solution.

Papers are ordered by first public date. Read the original sources before using a claim.

For example: A model processes several pause tokens before answering a math question. Those extra positions let it perform additional computation, even though they do not display a worked solution.

Version 0.9.5; literature checked through 2026-10-18.

[Topic guide](https://latent-implicit-thinking.github.io/topics/thinking-tokens/) · [JSON](https://latent-implicit-thinking.github.io/data/topics/thinking-tokens.json)

## How the methods connect and differ

**Pausing and planning**

Pause tokens add computation before the answer without choosing a different plan token for each step. Planning tokens instead select a learned token before generating each written reasoning step. The latter still produces a textual chain of thought. In the pause-token experiments, the training setup matters: gains were strongest when pauses were used in both pretraining and fine-tuning, not simply inserted into an unchanged model.

Sources: [Pause-token paper](https://proceedings.iclr.cc/paper_files/paper/2024/hash/76917808731dae9e6d62c2a7a6afb542-Abstract-Conference.html) · [Planning-token paper, §2](https://arxiv.org/html/2310.05707v3#S2)

Paper details: [Think before you speak: Training language models with pause tokens](https://latent-implicit-thinking.github.io/papers/2310.02226/) · [Guiding Language Model Reasoning with Planning Tokens](https://latent-implicit-thinking.github.io/papers/2310.05707/)

## Papers — oldest first

## Thinking Tokens for Language Modeling

Authors: David Herel, Tomas Mikolov.

Gives a recurrent language model extra hidden-state updates by inserting special tokens after each word. Tests whether these updates improve next-word prediction.

First public: 2023-09-04.
ArXiv initial submission: 2024-05-14.

[Original paper](https://arxiv.org/abs/2405.08644) · [Reading-list entry](https://latent-implicit-thinking.github.io/papers/2405.08644/)
[Preprint BibTeX](https://latent-implicit-thinking.github.io/papers/2405.08644/citation.bib) · [Preprint CSL-JSON](https://latent-implicit-thinking.github.io/papers/2405.08644/citation.csl.json) · [JSON record](https://latent-implicit-thinking.github.io/papers/2405.08644/record.json)

Date note: AITP presentation date; the earlier public release date of the conference abstract is unverified. [Source](https://aitp-conference.org/2023/).

### Approaches

- [Thinking tokens](https://latent-implicit-thinking.github.io/topics/thinking-tokens/): Give a model extra tokens for computation, such as pauses, filler symbols, or learned planning tokens. These tokens need not spell out the solution.

Summary and categories are editorial. Read the original paper to verify claims and limitations. A citation to a paper does not, by itself, establish that a later method builds on it.

Reading-list version 0.9.5; literature checked through 2026-10-18.


## Think before you speak: Training language models with pause tokens

Authors: Sachin Goyal, Ziwei Ji, Ankit Singh Rawat, Aditya Krishna Menon, Sanjiv Kumar, Vaishnavh Nagarajan.

Introduces pause tokens that give a language model additional computational positions before it must produce an answer.

First public: 2023-10-03.
ArXiv initial submission: 2023-10-03.

[Original paper](https://arxiv.org/abs/2310.02226) · [Reading-list entry](https://latent-implicit-thinking.github.io/papers/2310.02226/)
[Preprint BibTeX](https://latent-implicit-thinking.github.io/papers/2310.02226/citation.bib) · [Preprint CSL-JSON](https://latent-implicit-thinking.github.io/papers/2310.02226/citation.csl.json) · [JSON record](https://latent-implicit-thinking.github.io/papers/2310.02226/record.json)

### Method comparisons

**Pausing and planning**

Pause tokens add computation before the answer without choosing a different plan token for each step. Planning tokens instead select a learned token before generating each written reasoning step. The latter still produces a textual chain of thought. In the pause-token experiments, the training setup matters: gains were strongest when pauses were used in both pretraining and fine-tuning, not simply inserted into an unchanged model.

Sources: [Pause-token paper](https://proceedings.iclr.cc/paper_files/paper/2024/hash/76917808731dae9e6d62c2a7a6afb542-Abstract-Conference.html) · [Planning-token paper, §2](https://arxiv.org/html/2310.05707v3#S2)

Paper details: [Think before you speak: Training language models with pause tokens](https://latent-implicit-thinking.github.io/papers/2310.02226/) · [Guiding Language Model Reasoning with Planning Tokens](https://latent-implicit-thinking.github.io/papers/2310.05707/)


### Published version

Think before you speak: Training Language Models With Pause Tokens. Sachin Goyal, Ziwei Ji, Ankit Singh Rawat, Aditya Krishna Menon, Sanjiv Kumar, Vaishnavh Nagarajan. International Conference on Learning Representations, 2024.
[Publication](https://proceedings.iclr.cc/paper_files/paper/2024/hash/76917808731dae9e6d62c2a7a6afb542-Abstract-Conference.html) · [BibTeX](https://latent-implicit-thinking.github.io/papers/2310.02226/publication.bib) · [CSL-JSON](https://latent-implicit-thinking.github.io/papers/2310.02226/publication.csl.json)
Metadata checked 2026-10-18 against [the publication record](https://proceedings.iclr.cc/paper_files/paper/2024/hash/76917808731dae9e6d62c2a7a6afb542-Abstract-Conference.html).

### Approaches

- [Thinking tokens](https://latent-implicit-thinking.github.io/topics/thinking-tokens/): Give a model extra tokens for computation, such as pauses, filler symbols, or learned planning tokens. These tokens need not spell out the solution.

Summary and categories are editorial. Read the original paper to verify claims and limitations. A citation to a paper does not, by itself, establish that a later method builds on it.

Reading-list version 0.9.5; literature checked through 2026-10-18.


## Guiding Language Model Reasoning with Planning Tokens

Authors: Xinyi Wang, Lucas Caccia, Oleksiy Ostapenko, Xingdi Yuan, William Yang Wang, Alessandro Sordoni.

Adds a learned planning token before each written reasoning step, then uses that token to guide the step.

First public: 2023-10-09.
ArXiv initial submission: 2023-10-09.

[Original paper](https://arxiv.org/abs/2310.05707) · [Reading-list entry](https://latent-implicit-thinking.github.io/papers/2310.05707/)
[Preprint BibTeX](https://latent-implicit-thinking.github.io/papers/2310.05707/citation.bib) · [Preprint CSL-JSON](https://latent-implicit-thinking.github.io/papers/2310.05707/citation.csl.json) · [JSON record](https://latent-implicit-thinking.github.io/papers/2310.05707/record.json)
[Code](https://github.com/WANGXinyiLinda/planning_tokens)

### Method comparisons

**Pausing and planning**

Pause tokens add computation before the answer without choosing a different plan token for each step. Planning tokens instead select a learned token before generating each written reasoning step. The latter still produces a textual chain of thought. In the pause-token experiments, the training setup matters: gains were strongest when pauses were used in both pretraining and fine-tuning, not simply inserted into an unchanged model.

Sources: [Pause-token paper](https://proceedings.iclr.cc/paper_files/paper/2024/hash/76917808731dae9e6d62c2a7a6afb542-Abstract-Conference.html) · [Planning-token paper, §2](https://arxiv.org/html/2310.05707v3#S2)

Paper details: [Think before you speak: Training language models with pause tokens](https://latent-implicit-thinking.github.io/papers/2310.02226/) · [Guiding Language Model Reasoning with Planning Tokens](https://latent-implicit-thinking.github.io/papers/2310.05707/)


### Approaches

- [Thinking tokens](https://latent-implicit-thinking.github.io/topics/thinking-tokens/): Give a model extra tokens for computation, such as pauses, filler symbols, or learned planning tokens. These tokens need not spell out the solution.

Summary and categories are editorial. Read the original paper to verify claims and limitations. A citation to a paper does not, by itself, establish that a later method builds on it.

Reading-list version 0.9.5; literature checked through 2026-10-18.


## Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking

Authors: Eric Zelikman, Georges Harik, Yijia Shao, Varuna Jayasiri, Nick Haber, Noah D. Goodman.

Trains on ordinary text by generating written rationales before predicting what comes next. Learns from how much those rationales improve subsequent token predictions.

First public: 2024-03-14.
ArXiv initial submission: 2024-03-14.

[Original paper](https://arxiv.org/abs/2403.09629) · [Reading-list entry](https://latent-implicit-thinking.github.io/papers/2403.09629/)
[Preprint BibTeX](https://latent-implicit-thinking.github.io/papers/2403.09629/citation.bib) · [Preprint CSL-JSON](https://latent-implicit-thinking.github.io/papers/2403.09629/citation.csl.json) · [JSON record](https://latent-implicit-thinking.github.io/papers/2403.09629/record.json)

### Approaches

- [Thinking tokens](https://latent-implicit-thinking.github.io/topics/thinking-tokens/): Give a model extra tokens for computation, such as pauses, filler symbols, or learned planning tokens. These tokens need not spell out the solution.

Summary and categories are editorial. Read the original paper to verify claims and limitations. A citation to a paper does not, by itself, establish that a later method builds on it.

Reading-list version 0.9.5; literature checked through 2026-10-18.


## Let's think dot by dot: Hidden computation in transformer language models

Authors: Jacob Pfau, William Merrill, Samuel R. Bowman.

Studies whether meaningless filler-token positions can provide transformers with additional hidden computation without carrying readable reasoning content.

First public: 2024-04-24.
ArXiv initial submission: 2024-04-24.

[Original paper](https://arxiv.org/abs/2404.15758) · [Reading-list entry](https://latent-implicit-thinking.github.io/papers/2404.15758/)
[Preprint BibTeX](https://latent-implicit-thinking.github.io/papers/2404.15758/citation.bib) · [Preprint CSL-JSON](https://latent-implicit-thinking.github.io/papers/2404.15758/citation.csl.json) · [JSON record](https://latent-implicit-thinking.github.io/papers/2404.15758/record.json)
[Code](https://github.com/JacobPfau/fillerTokens)

### Approaches

- [Thinking tokens](https://latent-implicit-thinking.github.io/topics/thinking-tokens/): Give a model extra tokens for computation, such as pauses, filler symbols, or learned planning tokens. These tokens need not spell out the solution.

Summary and categories are editorial. Read the original paper to verify claims and limitations. A citation to a paper does not, by itself, establish that a later method builds on it.

Reading-list version 0.9.5; literature checked through 2026-10-18.


## Token Assorted: Mixing Latent and Text Tokens for Improved Language Model Reasoning

Authors: DiJia Su, Hanlin Zhu, Yingchen Xu, Jiantao Jiao, Yuandong Tian, Qinqing Zheng.

Interleaves latent and textual tokens so a model can combine compressed internal computation with readable reasoning steps.

First public: 2025-02-05.
ArXiv initial submission: 2025-02-05.

[Original paper](https://arxiv.org/abs/2502.03275) · [Reading-list entry](https://latent-implicit-thinking.github.io/papers/2502.03275/)
[Preprint BibTeX](https://latent-implicit-thinking.github.io/papers/2502.03275/citation.bib) · [Preprint CSL-JSON](https://latent-implicit-thinking.github.io/papers/2502.03275/citation.csl.json) · [JSON record](https://latent-implicit-thinking.github.io/papers/2502.03275/record.json)

### Approaches

- [Thinking tokens](https://latent-implicit-thinking.github.io/topics/thinking-tokens/): Give a model extra tokens for computation, such as pauses, filler symbols, or learned planning tokens. These tokens need not spell out the solution.
- [Continuous-state reasoning](https://latent-implicit-thinking.github.io/topics/continuous-states/): Use vectors—lists of numbers—to carry intermediate reasoning from one step to the next, instead of selecting a text token at every step. Some methods use the model's hidden states; others combine several candidate token representations.

Summary and categories are editorial. Read the original paper to verify claims and limitations. A citation to a paper does not, by itself, establish that a later method builds on it.

Reading-list version 0.9.5; literature checked through 2026-10-18.


## Efficient Pretraining Length Scaling

Authors: Bohong Wu, Shen Yan, Sijun Zhang, Jianqiao Lu, Yutao Zeng, Ya Wang, Xun Zhou.

Adds hidden decoding tokens during pretraining, then discards their key/value cache entries after use. Keeps the original tokens' entries to preserve longer-range context.

First public: 2025-04-21.
ArXiv initial submission: 2025-04-21.

[Original paper](https://arxiv.org/abs/2504.14992) · [Reading-list entry](https://latent-implicit-thinking.github.io/papers/2504.14992/)
[Preprint BibTeX](https://latent-implicit-thinking.github.io/papers/2504.14992/citation.bib) · [Preprint CSL-JSON](https://latent-implicit-thinking.github.io/papers/2504.14992/citation.csl.json) · [JSON record](https://latent-implicit-thinking.github.io/papers/2504.14992/record.json)

### Approaches

- [Thinking tokens](https://latent-implicit-thinking.github.io/topics/thinking-tokens/): Give a model extra tokens for computation, such as pauses, filler symbols, or learned planning tokens. These tokens need not spell out the solution.
- [Recurrent reasoning](https://latent-implicit-thinking.github.io/topics/recurrent-reasoning/): Repeat parts of the model or keep intermediate information in memory while working through a problem. Some methods also learn how many passes to run.

Summary and categories are editorial. Read the original paper to verify claims and limitations. A citation to a paper does not, by itself, establish that a later method builds on it.

Reading-list version 0.9.5; literature checked through 2026-10-18.


## Enhancing Latent Computation in Transformers with Latent Tokens

Authors: Yuchang Sun, Yanxi Chen, Yaliang Li, Bolin Ding.

Adds trainable dummy tokens to a pretrained transformer so attention can use them for extra computation during decoding. The tokens can be inserted in different positions at inference time.

First public: 2025-05-19.
ArXiv initial submission: 2025-05-19.

[Original paper](https://arxiv.org/abs/2505.12629) · [Reading-list entry](https://latent-implicit-thinking.github.io/papers/2505.12629/)
[Preprint BibTeX](https://latent-implicit-thinking.github.io/papers/2505.12629/citation.bib) · [Preprint CSL-JSON](https://latent-implicit-thinking.github.io/papers/2505.12629/citation.csl.json) · [JSON record](https://latent-implicit-thinking.github.io/papers/2505.12629/record.json)

### Approaches

- [Thinking tokens](https://latent-implicit-thinking.github.io/topics/thinking-tokens/): Give a model extra tokens for computation, such as pauses, filler symbols, or learned planning tokens. These tokens need not spell out the solution.

Summary and categories are editorial. Read the original paper to verify claims and limitations. A citation to a paper does not, by itself, establish that a later method builds on it.

Reading-list version 0.9.5; literature checked through 2026-10-18.


## Machine Mental Imagery: Empower Multimodal Reasoning with Latent Visual Tokens

Authors: Zeyuan Yang, Xueyang Yu, Delin Chen, Maohao Shen, Chuang Gan.

Uses latent visual tokens as an internal imagery channel for multimodal reasoning without rendering every intermediate state.

First public: 2025-06-20.
ArXiv initial submission: 2025-06-20.

[Original paper](https://arxiv.org/abs/2506.17218) · [Reading-list entry](https://latent-implicit-thinking.github.io/papers/2506.17218/)
[Preprint BibTeX](https://latent-implicit-thinking.github.io/papers/2506.17218/citation.bib) · [Preprint CSL-JSON](https://latent-implicit-thinking.github.io/papers/2506.17218/citation.csl.json) · [JSON record](https://latent-implicit-thinking.github.io/papers/2506.17218/record.json)

### Approaches

- [Thinking tokens](https://latent-implicit-thinking.github.io/topics/thinking-tokens/): Give a model extra tokens for computation, such as pauses, filler symbols, or learned planning tokens. These tokens need not spell out the solution.
- [Continuous-state reasoning](https://latent-implicit-thinking.github.io/topics/continuous-states/): Use vectors—lists of numbers—to carry intermediate reasoning from one step to the next, instead of selecting a text token at every step. Some methods use the model's hidden states; others combine several candidate token representations.

Summary and categories are editorial. Read the original paper to verify claims and limitations. A citation to a paper does not, by itself, establish that a later method builds on it.

Reading-list version 0.9.5; literature checked through 2026-10-18.


## Chain Of Thought Compression: A Theoretical Analysis

Authors: Juncai Li, Ru Li, Yuxiang Zhou, Boxiang Ma, Jeff Z. Pan.

Analyzes how skipping intermediate steps weakens learning signals in a simplified parity-task model. Introduces NatBool-DAG to test logical dependencies and ALiCoT, which supervises learned token states with intermediate logical labels or explicit-CoT representations.

First public: 2026-01-29.
ArXiv initial submission: 2026-01-29.

[Original paper](https://arxiv.org/abs/2601.21576) · [Reading-list entry](https://latent-implicit-thinking.github.io/papers/2601.21576/)
[Preprint BibTeX](https://latent-implicit-thinking.github.io/papers/2601.21576/citation.bib) · [Preprint CSL-JSON](https://latent-implicit-thinking.github.io/papers/2601.21576/citation.csl.json) · [JSON record](https://latent-implicit-thinking.github.io/papers/2601.21576/record.json)

### Approaches

- [Internalized reasoning](https://latent-implicit-thinking.github.io/topics/internalized-reasoning/): Learn from worked solutions so that some or all of the reasoning can later happen without writing those steps. Training may use a separate teacher model or progressively remove written steps from the model being trained.
- [Thinking tokens](https://latent-implicit-thinking.github.io/topics/thinking-tokens/): Give a model extra tokens for computation, such as pauses, filler symbols, or learned planning tokens. These tokens need not spell out the solution.
- [Analysis and evaluation](https://latent-implicit-thinking.github.io/topics/implicit-reasoning-evaluation/): Measure how well these methods work, investigate what their internal computation does, and identify where they fail. This includes experiments, studies of model internals, and mathematical results.

Summary and categories are editorial. Read the original paper to verify claims and limitations. A citation to a paper does not, by itself, establish that a later method builds on it.

Reading-list version 0.9.5; literature checked through 2026-10-18.


## Latent Reasoning with Supervised Thinking States

Authors: Ido Amos, Avi Caciularu, Mor Geva, Amir Globerson, Jonathan Herzig, Lior Shani, Idan Szpektor.

Supervises intermediate thinking states that bridge textual reasoning traces and latent computation.

First public: 2026-02-09.
ArXiv initial submission: 2026-02-09.

[Original paper](https://arxiv.org/abs/2602.08332) · [Reading-list entry](https://latent-implicit-thinking.github.io/papers/2602.08332/)
[Preprint BibTeX](https://latent-implicit-thinking.github.io/papers/2602.08332/citation.bib) · [Preprint CSL-JSON](https://latent-implicit-thinking.github.io/papers/2602.08332/citation.csl.json) · [JSON record](https://latent-implicit-thinking.github.io/papers/2602.08332/record.json)

### Approaches

- [Thinking tokens](https://latent-implicit-thinking.github.io/topics/thinking-tokens/): Give a model extra tokens for computation, such as pauses, filler symbols, or learned planning tokens. These tokens need not spell out the solution.
- [Continuous-state reasoning](https://latent-implicit-thinking.github.io/topics/continuous-states/): Use vectors—lists of numbers—to carry intermediate reasoning from one step to the next, instead of selecting a text token at every step. Some methods use the model's hidden states; others combine several candidate token representations.

Summary and categories are editorial. Read the original paper to verify claims and limitations. A citation to a paper does not, by itself, establish that a later method builds on it.

Reading-list version 0.9.5; literature checked through 2026-10-18.


## ThinkRouter: Efficient Reasoning via Routing Thinking between Latent and Discrete Spaces

Authors: Xin Xu, Tong Yu, Xiang Chen, Haoliang Wang, Julian McAuley, Saayan Mitra.

Routes intermediate computation between latent and discrete representations to trade off reasoning quality, readability, and token cost.

First public: 2026-02-12.
ArXiv initial submission: 2026-02-12.

[Original paper](https://arxiv.org/abs/2602.11683) · [Reading-list entry](https://latent-implicit-thinking.github.io/papers/2602.11683/)
[Preprint BibTeX](https://latent-implicit-thinking.github.io/papers/2602.11683/citation.bib) · [Preprint CSL-JSON](https://latent-implicit-thinking.github.io/papers/2602.11683/citation.csl.json) · [JSON record](https://latent-implicit-thinking.github.io/papers/2602.11683/record.json)

### Approaches

- [Thinking tokens](https://latent-implicit-thinking.github.io/topics/thinking-tokens/): Give a model extra tokens for computation, such as pauses, filler symbols, or learned planning tokens. These tokens need not spell out the solution.
- [Continuous-state reasoning](https://latent-implicit-thinking.github.io/topics/continuous-states/): Use vectors—lists of numbers—to carry intermediate reasoning from one step to the next, instead of selecting a text token at every step. Some methods use the model's hidden states; others combine several candidate token representations.

Summary and categories are editorial. Read the original paper to verify claims and limitations. A citation to a paper does not, by itself, establish that a later method builds on it.

Reading-list version 0.9.5; literature checked through 2026-10-18.


## Almost Free State Prediction Separation

Authors: John Langford, Nathan Godey, Giovanni Monea, Yoav Artzi, Harry Dong, Ying Fan, Gustavo de Rosa, Zheng Zhan.

Moves pause-token computation into a parallel, weight-shared prediction stream, providing additional internal compute without increasing generated sequence length or key/value-cache length.

First public: 2026-09-03.
ArXiv initial submission: 2026-09-03.

[Original paper](https://arxiv.org/abs/2609.03807) · [Reading-list entry](https://latent-implicit-thinking.github.io/papers/2609.03807/)
[Preprint BibTeX](https://latent-implicit-thinking.github.io/papers/2609.03807/citation.bib) · [Preprint CSL-JSON](https://latent-implicit-thinking.github.io/papers/2609.03807/citation.csl.json) · [JSON record](https://latent-implicit-thinking.github.io/papers/2609.03807/record.json)

### Approaches

- [Thinking tokens](https://latent-implicit-thinking.github.io/topics/thinking-tokens/): Give a model extra tokens for computation, such as pauses, filler symbols, or learned planning tokens. These tokens need not spell out the solution.
- [Alternative generation methods](https://latent-implicit-thinking.github.io/topics/alternative-reasoning-schedules/): Change how intermediate steps are generated, revised, or searched—for example, by refining several steps in parallel. These are related approaches: some still produce written chains of thought.

Summary and categories are editorial. Read the original paper to verify claims and limitations. A citation to a paper does not, by itself, establish that a later method builds on it.

Reading-list version 0.9.5; literature checked through 2026-10-18.


## Towards Understanding Pause Token Fine-Tuning Dynamics: A Mode Retention Perspective

Authors: Jaehyeon Kim, Suhwan Kim, Nakyung Lee, Yeongoon Kim, Jimin Seo, Giho Lee, Jungwoo Lee.

Places pause tokens at reasoning-step boundaries and masks their training loss. Studies how this affects learning new tasks while retaining earlier capabilities, including during reinforcement learning.

First public: 2026-09-03.
ArXiv initial submission: 2026-09-03.

[Original paper](https://arxiv.org/abs/2609.04489) · [Reading-list entry](https://latent-implicit-thinking.github.io/papers/2609.04489/)
[Preprint BibTeX](https://latent-implicit-thinking.github.io/papers/2609.04489/citation.bib) · [Preprint CSL-JSON](https://latent-implicit-thinking.github.io/papers/2609.04489/citation.csl.json) · [JSON record](https://latent-implicit-thinking.github.io/papers/2609.04489/record.json)

### Approaches

- [Thinking tokens](https://latent-implicit-thinking.github.io/topics/thinking-tokens/): Give a model extra tokens for computation, such as pauses, filler symbols, or learned planning tokens. These tokens need not spell out the solution.
- [Analysis and evaluation](https://latent-implicit-thinking.github.io/topics/implicit-reasoning-evaluation/): Measure how well these methods work, investigate what their internal computation does, and identify where they fail. This includes experiments, studies of model internals, and mathematical results.

Summary and categories are editorial. Read the original paper to verify claims and limitations. A citation to a paper does not, by itself, establish that a later method builds on it.

Reading-list version 0.9.5; literature checked through 2026-10-18.


## Trading Depth for Time in Recurrent Transformers

Authors: Zeyi Huang, Xuehai He, Yelong Shen, Yong Jae Lee.

Inserts a hidden-state refinement step between consecutive text tokens in a recurrent transformer. Compares this extra thinking time with adding physical layers or transformer loops under matched computation budgets.

First public: 2026-09-18.
ArXiv initial submission: 2026-09-18.

[Original paper](https://arxiv.org/abs/2609.21605) · [Reading-list entry](https://latent-implicit-thinking.github.io/papers/2609.21605/)
[Preprint BibTeX](https://latent-implicit-thinking.github.io/papers/2609.21605/citation.bib) · [Preprint CSL-JSON](https://latent-implicit-thinking.github.io/papers/2609.21605/citation.csl.json) · [JSON record](https://latent-implicit-thinking.github.io/papers/2609.21605/record.json)

### Approaches

- [Thinking tokens](https://latent-implicit-thinking.github.io/topics/thinking-tokens/): Give a model extra tokens for computation, such as pauses, filler symbols, or learned planning tokens. These tokens need not spell out the solution.
- [Recurrent reasoning](https://latent-implicit-thinking.github.io/topics/recurrent-reasoning/): Repeat parts of the model or keep intermediate information in memory while working through a problem. Some methods also learn how many passes to run.

Summary and categories are editorial. Read the original paper to verify claims and limitations. A citation to a paper does not, by itself, establish that a later method builds on it.

Reading-list version 0.9.5; literature checked through 2026-10-18.


## Skip the Talk, Re-Focus on Vision: Latent Reasoning for Reasoning Segmentation in Multimodal Large Language Models

Authors: Tianhang Guo, Yulin He, Wei Chen, Wenjuan Zhou, Yuhang Li, Xinbiao Gan.

LIRSeg generates a fixed sequence of learned special tokens before a target description and bounding box used for segmentation. Spatial alignment grounds the tokens' hidden states in relevant image regions, followed by training with segmentation rewards.

First public: 2026-09-25.
ArXiv initial submission: 2026-09-25.

[Original paper](https://arxiv.org/abs/2609.30783) · [Reading-list entry](https://latent-implicit-thinking.github.io/papers/2609.30783/)
[Preprint BibTeX](https://latent-implicit-thinking.github.io/papers/2609.30783/citation.bib) · [Preprint CSL-JSON](https://latent-implicit-thinking.github.io/papers/2609.30783/citation.csl.json) · [JSON record](https://latent-implicit-thinking.github.io/papers/2609.30783/record.json)

### Approaches

- [Thinking tokens](https://latent-implicit-thinking.github.io/topics/thinking-tokens/): Give a model extra tokens for computation, such as pauses, filler symbols, or learned planning tokens. These tokens need not spell out the solution.

Summary and categories are editorial. Read the original paper to verify claims and limitations. A citation to a paper does not, by itself, establish that a later method builds on it.

Reading-list version 0.9.5; literature checked through 2026-10-18.


## Ceiling of a Task: When Can a Transformer Succeed Without Its Chain of Thought?

Authors: Jiashu He, Jinxuan Fan, Xiao Xiao, Radu Marculescu, Alejandro Ribeiro.

Analyzes tasks for which a fixed-depth transformer needs an informative chain of thought to exceed a task-specific accuracy limit. Combines circuit-based results with experiments that replace, erase, or shuffle reasoning tokens, including comparisons with filler tokens.

First public: 2026-09-27.
ArXiv initial submission: 2026-09-27.

[Original paper](https://arxiv.org/abs/2609.33134) · [Reading-list entry](https://latent-implicit-thinking.github.io/papers/2609.33134/)
[Preprint BibTeX](https://latent-implicit-thinking.github.io/papers/2609.33134/citation.bib) · [Preprint CSL-JSON](https://latent-implicit-thinking.github.io/papers/2609.33134/citation.csl.json) · [JSON record](https://latent-implicit-thinking.github.io/papers/2609.33134/record.json)

### Approaches

- [Thinking tokens](https://latent-implicit-thinking.github.io/topics/thinking-tokens/): Give a model extra tokens for computation, such as pauses, filler symbols, or learned planning tokens. These tokens need not spell out the solution.
- [Analysis and evaluation](https://latent-implicit-thinking.github.io/topics/implicit-reasoning-evaluation/): Measure how well these methods work, investigate what their internal computation does, and identify where they fail. This includes experiments, studies of model internals, and mathematical results.

Summary and categories are editorial. Read the original paper to verify claims and limitations. A citation to a paper does not, by itself, establish that a later method builds on it.

Reading-list version 0.9.5; literature checked through 2026-10-18.


## AbridgeKV: Compressing Long Reasoning into the KV Cache of Learned Pause Tokens

Authors: Jack Cai, Lijie Yang, Vasilis Kontonis, Dimitris Papailiopoulos, Shivam Garg.

Compresses each completed block of generated reasoning into the key/value cache of learned pause tokens, then discards the block's original cache. Later tokens attend to the compact replacement.

First public: 2026-10-07.
ArXiv initial submission: 2026-10-07.

[Original paper](https://arxiv.org/abs/2610.20641) · [Reading-list entry](https://latent-implicit-thinking.github.io/papers/2610.20641/)
[Preprint BibTeX](https://latent-implicit-thinking.github.io/papers/2610.20641/citation.bib) · [Preprint CSL-JSON](https://latent-implicit-thinking.github.io/papers/2610.20641/citation.csl.json) · [JSON record](https://latent-implicit-thinking.github.io/papers/2610.20641/record.json)

### Approaches

- [Thinking tokens](https://latent-implicit-thinking.github.io/topics/thinking-tokens/): Give a model extra tokens for computation, such as pauses, filler symbols, or learned planning tokens. These tokens need not spell out the solution.

Summary and categories are editorial. Read the original paper to verify claims and limitations. A citation to a paper does not, by itself, establish that a later method builds on it.

Reading-list version 0.9.5; literature checked through 2026-10-18.