# Lightweight Latent Reasoning for Narrative Tasks

Authors: Alexander Gurung, Esmeralda S. Whitammer, Mirella Lapata.

Trains a small projector to replace parts of a written reasoning trace with continuous embeddings. Reinforcement learning teaches the language model when to use these latent steps for plot-hole detection and story generation.

First public: 2025-12-01.
ArXiv initial submission: 2025-12-01.

[Original paper](https://arxiv.org/abs/2512.02240) · [PDF](https://arxiv.org/pdf/2512.02240) · [HTML](https://arxiv.org/html/2512.02240v2) · [Reading-list entry](https://latent-implicit-thinking.github.io/papers/2512.02240/)
[Preprint BibTeX](https://latent-implicit-thinking.github.io/papers/2512.02240/citation.bib) · [Preprint CSL-JSON](https://latent-implicit-thinking.github.io/papers/2512.02240/citation.csl.json) · [JSON record](https://latent-implicit-thinking.github.io/papers/2512.02240/record.json)

## Published version

Lightweight Latent Reasoning for Narrative Tasks. Alexander Gurung, Esmeralda S. Whitammer, Mirella Lapata. Transactions of the Association for Computational Linguistics, 2026.
[Publication](https://aclanthology.org/2026.tacl-1.98/) · [BibTeX](https://latent-implicit-thinking.github.io/papers/2512.02240/publication.bib) · [CSL-JSON](https://latent-implicit-thinking.github.io/papers/2512.02240/publication.csl.json)
[Publication record](https://aclanthology.org/2026.tacl-1.98/).

## 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.
- [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.

Reading-list version 0.15.2; literature checked through 2026-11-03.