[ICML 2026 Spotlight] Latent Collaboration in Multi-Agent Systems

📖 This is a repository for organizing papers, codes, and other resources related to Latent Reasoning.

[ECCV 2026] Official repo of "Chain-of-Visual-Thought: Teaching VLMs to See and Think Better with Continuous Visual Tokens"

Official implementation of the NeurIPS 2025 paper "Soft Thinking: Unlocking the Reasoning Potential of LLMs in Continuous Concept Space"

[CVPR2026] Official codebase for the paper "Reasoning Within the Mind: Dynamic Multimodal Interleaving in Latent Space"

Official implementation of Latent-SFT: teaching LLMs to reason with vocabulary-space latent chains.

Official PyTorch implementation of "Latent Reasoning in TRMs is Secretly a Policy Improvement Operator" (ICML 2026)

Imagine Before You Predict: Interleaved Latent Visual Reasoning for Video Event Prediction

An open-source attempt at training a variant of LeCun's energy-based models (EBM) to reason in latent space and solve Sudoku.

Official implementation of Latent-GRPO: reinforcement learning for vocabulary-space latent reasoning.

Repository to train multimodal latent reasoning tokens for Qwen 2.5 VL.

🔥 Must-read papers for LLM-based Latent Reasoning

LaTER: Efficient Test-Time Reasoning via Latent Exploration and Explicit Verification

This is the codebase for "Large Reasoning Models Are (Not Yet) Multilingual Latent Reasoners"

Official Implementation for Pre-print "T2MLR: Transformer with Temporal Middle-Layer Recurrence"

Training-free entropy-gated explicit<->latent (SwiReasoning) decoding for llama.cpp — bit-exact, Pareto-superior on a 27B

Latent reasoning with LLMs using abstract thought tokens

Curated Awesome List for latent world models, JEPA-style predictive architectures, latent reasoning, and latent planning/control

Lightweight pairwise evaluator for relational signals in Ouro-2.6B-Thinking loop-state trajectories.

Implemented a recurrent-depth LLM (PyTorch) based on arXiv:2502.05171. Demonstrated that scaling inference compute increased arithmetic reasoning accuracy from 8% to 100% without additional parameters.