Search - metacurate.io

GitHub - Gen-Verse/LatentMAS: [ICML 2026 Spotlight] Latent Collaboration in Multi-Agent Systems · GitHub

LatentMAS is a multi-agent reasoning framework that shifts collaboration into latent space for efficient, training-free alignment and faster reasoning. The project, its extensions, and related tools are showcased with an ICML 2026 spotlight, alongside complementary systems like OpenClaw-RL, Science-LatentMAS, and AVP.

6 mins read Listed 18 days ago in reddit/LocalLLaMA/OpenAI's major results over the last 3 months came from scaling test-time compute with agent swarms. I've been building an open-source equivalent for the past year, and I'd love for you guys to try it.

arxiv.org Tutorial

[2610.25296] World Model Interpretability: From Decoding Representations to Understanding Latent Worlds

The article argues that as JEPA latent states become internal predictors, interpretability should focus on interactions, evolution, and causal influence of latent variables. It proposes a four-stage agenda—Ground, Structure, Verify, Act—to move from decoding representations to understanding and acting on latent worlds, noting that current work provides early evidence for Ground but lacks systematic analysis and intervention studies.

1 min read Listed 28 days ago in Hugging Face Daily Papers

arxiv.org Tutorial

[2610.36791] RynnWorld-Latent: A Cross-Embodiment World Model Conditioned on Universal Latent Actions

RynnWorld-Latent introduces a cross-embodiment video world model that conditions on universal latent actions to learn controllable visual dynamics across human, robot, and simulated data. It pretrains on massive multi-domain video data, uses a 3D-grounded latent-action interface with a video transformer, and supports two data-generation routes to improve real-world policy performance.

1 min read Listed 19 days ago in Hugging Face Daily Papers

kevinkelly.substack.com Blog post

Latent Space as a New Medium - by Kevin Kelly - KK

The text argues that latent spaces inside AI models, especially large language models, act as a new creative medium. These spaces compress vast human knowledge into compact representations, storing abstract information rather than exact texts or images, and their power grows with more data and larger models. Language emerges as the core mechanism, enabling translation, pattern recognition, and imaginative use of stored knowledge.

20 mins read Listed 13 days ago in reddit/LocalLLaMA/Heretic 2.0 released: Model modification for everyone

arxiv.org Scientific research

[2611.09055] OmniLatent: Unified Latent for 3D Part Modeling

OmniLatent generates 3D assets with part labels from a single image by encoding geometry, appearance, deformation, and part IDs in one shared latent space. Its extensions support point-guided editing and completion of partial part masks into full-object decompositions.

1 min read Listed 9 days ago in Hugging Face Daily Papers

arxiv.org Scientific research

[2611.05088] LatentTriad: Pre-Action Risk Representations for Malicious Agent Skill Detection

LatentTriad adds a pre-action risk detector for malicious agent skills, using boundary and operation-direction features to screen skills without executing them; it achieves high accuracy across many skill packages and operates with a fast, execution-free process.

1 min read Listed 9 days ago in Bluesky

arxiv.org Scientific research

[2610.21182] Does Latent Chain-of-Thought Actually Reason? A Single, Unifying Taxonomy for Latent Reasoning

The study investigates whether latent chain-of-thought reflects genuine reasoning by testing various model setups, tasks, and training methods. It finds that readable hidden states do not reliably indicate reasoning, with evidence varying by task; a taxonomy is proposed to distinguish what a chain represents from how it is used.

2 mins read Listed 29 days ago in Bluesky

arxiv.org Scientific research

[2610.42149] Latent Reasoning in Continuous Space for Unified Multimodal Models

LARC introduces latent continuous-space reasoning to bridging token-based reasoning and image generation in multimodal models, using curriculum SFT and information-gain RL; it improves output quality and scaling, achieving state-of-the-art GenEval performance over BAGEL.

1 min read Listed 16 days ago in Hugging Face Daily Papers

arxiv.org Scientific research

[2611.00290] Quantifying Local and Global Fidelity of Nonlinear Embeddings

The paper introduces local and global fidelity metrics for evaluating dimension reduction, uses an LG reversal test to compare maps, and reports strong global alignment across many datasets, with specific crossovers in large-scale and retail data scenarios.

2 mins read Listed 11 days ago in Bluesky

enricobottazzi.github.io Tools and repositories

Interconcept Space

The document describes exploring the space between two concept vectors from a large language model, using linear interpolation and model-provided descriptions to interpret each intermediate vector.

1 min read Listed 15 days ago in reddit/learnmachinelearning/The interconcept space - what lies between two concepts? (According to an LLM)