Tag
CausalWAM proposes adding a sparse routing path to a pretrained World Action Model's action head so that generated actions depend only on action-relevant latent variables, with provable identification of causal structure up to permutation without temporal independence assumptions. Experiments on synthetic data and real robots show improved closed-loop success efficiency.
LAMB is a recurrent framework for latent multi-hypothesis beliefs in long-horizon LLM agents, factorizing memory into a shared Common state and weighted Delta components, achieving notable gains over PABU on ALFWorld, ScienceWorld, and Crafter.
该论文研究了条件扩散模型中类几何(class geometry)对 classifier-free guidance 性能的影响，提出轻量的 Prototype Regularization 方法，在 ImageNet 上将引导 gFID 从 3.31 降至 2.28，超越了 DINOv3 对齐的 VA-VAE。
The paper proposes Predictive-State Flows (PSF), a continuous diffusion language model design where a rectified-flow prior generates block-level predictive states and a block-causal decoder generates tokens in parallel, significantly reducing diffusion states and sequential depth while improving efficiency and generative quality over models like ELF-B.
这篇论文研究生成模型在何种条件下可以充当可编辑的模拟器，提出通过探测器解码与逆映射两种方式将模型隐状态映射到模拟器状态，并在Othello与自建的连续世界Rayworld中发现映射成功与否强烈依赖环境，即使状态始终可解码。
该论文提出将 GUI 世界模型的预测目标从全屏像素重建转移到潜空间中的动作诱导未来信息，通过共享的多模态模型表示配合动作头与转移头，并引入动作感知对比目标来缓解身份捷径问题，相比像素空间世界模型最高实现 31.9% 和 40.6% 的成功率相对提升，并将推理延迟降低最多 98.9%。
The paper introduces Directed Temporal Representations for Control (DTRC), which learns a directed temporal quasimetric over frozen LeWorldModel features from offline visual trajectories to serve as a goal-relative progress signal for direct goal-conditioned policy learning, achieving strong control performance across ten visual tasks.
RoboJEPA 基于 JEPA 架构，在涵盖 12 种机器人具身的大型真实机器人数据集上训练，首次为多具身机器人世界模型建立了 scaling law：其 latent rollout 的“想象力误差”随计算量呈二阶幂律变化，并可作为真实机器人规划性能的可靠代理。该模型参数量达 8B，是迄今最大的 JEPA 预测模型，支持零样本部署为长时程规划的机器人智能体，并开源全部检查点与训练/部署代码。
This paper investigates whether clinical concepts exist as causally used, locatable representations inside open-weight LLMs by applying dictionary learning over clinical notes with ICD billing codes as proxies. The authors identify clinical concept centers across eleven open models, show they drive behavior under adversarial and aligned priming, demonstrate steering-based performance gains, and validate them via blinded clinician preference testing.
ATLAS is a training objective for latent world models that preserves relational geometry from encoder representations into the planning latent while calibrating its marginal distribution via Wasserstein-2 transport (WEMReg). Instantiated in LeWM, it improves goal-reaching success on PushT, TwoRoom, and OGBench-Cube benchmarks, particularly on higher-novelty episodes.
提出了 DC-SAE（解耦紧凑语义自编码器），在保持高保真重建的同时实现 32 倍空间压缩并加速扩散模型训练收敛，显著超越此前 SOTA 高压缩分词器 DC-AE。
This paper proposes Task-Conditioned Latent Alignment (TCLA) to stabilize neural decoding across sessions in brain-machine interfaces by learning a shared latent space. Evaluated on nonhuman primate data, TCLA shows improved robustness compared to existing methods.
BLARM is a feed-forward method for animating 3D meshes from monocular video by blending latent rigid motion primitives, enabling temporally coherent animations without explicit rigs.
This paper proposes Intention Distillation (INDI) to distill behavior intent into the action decoder of Vision-Language-Action models, improving performance on benchmarks like SimplerEnv-Bridge and real-world tasks.
The paper introduces 'Kilobyte models', an extreme compression paradigm where neural network weights are regenerated from a seed and a quantized latent vector, reducing storage to kilobytes by leveraging Mapping Networks and seeded block-wise bases.
AURORA-LM introduces a continuous-latent diffusion language model that separates decodable text representation from distribution modeling, achieving strong performance on OpenWebText and XSum while scaling to 1B parameters.
MUGEN introduces a unified motion-language framework that avoids discrete codebooks and iterative decoding, using a single adaptive-length autoencoder with continuous latent slots and one-shot generation to achieve efficient, high-quality text-to-motion and motion-to-text performance across HumanML3D and SnapMoGen benchmarks.
LATO.2 is a factorized flow matching framework for 3D mesh generation that decomposes the process into a vertex flow followed by a connectivity flow conditioned on the vertices, achieving state-of-the-art geometric fidelity and connectivity quality.
Identifies two coupled causes of posterior collapse in VAEs and introduces λ-VAE, a modification to the reparameterization step that equalizes variance across latent dimensions, reducing collapse and improving information capacity.
AdaJEPA introduces an adaptive latent world model that continuously updates during test-time via closed-loop model predictive control, significantly improving planning success under distribution shift.