Articles & Guides

DiffusionGemma is an experimental open-weight language model that uses discrete diffusion to generate text in parallel blocks rather than token by token. Built by fine-tuning Gemma 4, it explores whether diffusion-style decoding can deliver a better speed-capability trade-off for LLMs.

WCM targets a core bottleneck in robotic VLA post-training: critics that rely on single-frame observations struggle under partial observability. By jointly predicting future latent states and values, it trains the critic to encode temporal dynamics rather than only regress scalar returns.

InfiniSplat targets large-baseline novel view synthesis from a single image by replacing fixed pixel-aligned Gaussian prediction with geometry-guided support sampling and implicit Gaussian decoding.

GradCuit introduces a test-time latent optimization method that inserts trainable continuous states into an intermediate Transformer layer. By routing reward-weighted gradients from the generated continuation back to those states, it aims to make latent reasoning more robust and interpretable.

Motion Beyond Morphology (MBM) reframes video motion transfer as a problem of preserving transferable dynamics rather than matching fixed structures. The framework aims to animate targets whose morphology may differ substantially from the reference video.

A new study argues that many AI-generated patches are not merge-ready because models often keep code that should be removed.

GPTQ-2D addresses a more general form of adaptive matrix rounding where basis matrices act on both sides of the residual. Instead of vectorizing the problem and paying quartic cost, it preserves the same rounded output with a cubic-time anti-diagonal schedule.

Roomer reframes residual errors in generated indoor layouts as object-grounded repair tasks, identifying the items responsible for collisions, blocked openings, out-of-bounds placements, and unusable circulation.

DeepVoyager-VL aims to move multimodal search beyond one-shot image understanding. It treats visual evidence as an active driver of intermediate reasoning, retrieval, and agent decisions.

N₀-TWAM is a tactile-native world-action model for contact-rich manipulation, designed to predict not only future vision but also future contact before generating actions.

Weak-to-Strong On-Policy Distillation proposes a way to improve a strong LLM without relying on an even stronger teacher. The method builds a proxy teacher from contrasts between smaller models in logit space.

Mental World Modeling argues that a world model can correctly reconstruct a physical scene and still fail at predicting human action unless it tracks hidden mental states.

A new paper frames cross-session personalized ambiguity adaptation as a benchmark task for coding assistants. The goal is to see whether models can use a user’s resolved history to write the intended code with fewer repeated clarifications.

SAF-OPD tackles a subtle failure mode in LLM post-training: reinforcement learning with verifiable rewards and on-policy distillation are complementary, but a fixed mixture can collapse entropy and suppress exploration. The proposed SAF framework regulates when and how strongly the teacher signal is used.

Rubric-based reinforcement learning can improve open-ended LLM behavior, but its feedback can disappear in subtle ways. CriPO addresses both unexplored and suppressed criteria without relying on external guidance at inference time.

EMBL AI Librarian reframes Europe PMC as a natural-language evidence interface for AI agents. Instead of returning only paper lists, it plans searches, reads selected papers, and surfaces snippets that support an answer.

ODEWorld moves beyond fixed-step prediction by learning a continuous latent velocity field governed by physical time. The approach targets long-horizon visual prediction, flexible temporal resolution, backward prediction, and planning-friendly dynamics for robotics.

RL²-VLA introduces an inference-time steering framework for VLA models that activates reinforcement-learning guidance only when the base policy is likely to fail.

A new paper introduces Collective-State JEPA, a decentralized predictive architecture that lets each robot in a swarm form a representation of the same future collective state. The results highlight label efficiency, topology transfer, and planning-relevant value estimation.

A new study tests whether value-based content inserted during midtraining can make alignment more durable than post-training alone. The results suggest that the presence of constitutional content matters more than elaborate curriculum structure.

SULAND v2 rebuilds a public RGB surface-landmine benchmark by correcting annotation errors and an inverted OOD class-ID convention. The results show that strong in-distribution scores do not necessarily translate into deployment robustness.