Articles & Guides

AlayaRenderer-Flash accelerates generative world rendering from 0.56 FPS to 31.54 FPS, making a physics-driven generative scene playable at around 30 FPS. Instead of inventing dynamics from prompts, it renders RGB frames from structured world states.

The paper tackles a major but often overlooked memory cost in MoE training: optimizer state. SkewAdam assigns different state formats to the backbone, experts, and router, sharply reducing memory without sacrificing validation performance in the reported experiment.

A new paper argues that structural text template tokens in text-to-image Diffusion Transformers are not passive placeholders. They become implicit semantic registers that help preserve object identity during generation.

The paper argues that RLVR does not substantially rewrite a model’s singular-value spectrum. Instead, the behavioral gains appear to live in the input and output singular frames, motivating a fixed-spectrum optimization stack called ISO.

This paper investigates whether diffusion language models internally track denoising progress. The authors find that a latent time-like signal can be decoded from residual streams even without explicit timestep conditioning.

AlayaWorld aims to turn text, images, or videos into explorable and evolving virtual environments. Its technical focus is not only visual quality, but also interaction, long-horizon consistency, and efficient response.

ABot-World-0 frames video generation as an action-conditioned world model that can run in a closed interactive loop. Its main contribution is not only visual synthesis, but the combination of data curation, long-horizon distillation, and streaming inference on a single RTX 5090.

SciForma reframes scientific diagram generation around structural fidelity rather than surface-level visual plausibility. It decomposes correctness into components, arrows, and text, then trains models to satisfy all three at once.

AutoIndex argues that retrieval quality can be improved before the retriever ever runs: by learning executable programs that transform documents prior to indexing. With BM25 kept fixed, the framework improves performance across all eight CRUMB retrieval tasks.

Mage-Flow argues that competitive image generation does not have to rely only on ever-larger parameter counts. Its 4B-scale stack combines an efficient latent tokenizer, a native-resolution diffusion transformer, and system-level CUDA optimizations.

This paper rethinks post-training for open-ended tasks by upgrading judge-style feedback into coach-style experiential knowledge. Instead of compressing everything into a scalar reward, it keeps richer guidance alive in the learning loop.

NexForge tackles a core bottleneck in agent post-training: scalable executable data. Instead of building tasks around fixed tools or repositories, it starts from high-level capability requirements and generates runnable tasks plus expert trajectories.

FlowMimic tackles a common bottleneck in video editing: expensive data collection and limited task diversity. Its key idea is to transform image editing samples into video editing samples in real time, while aligning image and video capabilities through mutual mimicry.

GigaAM Multilingual targets a problem many multilingual ASR systems still fail to solve: uneven performance for long-tail languages. Its main contribution is not only scale, but also training design that keeps dominant languages from overwhelming the rest.

SeerGuard tackles a key weakness of mobile GUI agents: one bad action can cause irreversible damage. Its answer is to screen instructions and assess action risk before anything reaches the device.

Rather than scaling a general model, this paper builds a training ground where search agents can check their own progress and recover from mistakes. The core idea is self-improvement through verifiable interaction traces.

GEPO is a lightweight extension to GRPO that adjusts advantage signals using group entropy. The goal is to better handle mixed-task RL, where different prompts can require different exploration levels.

Long context is useful only when the right information survives. SWE-Pruner Pro argues that a coding agent already has enough internal signal to decide what tool output should stay or go.

DiFA is a training-free inference framework for diffusion models that treats intermediate predictions as correlated observations rather than disposable integration outputs. By building a forward-aligned temporal consensus, it aims to improve stability and preserve fine details during sampling.

RynnBrain 1.1 is built to connect perception, spatial reasoning, localization, and planning in one embodied framework. The key shift is not just better understanding, but outputs that are more directly useful for robot manipulation.

API-calling agents usually need lots of high-quality trajectories, but building full executable environments is expensive and slow. This paper shows that LLMs can simulate stateful API interactions from specs alone.