Articles & Guides

The paper argues that masked diffusion language models can serve as stronger and more steerable text-based world models for agent training. By denoising bidirectionally, MDLMs better preserve global anchors such as tool schemas, prior turns, rules, and intended outcomes.

Open-AoE is more than a dataset release. It packages collection, structuring, and downstream adaptation into an open pipeline for embodied learning.

Distilled Reinforcement Learning proposes a middle path between outcome-only RL and unconditional online policy distillation. It uses a teacher model to redistribute token-level policy-gradient signals while keeping optimization anchored to rewards.

This paper targets one of the hardest problems in long-audio understanding: temporal grounding. The model can answer questions with explicit timestamps over inputs as long as 120 minutes. Its key idea is to interleave periodic time markers with continuous audio tokens and train with large-scale synthetic supervision.

HarmoHOI tackles a core bottleneck in hand-object interaction synthesis: keeping multiple views visually consistent while preserving plausible 3D motion. It does so by jointly modeling RGB videos and aligned 3D point tracks in one diffusion framework.

This paper proposes TOPL, a training paradigm that does not directly imitate off-policy outputs. Instead, it teaches the model to classify whether each token is correct, which helps improve faithful generation under distribution shift.

TimeLens2 reframes video temporal grounding as an interval-set prediction problem, enabling multimodal models to identify not only what happens in a video but when the supporting evidence appears. Its 2B, 4B, and 8B variants show substantial gains over their Qwen3-VL backbones.

A new benchmark, 3D-Fit, evaluates whether general-purpose LLMs can generate 3D ligands under realistic spatial constraints for structure-based drug design. The results suggest early promise, but also expose a clear gap versus specialized diffusion models.

EvolvingWorld moves interactive literary simulation beyond static persona imitation. It treats story generation as a long-horizon process in which characters, scenes, and world state keep changing together.

ReflectWorld-MM is designed to help AI systems remember open-ended video streams the way humans do: by tracking persistent people, objects, and changes over time. Its main shift is from frame-based storage to entity-oriented memory.

HOMIE targets human-object centric video personalization, aiming to preserve subject identity while improving interaction accuracy. It integrates multimodal model knowledge more effectively into video generation.

Apple-PI asks a deeper question than visual realism: does a video model actually follow physical laws when it generates? The benchmark evaluates not only outcomes, but also the reasoning path behind them.

DiffGI targets a long-standing weakness of many 3D generative models: representing thin-shell and non-manifold surfaces such as garments. By combining continuous 2D TSDF maps with differentiable Marching Squares, it links compact 2D latent representations directly to 3D surface optimization.

JoyNexus proposes a service-oriented framework for post-training Vision-Language-Action models, aiming to make supervised fine-tuning, reinforcement learning, and evaluation more efficient under shared infrastructure.

A study of 1.02 million reviewed pull requests across 207 GitHub projects finds that AI agent involvement is linked to faster review decisions, but not consistently better review quality.

The paper introduces Contrastive Policy Optimization, a method that replaces entropy-only advantage shaping in RLVR with a token-level signal based on distributional disagreement. The goal is to distinguish useful uncertainty from harmful confusion during model training.

SVR-R1 brings inference-time self-checking into the reinforcement learning loop. The model answers, judges its own response with a Yes/No verdict, and rethinks when it says No.

RecGPT-V3 presents an industrial LLM-based recommender that maintains evolving user memory, grounds intent with Semantic IDs, and compresses reasoning into latent tokens. In Taobao’s “Guess What You Like” feed, the report cites higher engagement and a 52.4% reduction in serving resource consumption.

A new benchmark shifts plasma diagnostic ML evaluation from clean sensor data to realistic failure conditions. The results show that sequence models can break down sharply when sensors fail near a disruption.

Agon proposes a competitive cross-model RL setup in which two comparable but behaviorally different models train by trying to outperform each other on the same problem.

RESOURCE2SKILL proposes a way to distill videos, repositories, articles, and reference artifacts into reusable skills that software agents can retrieve and execute.