# DiffusionNFT

3 related articles

Subscribe to updates

vLLM Blog Nov 6, 2026
VeRL-Omni v0.3.0 brings async execution to multimodal RL training: short answers still favor sync, long ones favor separate-async
VeRL-Omni v0.3.0 extends asynchronous execution to multimodal reinforcement-learning training across text, audio input and diffusion models, and makes the V1 trainer built on TransferQueue and ReplayBuffer the default. Previously a strictly alternating loop left generation and training idle in turn. In MiniCPM-o 4.5 audio-visual question answering, separate-async ran about 9% slower per prompt on natural short answers but about 6.7% faster per prompt at forced 1024-token responses, so the execution mode should follow the workload.
VeRL-Omni v0.3.0 extends asynchronous execution to multimodal reinforcement-learning training across text, audio input and diffusion models, and makes the V1 trainer built on TransferQueue and ReplayBuffer the default. Previously a strictly alternating loop left generation and training idle in turn. In MiniCPM-o 4.5 audio-visual question answering, separate-async ran about 9% slower per prompt on natural short answers but about 6.7% faster per prompt at forced 1024-token responses, so the execution mode should follow the workload.
VeRL-Omni Asynchronous Rollout DiffusionNFT Group Sequence Policy Optimization Post-training
arXiv Oct 12, 2026
One RNA design policy now serves many backbones: pairing MCC climbs from 0.61 to 0.73
Structure-conditioned RNA inverse design previously either trained a separate policy per target backbone or relied on offline preference optimization, leaving online post-training of a reusable backbone-conditioned model largely unexplored. MORAD maps six design measurements to bounded desirability scores, combines them with a weighted geometric mean, and uses group-relative normalization to drive a forward-process diffusion update. On 153 evaluation targets it raises GDT-TS from 0.3300 to 0.3827 and pairing MCC from 0.6110 to 0.7295, while cutting normalized ensemble defect from 0.3906 to 0.2830.
Structure-conditioned RNA inverse design previously either trained a separate policy per target backbone or relied on offline preference optimization, leaving online post-training of a reusable backbone-conditioned model largely unexplored. MORAD maps six design measurements to bounded desirability scores, combines them with a weighted geometric mean, and uses group-relative normalization to drive a forward-process diffusion update. On 153 evaluation targets it raises GDT-TS from 0.3300 to 0.3827 and pairing MCC from 0.6110 to 0.7295, while cutting normalized ensemble defect from 0.3906 to 0.2830.
Fudan University
Jixin Tang（First author）
MORAD RNA Inverse Folding Reinforcement Learning DiffusionNFT Reward Design
arXiv Oct 2, 2026
Tencent team proposes Adaptive Reward Routing, taking best result on nine of ten metrics across LTX-2 and LTX-2.3
The work proposes Adaptive Reward Routing, which jointly adapts where reward-driven updates act and how competing rewards are coordinated during forward-process RL (DiffusionNFT) of joint audio-video diffusion models: it uses bidirectional cross-attention responses as a proxy for evolving cross-modal influence to dynamically reweight token-aware losses and scale gradients across cross-modal layers, and after warm-up uses branch-specific reward-gradient interactions as residual corrections to predefined preference weights, training on 19,487 audio-video prompts over the LTX-2 (19B) and LTX-2.3 (22B) backbones and achieving the best result on nine of ten metrics on the 10,140 prompts of JavisBench.
The work proposes Adaptive Reward Routing, which jointly adapts where reward-driven updates act and how competing rewards are coordinated during forward-process RL (DiffusionNFT) of joint audio-video diffusion models: it uses bidirectional cross-attention responses as a proxy for evolving cross-modal influence to dynamically reweight token-aware losses and scale gradients across cross-modal layers, and after warm-up uses branch-specific reward-gradient interactions as residual corrections to predefined preference weights, training on 19,487 audio-video prompts over the LTX-2 (19B) and LTX-2.3 (22B) backbones and achieving the best result on nine of ten metrics on the 10,140 prompts of JavisBench.
Tencent (China)Hong Kong University of Science and Technology University of Hong Kong
Songlin Yang（First author）
Adaptive Reward Routing Multi-Reward Optimization DiffusionNFT Joint Audio-Video Diffusion Models LTX-2

You’re all caught up

VeRL-Omni v0.3.0 brings async execution to multimodal RL training: short answers still favor sync, long ones favor separate-async

One RNA design policy now serves many backbones: pairing MCC climbs from 0.61 to 0.73

Tencent team proposes Adaptive Reward Routing, taking best result on nine of ten metrics across LTX-2 and LTX-2.3