Research papers, read to you in five minutes.
Research papers, read to you in five minutes.

Joint audio-video generation usually desyncs on lips and impacts because fusion is applied uniformly across tokens. DreamX-Creator learns a per-token, per-head sigmoid gate on cross-modal attention, and at 7B beats 10-15B open baselines on DeSync by ~40%, though it still trails 22-33B models on raw audio aesthetics.

Diffusion · Aug 29

GenFirst: Generation Before Reconstruction for Stable End-to-End Latent Generative Modeling

Joint VAE+generator training usually collapses, so everyone freezes the VAE and eats a latent space optimized for pixels, not sampling. GenFirst fixes it by splitting KL into prior-fit and entropy terms, then training generation before reconstruction — hitting gFID 0.97 on ImageNet 256 at 2× less compute than REPA-E.

Agents · Aug 28

LoopArena: Benchmarking Models as Runtime Controllers for Loop Engineering

LoopArena isolates the outer "manager" model in coding agents by freezing the Worker and swapping only the Controller that reads evidence and drafts the next instruction. The kicker: a Controller that just restates the goal each round scores 18.52% — identical to no manager at all. State-blind persistence isn't control.

Agents · Aug 28

Agentic Artifact Creation: Systems, Evaluation, Principles, and Opportunities

Agent quality isn't about smarter planners or more self-critique — it's whether the artifact's representation, edit interface, and verifier granularity line up. Across 230 systems, when any link breaks, agents fall back to whole-artifact regeneration. Precise diagnosis is wasted if your only edit is "regenerate."

Multimodal · Aug 27

Beyond Data Scaling: Representation-Centric Continued Pre-training for Vision-Language-Action Models

VLA backbones overfit to whichever action head you pretrain with, so swapping decoders or robots forces retraining. VLAct freezes lower layers and supervises three heads in parallel during continue-pretraining, yielding a backbone that beats full-data baselines on an unseen humanoid using just 20% of trajectories.

Never fall behind the literature again.

Free account. Follow topics, build your queue.

Sign up free

Audio/Speech · Aug 31

DreamX-Creator: Democratizing Native Audio-Video Generation at 2K Resolution

Joint audio-video generation usually desyncs on lips and impacts because fusion is applied uniformly across tokens. DreamX-Creator learns a per-token, per-head sigmoid gate on cross-modal attention, and at 7B beats 10-15B open baselines on DeSync by ~40%, though it still trails 22-33B models on raw audio aesthetics.

Diffusion · Aug 29

GenFirst: Generation Before Reconstruction for Stable End-to-End Latent Generative Modeling

Joint VAE+generator training usually collapses, so everyone freezes the VAE and eats a latent space optimized for pixels, not sampling. GenFirst fixes it by splitting KL into prior-fit and entropy terms, then training generation before reconstruction — hitting gFID 0.97 on ImageNet 256 at 2× less compute than REPA-E.

Agents · Aug 28

LoopArena: Benchmarking Models as Runtime Controllers for Loop Engineering

LoopArena isolates the outer "manager" model in coding agents by freezing the Worker and swapping only the Controller that reads evidence and drafts the next instruction. The kicker: a Controller that just restates the goal each round scores 18.52% — identical to no manager at all. State-blind persistence isn't control.

Agents · Aug 28

Agentic Artifact Creation: Systems, Evaluation, Principles, and Opportunities

Agent quality isn't about smarter planners or more self-critique — it's whether the artifact's representation, edit interface, and verifier granularity line up. Across 230 systems, when any link breaks, agents fall back to whole-artifact regeneration. Precise diagnosis is wasted if your only edit is "regenerate."

Multimodal · Aug 27

Beyond Data Scaling: Representation-Centric Continued Pre-training for Vision-Language-Action Models

VLA backbones overfit to whichever action head you pretrain with, so swapping decoders or robots forces retraining. VLAct freezes lower layers and supervises three heads in parallel during continue-pretraining, yielding a backbone that beats full-data baselines on an unseen humanoid using just 20% of trajectories.

Never fall behind the literature again.

Free account. Follow topics, build your queue.

Sign up free

Browse topics

Inference Optimization 42

Computer Vision 55

Agents 67

Reinforcement Learning 32

Multimodal 45

Reasoning 29

All topics →

Research papers, read to you in five minutes.Research papers, read to you in five minutes.

Research papers, read to you in five minutes.
Research papers, read to you in five minutes.