Streaming video models drift within minutes because training sees clean history while inference sees noisy, error-accumulated history. Vidu S2's fix, Self-Replay Forcing, detaches a full rollout, re-noises every segment, then backprops across segments in one pass — hitting 720p at 25–42 FPS with AVSync 0.617 vs 0.855.

Agents · Sep 10

Scaling Automatic Research Agents via World Models

RL for code-writing agents burns 1000+ GPU-hours because every trajectory needs sandboxed execution. WMRL replaces most executions with an LLM predicting the reward, keeps 10% real runs as anchors for isotonic debiasing plus inverse-variance fusion, and matches GRPO at 3–4× less compute — degrading gracefully when predictions fail.

Evaluation · Sep 10

Mi-Ripple: Restoring Images Degraded by Iterative AI Editing

Iterative reference-based image editing accumulates grid and grain artifacts that prompts can't fix and generic denoising destroys. Mi-Ripple diagnoses first—spectral peaks versus content-entangled granularity—then routes to notch filtering, masked suppression, or regeneration from a cleaned reference, keeping high-frequency retention above 98.9%.

Agents · Sep 9

Programmable World Model

Interactive video world models forget: kill an NPC and it walks back on screen. The fix is architectural — a rule engine owns state, a frozen video backbone just renders from projected 3D bounding boxes plus identity/direction overlays. Hits 98% post-death state accuracy versus 8% for prompt-swap baselines.

Inference Optimization · Sep 9

NCP-ArchPreview Technical Report: Moving towards Latent Space Language Models through Next Concept Prediction

NCP-ArchPreview attacks the fact that LLM pretraining only supervises surface tokens, never higher-level structure. Adding a product-quantized "next concept" prediction objective alongside NTP, an 8.9B model matches OLMo-3-7B's final loss on 51.3% of tokens, and enables 17M-param domain adaptation that beats LoRA without forgetting.

Never fall behind the literature again.

Free account. Follow topics, build your queue.

Sign up free

Inference Optimization · Sep 10

Vidu S2: Real-Time Interactive, Editable, and Spatial Video Generation

Streaming video models drift within minutes because training sees clean history while inference sees noisy, error-accumulated history. Vidu S2's fix, Self-Replay Forcing, detaches a full rollout, re-noises every segment, then backprops across segments in one pass — hitting 720p at 25–42 FPS with AVSync 0.617 vs 0.855.

Agents · Sep 10

Scaling Automatic Research Agents via World Models

RL for code-writing agents burns 1000+ GPU-hours because every trajectory needs sandboxed execution. WMRL replaces most executions with an LLM predicting the reward, keeps 10% real runs as anchors for isotonic debiasing plus inverse-variance fusion, and matches GRPO at 3–4× less compute — degrading gracefully when predictions fail.

Evaluation · Sep 10

Mi-Ripple: Restoring Images Degraded by Iterative AI Editing

Iterative reference-based image editing accumulates grid and grain artifacts that prompts can't fix and generic denoising destroys. Mi-Ripple diagnoses first—spectral peaks versus content-entangled granularity—then routes to notch filtering, masked suppression, or regeneration from a cleaned reference, keeping high-frequency retention above 98.9%.

Agents · Sep 9

Programmable World Model

Interactive video world models forget: kill an NPC and it walks back on screen. The fix is architectural — a rule engine owns state, a frozen video backbone just renders from projected 3D bounding boxes plus identity/direction overlays. Hits 98% post-death state accuracy versus 8% for prompt-swap baselines.

Inference Optimization · Sep 9

NCP-ArchPreview Technical Report: Moving towards Latent Space Language Models through Next Concept Prediction

NCP-ArchPreview attacks the fact that LLM pretraining only supervises surface tokens, never higher-level structure. Adding a product-quantized "next concept" prediction objective alongside NTP, an 8.9B model matches OLMo-3-7B's final loss on 51.3% of tokens, and enables 17M-param domain adaptation that beats LoRA without forgetting.

Never fall behind the literature again.

Free account. Follow topics, build your queue.

Sign up free

Browse topics

Multimodal 77

Image Generation 32

Inference Optimization 80

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.