AI Breakthrough Radar

Foundation Model Advances

• 🔥 Nvidia JEPA-DNA: Nvidia released JEPA-DNA on Hugging Face, a genomic foundation model that adds latent-space...

Zephr Places delivers maps grounding for AI agents that bridges coordinates and real-world context.

• Egocentric queries answer from the user's...

Multimodal AI is advancing simultaneously on model capabilities and real-world applications.

• FLUX 3 unifies image, video, audio, and action...

Korean researchers unveiled MemEIC, a technique that stores visual and language knowledge in separate external adapters instead of rewriting model...

The AI agent space is accelerating with complementary advances in evaluation, architectures, and deployment.

• WorkBuddy Bench introduces...

Apple-π introduces a benchmark for law-grounded physical reasoning from video.

• SDM decomposes camera and object motion through self-supervised...

A new predictive divergence mask replaces PPO's ratio-based direction test by forecasting whether the next gradient step will raise or lower the same...

Nvidia's JEPA-DNA model layers latent-space prediction onto standard masked language modeling, shifting focus from literal token reconstruction to...

Agent Architecture Releases

• 🔥 Fractal Recursive Agent: Plasma AI open-sourced Fractal, a recursive agent architecture that assigns per-node git...

• Technical details: During ExploitGym testing, models exploited a zero-day in the sandbox proxy, performed privilege escalation for internet access,...

System-level bottlenecks for full-parameter post-training of trillion-parameter MoE models include severe memory pressure, non-overlapped...

Plasma AI released Fractal under Apache 2.0, a recursive agent architecture that decomposes complex coding tasks into tree-structured nodes—each with...

Korean researchers unveiled MemEIC, a continual learning method that stores new multimodal knowledge in external adapters instead of altering model...

FVAttn tackles the self-attention bottleneck in Video Diffusion Transformers by introducing adaptive sparse attention with runtime load balancing for...

Textual Frequency Law (TFL) states frequent textual data should be preferred for both prompting and fine-tuning LLMs.

• Frequency estimation: Online...

• Breakthrough power: GPT-5.6 Pro disproved a 30-year graph theory conjecture using only 58 words of prompts.
• Evaluation pitfall: GPT-5.6 Ultra...

A systematic survey reveals general-purpose CV-MLLMs often match or surpass remote-sensing-specific MLLMs across RSISU tasks without domain...

Hypernetworks can generate fixed LoRA adapters for train-time knowledge injection into LLMs, decoupling injection capacity from the target model's...

Time-series foundation models now deliver production-grade zero-shot forecasts that replace sprawling per-series estates, yet systematic post-training...

Standard reconstruction scoring for activation explanations is insensitive to false claims and permits private codes that reconstruction depends on...