AI Research Pulse

Key Research Papers

• RepWAM: Introduces Representation Visual-Action Tokenizers for world action modeling, inducing latent actions as transitions...

Noam Shazeer's move to OpenAI as Lead for Architecture Research underscores the intensifying talent war at the frontier, where even a $2.7B licensing...

World model research is accelerating across robustness, robotics, and capital.

• Adversarial exposure: BadWorld reveals catastrophic failures in...

• MoE architectures like Mixtral activate just 12.9B of 46.7B parameters per token, slashing VRAM needs for consumer GPUs.
• SubQ's SSA delivers...

Sovereign AI infrastructure funding is accelerating alongside a fundamental rebuild of data architectures for AI workloads.

• CNTXT AI secured $60M...

Two fresh RL techniques are reshaping generative models:

• A2D2 enables reward-guided fine-tuning of discrete diffusion models, letting them...

PixelRAG renders pages as images, tiles them, and embeds via vision models for direct VLM queries—bypassing text extraction that causes over a third...

FPRM tackles signal degradation in looped transformers through pre-norm layers and residual scaling, enabling stable deep iterations while using...

Open-Source Model Releases

• 🔥 GLM-5.2: Z.ai introduces GLM-5.2, an open-source LLM with 1M token context window for long-horizon tasks and...

Fresh capital flows into India's AI data centers while OpenAI's Stargate project encounters costly setbacks.

• India attracts funding: Canadian...

• China enables IPOs on the STAR Market for large-model firms and future-industry startups to fund long-term R&D.
• Poland makes a direct $11M...

Two novel RL methods advance generative models through targeted constraint innovations:

• Flow-DPPO swaps ratio clipping for a divergence proximal...

SpaceX's $60 billion acquisition of AI coding startup Cursor marks a significant step in integrating advanced AI tools into aerospace operations.

Two new models highlight how LLMs are optimizing for specific coding workflows rather than general performance.

• LoopCoder-v2 uses a two-loop...

• DeLM lets multiple AI agents collaborate asynchronously via shared task queues and verified context, ditching central orchestrators.
• It delivered...

Larger models develop more Rosetta neurons with recurring activation patterns across independent trainings, yet their count grows only sublinearly...

• UniAR unifies image understanding, generation, editing and reasoning in a single model via one visual tokenizer and shared representation.
• It...

TDV replaces hand-crafted augmentations in self-supervised visual learning with a minimal causal assumption: the past causes the future. By jointly...

Researchers embedded core computer-vision operations into a light-manipulating optical metasurface, creating a prototype that delivers accurate real-time perception across tasks with minimal energy use.

Lakshay Sharma's Sub-Image Overlap Prediction method delivers competitive semantic segmentation results in remote sensing using self-supervised pretraining on just thousands of images, making advanced vision models accessible with only a single GPU.