AI Research Radar

• 4Director conditions generation on explicit 4D rigid meshes and per-frame transformations for precise 3D camera/object control, replacing coarse...

Representation spaces are emerging as the core limitation for continuous diffusion language models.

• Stronger embeddings boost performance yet...

The viral "AI models creating children" claim correctly describes neuroevolution: a real, gradient-free method that evolves neural networks via...

Memory from interaction history cannot ensure a coherent understanding of the current world in LLM agents, often causing Belief Trapping where...

DARA applies an inverse-square-root density correction to reward aggregation in multi-objective RL, dynamically up-weighting sparse signals based on...

SETA's NeurIPS 2026 acceptance includes full training scripts for DeepSeek-V4-Flash, GLM-5.2, and Qwen3.8-27B using both GRPO and PPO. The release of...

A surrogate-constrained deep learning framework is proposed to address limitations in simulation-driven inverse design for engineering systems.

GraphForge tackles the core missing infrastructure for training agents that must navigate real files and tools: it builds evidence graphs from...

AREION introduces a foundation model for atlas-agnostic fMRI representation learning, extending large-scale pretraining to FC-based models that...

APPL embeds each skill policy's structural prior directly into the composition interface, closing the information gap that previously blocked...

Gacha Decoding generates genuinely diverse LM outputs by combining instruction following with external RNG instead of token-level entropy. The...

• Open weights edge: Clef models released under Apache-2.0 on Hugging Face for local download, unlike Jev's closed architecture.
• Capability gains:...

SILSA introduces compact sliding-window slice latents that replace expensive voxel tokens, enabling single-stage rectified-flow generation while...

The Institute of Foundation Models at MBZUAI is hiring research scientists with PhD credentials, first-author papers at NeurIPS/ICML/ICLR/CVPR, and...

The rapid rise of AI agents presents urgent challenges in authentication, authorization, and identity management that exceed conventional user-centric...

PixelDense routes DINOv2/SAM2 through a semantic stream and Depth Anything v2/Metric3D v2 through a geometric stream, adding a weight-space...

• AutoDataBench holds training frameworks and compute fixed to isolate data intelligence, letting researchers attribute gains to specific capabilities...

GRPO assigns identical advantage to every token in a trajectory, diluting the signal for decisive steps in long-horizon tasks. ProVer instead uses an...

A 27B agent trained on verifiable ML and algorithmic tasks converts extra test-time rounds into stronger solutions, reaching 81.8 on MLE-bench Lite and 70.7 on Frontier-CS while transferring to deep research.

BeyondSCe is a zero-shot system that lets robots retrieve objects by their role in past interactions—even when hidden—by combining event history with...