AI Research Radar

New Agent Evaluation Benchmarks

• SWE-Skills-Bench: SWE-Skills-Bench questions whether agent skills help in real-world software engineering...

• Teleop data is expensive and hard to scale—enter simulation 🖥️, human videos 🎥, AC-WMs 🤖, and WAMs 🌍.
• #ICRA2026 workshop tackles how these...

InCoder-32B advances industrial AI with a 32B-parameter model for chip design and GPU kernel optimization.

Key breakthroughs:

• Bridges general LLMs...

Emerging infra blends decentralized discovery with vulnerability safeguards:

• ScienceClaw coordinates independent agents for scientific tasks via...

• LLM evals are expensive, so sub-sampling eval data is common to cut costs and boost efficiency in tuning/ablation experiments.
• Wrong sampling approaches lead to noisy or incorrect results.
• Prioritize rank correlation for reliable subsampling.

• Agentic split: Most AI tools merge planning and execution; Capy uses specialized captain (planning) and build (execution) agents.
• Quality edge:...

WorldCam enhances interactive 3D gaming with video diffusion transformers augmented by camera pose representation, enabling precise action control and long-term 3D consistency.

M^3 advances monocular SLAM:

• Integrates dense matching with multi-view foundation models
• Targets Gaussian splatting SLAM for multimodal 3D reconstruction
  Join the discussion to explore sim-to-real implications.

Emerging trend in AI research: Benchmarks exposing gaps in LLM agent performance.

• FinToolBench evaluates agents on real-world financial tool use
  -...

New paper outlines a cognitive framework for measuring AGI progress, quickly gaining traction with 58 points on Hacker News. Explores milestones beyond pure scaling.

• MiroThinker-1.7 & H1 targets heavy-duty research agents via verification.
• One-Eval enables automated, traceable LLM evaluation with agentic systems.
• Emerging pattern: Verification primitives boost reproducible agent reasoning for AI self-assessment.

Efficient Sequence Modeling

• 🔥 Mamba-3: Mamba-3 is a new state space model (SSM) by Together AI designed with inference efficiency as the...

AI systems don't truly learn autonomously, per this cognitive science analysis that's buzzing on Hacker News with 62 points. Essential reading for agentic AI progress.

HSImul3R closes the gap with a novel simulation-ready Human–Scene Interaction 3D reconstruction framework, using physics-in-the-loop to formulate reconstruction as a bi-directional process. Key for agentic environments bridging perception to sim.

• AttnRes by Kimi Team swaps fixed residuals for depth-wise ones, revolutionizing Transformer architecture.
• MoDA lets each attention head access...

New survey explores deep generative modeling for tabular data:

• Reviewed through five key requirements, starting with utility of synthetic data
  -...

Grokking in neural networks is framed as a variance-limited phase transition driven by spectral gating, via tail-index analysis of stochastic gradient noise. Presented at ICML 2019—key mechanistic insight into training dynamics.

Breakthrough in multimodal captioning: New paper introduces a novel deep learning model for image captioning using an advanced vision transformer architecture with a powerful LLM.

Key inference efficiency breakthroughs:

• TiDAR hybrids diffusion drafting with AR verification for 6x speedup, bypassing memory bottlenecks in...

New research tackles the consensus problem in multi-agent systems (MAS) subject to external disturbances. Key focus: strategies based on dynamical approaches for robust coordination.