AI & ML Daily Digest

Efficient Agentic Models

• 🔥 AgenticQwen-30B-A3B: NEW paper from Alibaba presents a 30B MoE with only 3B active params that matches Qwen3-235B on...

Sessa proposes selective state space attention in new ML research. Join the discussion on this paper page.

Two fresh takes on frozen LLM limits:

• Google's Hope: Nested architecture with NLMs, Self-Modifying Titans, and CMS enables continual learning,...

• FIT uses Fourier transform to decouple image style and semantics in frequency domain, addressing gaps in prior input transformations.
• Pipeline:...

Emerging trend in agentic world modeling for robotics:

• New taxonomy synthesizes 400 works into L1 Predictor, L2 Simulator, L3 Evolver levels
  -...

Hacker News thread exposes peer review drama around TurboQuant paper:

• TurboQuant = restricted EDEN (NeurIPS '21, ICML '22) on KV-cache, not novel...

AgentSearchBench launches as a benchmark for AI agent search in the wild, highlighting real-world challenges. Join the discussion on the paper page.

Emerging RL trend boosts small LLMs to rival massive models:

• Poly-EPO uses set RL to fight diversity collapse, generating diverse reasoning paths...

Emerging lightweight approaches for LLM safety:

• SIREN probes internal layers with linear probing and L1 regularization for efficient harm detection,...

Breakthrough multimodal ML paper fuses WSIs and CT images via multi-stage attention (MSAT) and co-attention for precise gastric cancer recurrence...

PaperCoder auto-generates functional code repos from ML papers using a multi-agent lifecycle: planning, analysis, generation.

• Specialized agents...

ResVLA innovates Vision-Language-Action models by anchoring generation on low-frequency intent signals:

• Semantic-to-action bridge: Links high-level...

MARCO revolutionizes semantic correspondence as a minimalist DINOv2-based model, slashing params while boosting object region matching.

Key...

New vulnerability in LLM serving: GPU DRAM bit-flips target shared KV-cache blocks, causing silent divergence where outputs look legit but err.

-...

• Modern NNs deliver high accuracy but are overconfident due to poor calibration.
• Strong link: Curvature and margins during optimization drive...

DiffNR leverages diffusion enhancement to optimize neural representations for sparse-view 3D tomographic reconstruction, tackling few-shot challenges in the field.

ICLR 2026 Workshops

• GRaM Workshop: GRaM Workshop at ICLR 2026 lists speakers including Kathlén Kohn from KTH and Gabriel Loaiza-Ganem from Layer...

ICLR 2026 workshops spotlight foundational DL trends toward interpretable, physics-aware models:

• AI&PDE: Tackles PDE solvers via neural operators...

Key developments in agent infrastructure:

• AMA-Bench targets long-horizon memory critical for LLMs as autonomous agents in complex apps
  -...

• Omni is a unified foundation model trained natively on text, images, videos, and 3D geometry.
• Introduces Context Unrolling: Explicit reasoning...