AI Research Roundup

Efficient Agent Frameworks

• 🔥 DR-Venus: Towards frontier edge-scale deep research agents with only 10K open data.
• Self-Evolving Framework: For...

Agentic AI frontiers push forward with DR-Venus enabling frontier edge-scale deep research agents using only 10K open data, alongside a self-evolving framework for efficient terminal agents via observational context compression.

Practical robotics advances rapidly:

• Cortex 2.0 grounds world models in real-world industrial deployment.
• Sony's Ace beats elite humans at table...

• Diamond Sponsor with 95+ papers from Google Research/DeepMind
• 7 orals on privacy, 3D generation, Mamba learning, visual planning, Text-to-SQL, RL...

SAVOIR proposes Shapley-based reward attribution to learn social savoir-faire in AI agents, enhancing RL with fair modeling for socially intelligent behaviors.

ReImagine rethinks controllable high-quality human video generation pipelines via an image-first synthesis approach. Join the discussion on the paper page for deeper insights.

Efficient Inference Advances

• 🔥 Micro Language Models Enable Instant Responses: Ultra-compact models (8M–30M parameters) instantly generate the...

ABMMDLF framework applies attention-based multimodal deep learning to achieve high-accuracy spatio-temporal crop yield prediction, enhancing precision agriculture.

• Top claim: Kimi K2.6 is the most powerful open-source LLM yet.
• Prior models excel at chatting (good), coding (decent), reasoning (okay).
• But they fail badly at long, complex, real-world tasks – K2.6 aims to fix this.

Key papers spotlighting test-time adaptation trends:

• TEMPO scales test-time training for large reasoning models.
• Evaluation-driven scaling for scientific discovery.
  Watch eval methods enhance large models on complex tasks.

• Chat2Workflow introduces a benchmark for generating executable visual workflows from natural language
• ClawNet proposes a human-symbiotic agent...

Emerging techniques are accelerating autoregressive inference for real-world use:

• Speculative decoding targets autoregressive video generation
• KV...

Quick scan of today's top AI papers:

• Speculative Decoding for Autoregressive Video Generation from UC Berkeley
• Submitted by SteveZeyuZhang
• UniMesh featured

New arXiv paper 2604.19738 delves into phase transitions in the fluctuations of functionals of random neural networks, a key theoretical advance for understanding random NN dynamics.

New arXiv paper proposes a deep-learning framework for environmental sound deepfake detection (ESDD)—identifying real vs. fake sound scenes, bolstering defenses against audio deepfakes.

New tools highlight trends in probing model weaknesses for better safety:

• Contrastive Attribution analyzes LLM failures on realistic benchmarks.
  -...

Deep learning can now train neural networks built from discrete logic gate functions, bridging ML software with silicon hardware for efficient, native AI designs.

MM-ADF framework uses multimodal deep learning for anomaly detection in urban infrastructure networks, enabling smarter city monitoring.

New Benchmarks

• 🔥 Precise Debugging Benchmark: Frontier LLMs like GPT-5.1-Codex pass unit tests over 76% of the time but touch fewer than 45% of...

Stratagem introduces transferable reasoning learned through trajectory-modulated game self-play, highlighting how self-play trajectories can enhance general reasoning skills.