AI Daily Brief

Two papers tackle perception robustness for autonomous systems:

• MOSAIC-GS reconstructs complex dynamic scenes from monocular video via geometric...

LatentOmni shows unified audio-visual latent reasoning outperforms explicit text CoT by preserving dense sensory signals and temporal consistency via...

• Mythos and GPT-5.5 show "game-changer" hacking abilities, finding thousands of vulnerabilities in major OS and browsers faster than expected.
  -...

FlowLong enables long, coherent video generation at inference time without retraining by blending predictions across overlapping windows via...

• Callosum shows mixing Qwen 3 VL8B with Kimi K2.5 beats GPT/Gemini by 18-25% on Video Web Arena at 3.7x lower cost
• Routing subtasks to smaller...

Two complementary strategies target LLM serving costs:

• KVServe dynamically compresses KV caches for disaggregated setups, cutting TTFT up to 32.8×...

• Spreadsheet-RL introduces an RL framework that trains LLM agents inside a realistic Microsoft Excel environment.
• An automated pipeline harvests...

Efficient Language Models

• 🔥 Mamba-3: Mamba-3 introduces improvements to sequence modeling using state space principles, including expressive...

Trend spotlight: LLM agents are advancing toward autonomous R&D.

• PostTrainBench evaluates agents like Claude Code automating fine-tuning on...

Key trend in production AI adaptation:

• Industry defaults to fine-tuning on small datasets as it seems cheaper, ignoring inference costs.
• Mixing...

Key trend in robotics: Teleop data is expensive and hard to scale, pushing alternatives like simulation, human videos, AC-WMs, and WAMs.

-...

Mamba-3 tackles Transformer inference woes with three innovations:

• Expressive recurrence from SSM discretization, complex-valued updates, and MIMO...

HorizonMath launches a benchmark for AI mathematical discovery:

• Over 100 unsolved problems in computational/applied math, immune to data...

Key highlights from the new paper:

• Introduces FinToolBench, a benchmark evaluating LLM agents on real-world financial tool use
• Targets...

New paper proposes Latent Entropy-Aware Decoding for MLRMs to mitigate hallucinations by thinking in uncertainty—a practical step toward reliable multimodal reasoning.

Rethinking UMM visual generation via masked modeling enables efficient image-only pre-training, shifting toward cost-effective multimodal systems.

New paper swaps static residuals for selective depth-wise attention across layers, tackling fundamental flaws in how traditional nets accumulate info...

Transformer Architecture Innovations

• 🔥 Moonshot AI's Attention Residuals: Moonshot AI released Attention Residuals repo as a drop-in...

AI systems don't truly learn, lacking the autonomous capabilities highlighted in cognitive science. This provocative take sparked 62 points of discussion on Hacker News, urging a rethink of data-driven paradigms versus human cognitive autonomy.

Emerging VLM trend targets compute asymmetry between vision and language for scalable efficiency:

• Penguin-VL deploys LLM-based vision encoders to...