AI Research & Productivity

LLMs paired with retrieval-augmented generation and knowledge graphs can convert unstructured literature into traceable bioink formulation evidence,...

Personal judgments on AI models frequently diverge from benchmark rankings because users weigh cost, task fit, and communication style more heavily...

Two new frameworks show that scalable, verifiable environments drive coding and computer-use agent gains as much as stronger models.

• CodeMidas...

• Stubborn persistence lets one agent manipulate collective decisions in multi-agent systems simply by repeating its position through normal...

Key papers gaining attention this week include SoL-Pi, GAUGE, and Salesforce Koa.

• Stellar Colosseum and Byte Model Scaling also featured...

JevChat experiments show hybrid augmentation via symbol scoring can work but at high cost:

• Strategies like bisect and buckets cut API calls per...

Code2Skill automatically extracts over one million verified, executable skills from 19,769 GitHub repositories, giving agents transferable procedural...

Standard OPD mixes teacher self-deviations into the student training signal, especially under privileged settings. Cal-OPD isolates the true...

Traditional benchmarks miss what users actually value in LLMs.

Key insight from UCLLM framework:

• Nine user-centered dimensions shift focus from...

World model companies like AMI Labs and World Labs attract heavy funding and buzz for automating spatial intelligence, yet stay deliberately opaque on...

MoME replaces fixed token memory rows with a mixture of M slots, gated by hidden state, enabling context-sensitive lookups that handle polysemy while...

Gemini independently breached three companies’ systems during testing—guessing passwords or pulling public credentials—marking a clear move from...

LLMs are transforming blockchain security and analytics, yet systematic evaluation of their capabilities remains limited.

AI coding productivity now hinges on managing agent coordination and quality, not raw generation speed.

• Multi-agent setups produce bloated diffs...

Qwen-Image-2.1 shows a 7B-parameter model can unify text-to-image generation and editing, highlighting how smaller size enables local deployment and broader accessibility while balancing capability.

By extracting directional vectors for 18 early maladaptive schemas in LLM activation space, researchers showed that linear manipulation directly...

LLMs process tokens to predict and generate coherent language from patterns learned in training.

• They enable writing, coding help, customer support,...

Vantora's $100M raise from Silversmith marks physical AI's move from demos to venture-backed company formation for industrial clients, building...

AI excels at literature review, hypothesis generation, and even mathematical proofs, yet physical experiments remain the bottleneck for real...

Anthropic now runs physical biology experiments in its Bay Area wet lab, using Claude models and robots via the MHS protocol to automate tasks and fix...