Technology & AI
LLM-as-a-Verifier hits 86.5% on Terminal-Bench V2
LLM-as-a-Verifier treats checking AI answers as a scaling axis, hitting 86.5% on Terminal-Bench V2 and 78.2% on SWE-Bench Verified without extra training.
Marcello Babbili
AI and technology, explained through what they actually mean. An AI-assisted newsroom under human editorial rules — every story cites its primary sources so you can check them yourself.
306 stories
Technology & AI
VAORA aligns vision-language model reasoning with physical actions
VAORA, a new reward design on arXiv, targets hallucinated reasoning and reasoning-action misalignment in vision-language models on physical tasks.
Technology & AI
NARAD voting protocol tallies 50,000 ballots in under one second
A new arXiv preprint from IIT Madras proposes a blockchain-verifiable voting system on Solana that keeps ballots private without a trusted key dealer
Technology & AI
Quantum stabilizer testing fails even with 99% memory
Limited quantum memory collapses the gap between stabilizer testing and learning, a new arXiv preprint shows, with both needing Theta(n) copies.
Technology & AI
Label-free AI sorts real cosmic signals from noise
A new arXiv preprint trains a Real-Bogus classifier without human labels, using simulated injections and dual-network co-teaching for noisy survey data.
Technology & AI
QCNN with path signature kernels targets time series
A new preprint pairs quantum convolutional networks with path signature kernels to tackle reparameterisation invariance in time series classification.
Technology & AI
Orcaella protocol: choose fast commit or 54% fault tolerance
New preprint Orcaella lets clients pick a fast 2-message-delay commit or a resilient path tolerating 54% equivocation at double the latency.
Technology & AI
CamVLA: robots work after cameras move, no recalibration
New arXiv preprint from NTU and Alibaba introduces a Vision-Language-Action model needing only a single RGB image to handle unseen camera angles.
Technology & AI
Agentic Data Environments: turning data into agent guardrails
Columbia researchers want data infrastructure to do more than store information — they want it to actively keep autonomous agents from causing harm
Technology & AI
LLM personality traits mapped and controlled in weight space
New arXiv paper shows low-rank adapters can dial up or suppress AI model personality traits like neuroticism and agreeableness, with measurable effects on safet
Technology & AI
Encrypted AI inference: layer-parallel cuts bootstraps 2.65×
New preprint cuts encrypted Transformer bootstraps 2.65× with 1.2% perplexity cost, bringing privacy-preserving AI inference closer to practical real-world depl
Technology & AI
xDECAF open-sources data flow security tool, 20+ models
xDECAF, a browser-based tool checking data flow diagrams against security constraints, shipped as a public preprint with a reusable dataset of over 20 models.