Do Existing Preconditioners Improve Biomedical Tabular Foundation Learning? An Empirical Study on TabPFN Optimization
A comprehensive study tests five AdamW-based preconditioning strategies when fine-tuning TabPFN v2.5 across 59 biomedical datasets, measuring both performance and statistical significance. The results are blunt – the original AdamW optimizer wins consistently, while curvature-aware preconditioners fail to deliver reliable gains across diverse healthcare tabular tasks.
Meta’s Muse reportedly leaked its entire filesystem with “almost no prompt injection resistance”
Two developers say they coaxed Meta’s Muse into exporting a zipped copy of its root filesystem, including Ubuntu system files, app templates, and internal documentation. Meta pushes back, saying Muse runs in persistent per-user Linux VMs, but the incident still spotlights how quickly agent “helpfulness” can collide with security boundaries.
Accelerating vision-language models with LFM2.5-VL-DSpark
This release focuses on making multi-modal vision-language modeling faster by optimizing the training and serving path for embedding-heavy workloads. The pitch is improved throughput without throwing away the retrieval and grounding quality that embedding-centric VLMs depend on for downstream applications.
4DGS-JEPA: Temporally Compositional Joint-Embedding Prediction for Dynamic Gaussian Splatting
Researchers propose 4DGS-JEPA, a Gaussian-native approach for learning predictive dynamics in evolving 3D scenes using joint embeddings rather than just reconstruction. The key idea is temporal composition – predictions should stay consistent even when different future-trajectory paths reach the same endpoint.
Gemini 3.8 Live with Live Avatar gives Google’s AI a face
Google is rolling out Gemini 3.8 Live with a “Live Avatar” that lip-syncs and changes facial expressions during conversations. It’s currently limited to Gemini Enterprise customers, but Google claims the avatar can switch between 97 languages while keeping video fidelity stable.
The AI Neuroscientist: An Interactive Agentic Interface for Neuroimaging Analysis
The AI Neuroscientist combines an LLM with neuroimaging tools so researchers can run quality control, modeling, and visualization through natural-language queries. Instead of opaque scripts, the agent is designed for interactive exploration, and the paper evaluates it on fNIRS data against general-purpose LLM agent baselines.
Google is sending an AI satellite into space next week
Google’s Project Suncatcher will launch a satellite carrying AI processors (including Tensor Processing Units) to measure how they handle space radiation, thermal extremes, and physical stress. The goal is to stress-test whether AI data-center hardware can survive and perform under orbital conditions, potentially informing future “AI-in-orbit” infrastructure.
Do Synthetic Personas Predict Real Audience Response? A Sim-to-Real Study Where a No-Persona Baseline Beats Persona-Based Copy Simulation
A sim-to-real study uses thousands of Upworthy real headline A/B tests to test whether persona-conditioned LLMs can predict which copy will perform. The surprising outcome – a no-persona baseline that estimates click likelihood for a typical reader ranks variants better, while persona simulation degrades predictive validity across multiple model tiers and replication splits.