Search

VLA — Robotics Research Hub

The document analyzes challenges in teaching humanoid robots to follow instructions using full-body actions, highlighting large, coupled action spaces, balance constraints, and scarce demonstrations, within reinforcement learning and learning-from-demonstration contexts.

[2611.12762] Sage-Taiyi: A Metacognitive Architecture for Scientific Discovery

Sage-Taiyi monitors evidence, assumptions, and unanswered questions to guide scientific research. It led the reported ideation benchmark and produced proposals tested in expert-guided collaborations, including chemistry experiments with higher reaction yields.

Scientisttwo, the artificial intelligence that tries to conduct scientific research without researchers – AI Magazine

Translated title. Original: Scientisttwo, l’intelligenza artificiale che prova a fare ricerca scientifica senza ricercatori – Rivista AI

ScientistTwo envisions autonomous AI conducting full scientific research cycles, from hypothesis to publication, using a multi-agent system that iterates through hypothesis generation, experiments, and simulated peer review; preliminary results show strong but not uniform performance compared with human benchmarks.

Machine learning models predict common cancer risks using routine health data

Translated title. Original: Les modèles d’apprentissage automatique prédisent les risques courants de cancer à l’aide de données de santé de routine

ML models using routine health data can predict risk for multiple common cancers, showing overall predictive power and variable accuracy by cancer type; the approach relies mainly on blood tests and medical history, with limited gain from genetics or metabolomics, and includes a tool (PRE-Screen) to standardize model use by AI agents.

I audited my own four-sector cosmology myself: what remains (Paper FP-0)

Translated title. Original: 自分で作った四セクター宇宙論を、自分で監査した：何が残るか（Paper FP-0）

The author audits their four-sector cosmology (FSC) against Planck and DESI data, showing FSC reduces to a form compatible with LCDM when constrained, but without distinctive predictions; statistical measures (AIC/BIC) favor LCDM, and no conclusive evidence for FSC remains. Reproducibility is claimed via provided scripts.

AI discovered the most dangerous form of cancer one and a half years in advance.

Translated title. Original: AI upptäckte den farligaste formen av cancer ett och ett halvt år i förväg.

An AI system named REDMOD can detect pancreatic cancer earlier than standard radiology by analyzing radiomic features in CT scans, achieving higher sensitivity and earlier lead times, but requires broader validation before clinical use.

[2611.00580] Evolving Scientific Intelligence: From Fragmented Evidence to Distributed Experimental Loops

The document proposes Evolving Scientific Intelligence (ESI), a blueprint for a networked system of intelligent laboratories that shares computable experimental memory, links results across domains, and uses feedback to iteratively generate testable hypotheses while guarding against biased optimization. It aims to create a cumulative, evidence-driven scientific process and advocates an Information Manhattan Project to build this infrastructure.

TI-0.1 Technical Report · Tutor Intelligence

The report presents TRACE, a system that uses LLM agents and robot telemetry to rebuild tasks in simulation, train policies with reinforcement learning, and improve them through hardware feedback. It reports 99.8% success in simulated Lego pick episodes and describes the Golem software stack.

[2610.32195] Principia: Test-Time Hypotheses to Principle Construction for Scientific Discovery

Principia is an LLM-based framework for test-time scientific discovery that builds and refines inductive principles from comparative hypotheses, iteratively revising them with new observations across molecular and biological tasks, and demonstrating improved performance over baselines.

[2611.08478] Scientist World: Grow a Research World for AI-Scientists

A proposal and result set for Scientist World, an executable environment built from research papers to evaluate AI-Scientists against human benchmarks; 1,000 environments released, showing AI outperforming humans in 74% of cases and enabling substantial model improvement in ideation.