The Archive

Claude users in India are starting to see Indian rupee-denominated subscription plans.

Study compares production vs. perception asymmetry in Llama-3.1-8B via token probability analysis, finding LLMs lack functional production-perception distinction.

Q2SAR applies quantum multiple kernel learning to QSAR modeling for improved prediction of compound toxicity and bioavailability in drug discovery.

Automated red-teaming system discovers reusable vulnerability patterns in production LLM agents (Claude Code, Codex) operating on untrusted content.

Hourglass reasoning enforces structural isolation between induction stages in LLMs, using compressed symbolic state to improve few-shot reasoning.

Roadmap for World Action Models (WAMs) connecting interventions to consequences; standardizes action spaces, datasets, and interfaces for embodied physical intelligence.

Camera-only ground vehicle recovery using visual odometry for line-following robots in warehouses.

Contextual bandit algorithm combining multinomial logit choice models with submodular diversity objectives.

RAGU: open-source modular GraphRAG system with two-stage extraction and compact domain-adapted LLM for knowledge graph construction.

Graph neural network with multi-scale features for accelerating fluid dynamics simulations in complex geometries.

Dataset and task for modeling semantic drift of compositionality in German and English noun compounds over decades.

Geometric Dimensionality Regularization (GeomDR) controls grokking timing by leveraging representation collapse as predictive signal.

"Context bombing" tricks hacking agents into shutting down before they can do harm.

NITROGEN: imputation-free transformer for Alzheimer's disease prediction from heterogeneous clinical data with uncertainty quantification.

Anytime-valid auditing framework for monitoring calibration of conditional quantile forecasters in data streams.

802.1X and RADIUS-based access control architecture for automated anomaly-driven revocation in IoT networks.

Across science, engineering, and finance, many of the most important risks come from low-likelihood, high-impact events. Estimating the probability of these... Across science, engineering, and finance, many of the most important risks come from low-likelihood, high-impact events. Estimating the probability of these events with brute-force Monte Carlo sampling—running a model repeatedly with randomly drawn inputs to estimate the probability of rare outcomes—can require an excessive volume of model iterations, especially when each sample comes from an… Source

Xiaomi-Robotics-U0: 38B multimodal autoregressive model for embodied synthesis leveraging foundation models with world physics.

Theoretical analysis of sampling-allocation limits in best-arm identification under fixed budgets.

LLMs parameterize cumulative prospect theory for modeling human route-choice biases at scale.

SKooP combines Koopman autoencoder with morphological symmetries to improve RL sample efficiency for legged robot locomotion.

LLaVA perturbations reproduce aphasic picture-naming error patterns, validating multimodal models as clinical simulation tools.

ARMT enables long-context LLM processing with constant memory scaling via associative recurrent architecture.

Empirical audit reveals trained distributional RL agents' risk estimates often violate first-order stochastic dominance.

Formal language theory on minimal coloring schemes for identifiability in Gold's language-learning model.

Differentially private collaborative Bayesian optimization achieves centralized performance without raw-data sharing.

Interaction-based test-time compute (model-environment feedback loops) outperforms longer reasoning or sampling at inference.

Bangla hate-speech detectors fail on implicit, context-dependent cases; benchmark-trained models miss cultural nuance.

Google DeepMind and AIM launch ATL Saathi, a Gemini-powered educational tool for Indian robotics labs.

Experts explain how they work, what they can do, and what's still unsettled.