A latent variable IRT framework decouples four safety-driving factors across 61 model configurations and 10 languages using 1.9 million evaluations, revealing that safety is largely unidimensional and that high cross-lingual gaps cluster in physical harm prompts and lower-resource languages.
Pyro: Deep Universal Probabilistic Programming
6 Pith papers cite this work. Polarity classification is still indexing.
6
Pith papers citing it
abstract
Pyro is a probabilistic programming language built on Python as a platform for developing advanced probabilistic models in AI research. To scale to large datasets and high-dimensional models, Pyro uses stochastic variational inference algorithms and probability distributions built on top of PyTorch, a modern GPU-accelerated deep learning framework. To accommodate complex or model-specific algorithmic behavior, Pyro leverages Poutine, a library of composable building blocks for modifying the behavior of probabilistic programs.
representative citing papers
Combined five-PTA dataset yields posterior on SGWB power-law amplitude and index consistent with nonzero signal but below 5-sigma significance, with reconstructed angular correlations matching the Hellings-Downs prediction.
Spectral-siren H0 constraints from GWTC-4.0 binary black holes remain robust when the mass spectrum is permitted to evolve with redshift at current detector sensitivity.
A new compact hierarchical triple main-sequence star system G1010 was discovered through combined low- and high-SNR spectroscopy, Gaia DR3 data, and TESS light curve analysis, showing an inner eclipsing binary rather than a compact object companion.
Symbolic emulators approximate key Lambda CDM functions to 0.001-0.05% accuracy across relevant redshifts and Omega_m values, enabling faster 3x2pt inference with consistent results.
Fine-tuning Qwen2.5-3B with GRPO on GSM8K to use Prolog yields competitive zero-shot MMLU performance but exposes an accuracy-auditability trade-off interpreted as reward hacking.
citing papers explorer
-
Why Do Safety Guardrails Degrade Across Languages?
A latent variable IRT framework decouples four safety-driving factors across 61 model configurations and 10 languages using 1.9 million evaluations, revealing that safety is largely unidimensional and that high cross-lingual gaps cluster in physical harm prompts and lower-resource languages.
-
Stochastic gravitational-wave background search using data from five pulsar timing arrays
Combined five-PTA dataset yields posterior on SGWB power-law amplitude and index consistent with nonzero signal but below 5-sigma significance, with reconstructed angular correlations matching the Hellings-Downs prediction.
-
Gravitational-wave constraints on $H_0$ are robust to (putative) redshift evolution in the binary black hole mass spectrum at current sensitivity
Spectral-siren H0 constraints from GWTC-4.0 binary black holes remain robust when the mass spectrum is permitted to evolve with redshift at current detector sensitivity.
-
Discovery of a compact hierarchical triple main-sequence star system while searching for binary stars with compact objects
A new compact hierarchical triple main-sequence star system G1010 was discovered through combined low- and high-SNR spectroscopy, Gaia DR3 data, and TESS light curve analysis, showing an inner eclipsing binary rather than a compact object companion.
-
Symbolic Emulators for Cosmology: Accelerating Cosmological Analyses Without Sacrificing Precision
Symbolic emulators approximate key Lambda CDM functions to 0.001-0.05% accuracy across relevant redshifts and Omega_m values, enabling faster 3x2pt inference with consistent results.
-
Training Language Models to Use Prolog as a Tool
Fine-tuning Qwen2.5-3B with GRPO on GSM8K to use Prolog yields competitive zero-shot MMLU performance but exposes an accuracy-auditability trade-off interpreted as reward hacking.