Search

Artificial intelligence is unpredictable for a mathematical reason.

Translated title. Original: L'Intelligenza artificiale è imprevedibile per una ragione matematica

A study argues that AI’s unpredictable outputs can stem from the mathematical structure of its algorithms, not just external randomness. Understanding how tiny changes in initial conditions lead to different results may help improve reliability and reproducibility.

Google DeepMind study finds LLMs can't reliably roll their own dice

A DeepMind/NUS study finds LLMs struggle to sample truly random outputs from specified distributions, showing bias and distribution mismatches; providing seeds from external randomness yields accurate sampling, highlighting a risk for agentic AI systems and urging external randomness generation for secure, faithful sampling.

[2610.23570] A general recipe for data-adaptive anytime-valid envelopes for iterative algorithms

The paper proposes a general method to build data-adaptive, anytime-valid envelopes for iterative algorithms, grounded in time-uniform noise concentration and a sub-psi condition, with constructions via mixtures, stitching, and martingales. It applies to SGD and other methods, offering ready-to-use envelopes with implementation details and numerical validation.

[2610.31041] Algorithmic drift in generative adaptive learning

The paper identifies algorithmic drift as a time-based fairness risk in adaptive, AI-assisted tutoring. It formalizes drift, validates it via simulations using online trackers or LSTM knowledge tracing, and proposes a Temporal Fairness Stabilizer to periodically re-assess learner ability, showing it can reduce the achievement gap compared with standard exploration, while delineating analytical, simulation, and real-world validation boundaries.

Beyond Likelihood Ratio Bias: Nested Multi-time Scale Stochastic Approximation for Likelihood-Free Parameter Estimation | Machine Learning | Springer Nature Link

The document proves the uniform convergence and distributional limits of an approximate posterior under a nested multi-time scale stochastic approximation, using Delta Method and Lipschitz assumptions on the variational family; it establishes both almost-sure uniform convergence to the projected posterior and probabilistic convergence rates for the finite-sample posterior error.

[2609.33543] OSCC: Certified Observation-Safe Coupling Optimization for Gradient-Noise Control in Imperfect-Information Learning

OSCC introduces observation-safe counterfactual coupling to control gradient-noise in imperfect-information learning, with a calibration-based OSCC-Select to choose rollout strategies; experiments show significant reduction in return-contrast variance and improved calibration.

[2610.33744] Estimating Stability of Individual-Specific Predictions from AI-based prediction models using Bootstrap and Cross-validation Resampling

The study evaluates how individual predictions from AI-based clinical models vary across resampled development datasets. It compares bootstrap and cross-validation methods on Elastic Net and Random Forest, showing instability rises in small samples and differs by model; no single resampling procedure is best, though out-of-bag bootstrap and 2-fold CV align closely with independent sampling, while repeated CV often underestimates instability.

OpenMOSS-Team/MOSS-Image-SFT · SFT vertical colour stripes: where they come from, plus a small fix

Users report persistent vertical color stripes in MOSS-Image SFT, linked to the width-wise positional encoding and specific SFT directions. Removing a subset of directions or using a targeted LoRA mitigates the effect; questions remain about training changes and reproducible fixes.

Quantum Fluctuations Stabilise An Unexpected 'electronic Quasicrystal' In Bilayer Wells

MIT reports a new quantum phase in bilayer quantum wells: an electronic quasicrystal formed by a 30° twisted Wigner crystal, stabilized by zero-point motion and quantum fluctuations. The finding, supported by neural-network variational Monte Carlo and an energy-landscape analysis, shows the quasicrystal as the ground state over the honeycomb arrangement in a wide parameter range and is complemented by an optical plasmon mode that reinforces stability.

Therefore AI almost always chooses the number 7: Gradient descent - Dagens PS

Translated title. Original: Därför väljer AI nästan alltid siffran 7: Gradient descent - Dagens PS

AI language models, when asked to pick a number between 1 and 10, disproportionately choose 7, mirroring human biases in randomness. Newer studies across languages confirm 7 as the most frequent response, highlighting that models optimize for likely outputs rather than true randomness, a reflection of their training on human language and patterns.