ieee p3109 arithmetic formats for machine learning
the ieee p3109 draft standard defines parameterized binary floating-point formats and operations to support efficient, consistent machine learning computations.
topic
research
peel framework exposes ai distortions in research
a semiotic scaffolding called peel reveals systematic distortions in ai-generated text condensations, showing fluency does not equal fidelity.
eeg region contributions to cognitive workload prediction
a large-scale analysis across four eeg datasets examines how different scalp regions contribute to predicting cognitive workload, revealing consistent patterns and practical implications for sensor selection.
five papers that explain llms clearly
a guide to five foundational papers covering transformer architecture, few-shot learning, scaling laws, instruction tuning, and retrieval-augmented generation for understanding large language models.
behaviorbench tests ai on real user decisions
a new benchmark uses public prediction-market and blockchain data to evaluate how well models predict individual beliefs and trades.
google open-sources hydrology model for flood forecasting
google research releases its ai flood forecasting framework on github, enabling national agencies to train and customize models with local data.
dpo cuts text degeneration in ocr models
direct preference optimization reduced text degeneration by an average of 59.4% across five ocr model families by using the model's own failure outputs as rejection pairs.
human-gated bandits speed up rental pricing
a new framework lets human agents approve or reject algorithmic price suggestions, using old pricing data to skip the slow start typical in sparse booking markets.
tera cuts derivative gp cost to linear scaling
a new method reduces the cubic complexity of gaussian processes with gradients by using exact gradient reduction and vecchia approximation.
structure over models for scientific discovery
a position paper argues that in high-dimensional settings, many different mechanisms can produce the same data, so predictive success does not prove a model has found the true mechanism, and large language models can hide this by giving a single fluent explanation.
action-gated memory keeps robot policies at constant vram
aura-mem uses a learned gate to write only when observations change actions, keeping memory fixed at 4,224 bytes regardless of episode length.
target updates may stabilize linear q-learning
periodic and soft target updates can guarantee convergence in linear q-learning under explicit spectral and step-size conditions.
gradient-based early stopping for boosted trees
a new method uses gradient tests instead of validation loss to decide when to stop training gradient boosted trees, avoiding the need for a patience parameter.
class-split anomaly detection can fail when classes overlap
a new diagnostic reveals that common anomaly detection benchmarks become unreliable when held-out classes overlap with normal data in representation space.
aligning ehr models with llms for clinical reasoning
a new framework aligns structured electronic health record representations with large language models to improve clinical prediction and reasoning.
microsoft releases mai models with low active parameters
microsoft announced two new language models, mai-thinking-1 and mai-code-1-flash, with low active parameter counts and claims of clean training data.
a gentle primer on llm explainability
an overview of advances in making large language models more interpretable through dynamic evaluation, statistical methods, and accessible tools.
microsoft assert tests ai behavior from text descriptions
microsoft's open source assert framework uses ai to turn plain-language rules into scored tests for application-specific ai behavior.