Skip to content

Research separate from news

What could become important next

A preprint is a signal, not a finished product or an independently confirmed result. We therefore track research papers separately and show the actual state of supporting evidence for each one.

1381 published research events

Latest work

A significant claim from a single source is published only after further confirmation.

Research only one source so far

MIT projects selected for funding through the Genesis Mission initiative

The US Department of Energy selected 15 MIT projects for funding through Genesis Mission, an initiative aimed at creating a scientific discovery platform combining AI, supercomputing and quantum systems. In phase 1, the projects focus on integrating AI into scientific research, including quantum sensors…

MIT News – Artificial intelligence Original source ↗
Research only one source so far

A new method for attributing contributions to individual agents in multi-agent LLM systems

Researchers introduced Semantic Cooperative Games (SCG) and the SLIC algorithm for measuring the contribution of individual agents in multi-agent LLM systems. According to the authors, the method reduced computational costs on a medical benchmark by 93.3 % compared with the Monte Carlo Shapley baseline while maintaining high agreement between results.

arXiv cs.AI (Artificial Intelligence) Original source ↗
Research only one source so far

Do AI-native biotech companies need departments? Comparing models for AI-driven drug development

Researchers on arXiv compare the organizational structures of AI-native biotech companies. They examine whether these should copy human organizational charts or operate around a "Company World Model" — a dynamic model representing assets, goals and constraints. The benchmark tested 45 decision cases; an architecture focused on…

arXiv cs.AI (Artificial Intelligence) Original source ↗
Research only one source so far

HypEMBER: A hypernetwork-based ensemble for robust control of parameterized dynamical systems

HypEMBER is introduced as an RL framework combining hypernetworks and ensemble learning for robust control of parameterized dynamical systems. The method addresses generalization and robustness in the presence of measurement and model uncertainties. Tests on the Kuramoto-Sivashinsky equation and 2D gyre navigation demonstrated higher…

arXiv cs.LG (Machine Learning) Original source ↗
Research only one source so far

Detecting contamination in near-OOD benchmarks using a leak fingerprint

A method for detecting contamination in OOD benchmarks (a test class in the training set). It proposes a "leak fingerprint": high supervised decodability (AUROC ~1) + failure of automatic detection (<0.65). Across 52 settings: sensitivity 18/20, specificity 31/32; in 1 of 24 standard pairs, it found…

arXiv cs.LG (Machine Learning) Original source ↗
Research only one source so far

Editing time series forecasts using expert feedback

Researchers introduced DEFT, a method for improving forecasts from time series foundation models. The method balances model exploitation with structured exploration of components (trend and seasonality). It was tested on 78 datasets with three foundation models and seven query budgets.

arXiv cs.LG (Machine Learning) Original source ↗
Research only one source so far

SkillSight: Accurate skill retrieval in agent libraries

The research paper introduces SkillSight, a method for skill retrieval in LLM agent libraries. It identifies a problem where shared descriptive patterns in skill descriptions distort relevance scores. Without additional training, it improves Recall@10 by up to 20 % and is 1248× faster than the combined…

arXiv cs.AI (Artificial Intelligence) Original source ↗
Research only one source so far

LAARA: Adaptive rank allocation for parameter-efficient fine-tuning

A research paper introduces LAARA, a framework for dynamically allocating adapter ranks during fine-tuning. The method outperforms uniform LoRA allocation and competes with AdaLoRA and DyLoRA on the GLUE and MathInstruct benchmarks with fewer trained parameters. The code is publicly available.

arXiv cs.LG (Machine Learning) Original source ↗
Research only one source so far

Marine Engine Fault Dataset: an open dataset for diagnosing engine faults

A research team published an open dataset on arXiv from tests of a turbodiesel marine engine with five simulated faults (cavitation, filter blockage, fouling, valve blockage, turbine degradation) and multi-sensor measurements. The dataset serves as a benchmark for anomaly detection and predictive maintenance models…

arXiv cs.LG (Machine Learning) Original source ↗
Research only one source so far

Phionyx: A deterministic runtime architecture for LLMs with state management and safety control

The Phionyx architecture enables deterministic LLM execution with governance mechanisms. It has three layers: a deterministic core, a safety layer and a memory system. Experiments show a 31% reduction in computational overhead, a 24% improvement in data retention and zero deviations across 100 runs.

arXiv cs.AI (Artificial Intelligence) Original source ↗
Research only one source so far

Federated Learning for breast cancer prognosis with privacy protection

An arXiv preprint study of federated learning for predicting breast cancer progression. It combines clinical data and MRI images, testing a multimodal model with attention to transparency, safety and fairness. It compares federated and centralized approaches with the aim of keeping sensitive data at the patients' location.

arXiv cs.LG (Machine Learning) Original source ↗
Research only one source so far

Fast construction, slow evaluation: Pipeline choices dominate the variance in autointerpretability scores

A scientific study demonstrated that, when comparing sparse autoencoders, autointerpretability scores are influenced more by methodological choices in the evaluation pipeline than by actual architectural differences. Testing four metrics on two LM models shows that methodological variability exceeds architectural variability…

arXiv cs.LG (Machine Learning) Original source ↗
Research only one source so far

Generating bearing vibration signals with target fault probabilities using PR-GAN and counterfactual methods

A study compares two approaches to generating bearing vibration signals with predetermined fault probabilities (0.25; 0.50; 0.75). PR-GAN modifies signals through training-driven generation with a mean error of 0.046–0.059. A training-free Wachter-style counterfactual achieves higher accuracy (error…

arXiv cs.LG (Machine Learning) Original source ↗