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

Neural operator approximations for two-dimensional neutron flux

The study extends one-dimensional neural operators to two dimensions. The authors test Fourier neural operators (FNO) and U-shaped neural operators (UNO) to approximate scalar flux in one-group neutron transport with isotropic scattering. They investigate three approaches, including a variation using…

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

Unified geometric operations on spatio-temporal adaptation vectors for federated incremental learning

An academic paper (arXiv:2607.19384) proposes the SUM method for combining federated and continual learning. It addresses catastrophic forgetting caused by client heterogeneity and sequential tasks through a purely server-side geometric operation on adaptation vectors, without imposing any additional load on clients. It achieves up to…

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

Orthogonalized read as a training scaffold for mLSTM recurrent memory

Research on arXiv describes how orthogonalizing the mLSTM memory matrix at read time (five iterations of the Newton-Schulz algorithm) improves associative recall but does not act as a memory improvement. The effect comes from reconditioning the learning problem during training; removable once convergence is reached, effects on…

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

Autoregressive models of animal behavior from an agent's perspective

Researchers introduced a framework for training generative models that predict animal behavior from the animal's own perspective. The models capture an individual's observations and movements, supporting social behavior in groups. They demonstrated it on fruit flies (Drosophila) and released a library for working with…

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

Deep reinforcement learning for the asymmetric game Baghchal

The research paper examines four deep reinforcement learning algorithms (DQN, REINFORCE, PPO, MuZero) trained on the traditional Nepali game Baghchal with asymmetric roles. MuZero achieved the best results: an 86% win rate when playing as the tigers and 62% for the goats. PPO was competitive with lower computational costs.

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

Cross-subject semantic decoding with shared-space alignment for generalized neural representation learning

A research study addresses the generalization of neural decoding across individuals. The method aligns neural responses from multiple subjects into a shared latent space and trains a decoder to predict semantic embeddings. On electrocorticographic data from natural language comprehension, it achieves consistent…

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

Bayesian model selection in transformers using involutions

The study examines how transformers (2.8M–316M parameters) perform Bayesian model selection. In a controlled setting with fixed-point-free involutions, they achieve an entropy match of 0.01 bits with the Bayesian optimum. Model selection works with both integer tokens and opaque symbols, but fails with…

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

The CruiseBench benchmark for predicting the remaining life of aircraft engines

A paper introduces CruiseBench, a new benchmark derived from the N-CMAPSS dataset for evaluating models that predict the remaining useful life (RUL) of aircraft engines. It introduces CPM-N-CMAPSS (a mask for identifying flight phases) and tests LSTM, GRU, TCN and TSMixer models. TSMixer achieved the lowest RMSE of 3.4±1.71.

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

State compression in two-stage LLM relays: a study of constraint preservation

The research compares four approaches to intermediate state compression in a two-stage LLM system: no compression, narrative summary, JSON schema and embedding-based pruning. JSON schema achieved the highest accuracy (0.96), narrative the lowest (0.48). Structured representations preserve constraints better than compact…

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

An audit of ML models for trading on Binance: All strategies lost money

An arXiv study audits ML models for predicting cryptocurrency extrema for trading on Binance Spot. The main result: the strongest model lost 6.72 % over 19 cycles, while validation models lost 1.79 % and 2.80 % compared with buy-and-hold. One model lost 44.30 % versus −41.20 % for the passive strategy. Conclusion: no…

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

STN-TGAT: Stock portfolio construction through prior-guided graph attention and soft-threshold sparsification

A research paper on arXiv describes STN-TGAT, a model combining Transformer and Graph Attention Network for predicting and ranking stocks in the S&P 500. The approach integrates stock price dynamics with relationships between stocks and incorporates realistic trading conditions such as transaction costs. According to the authors, it achieves…

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

MILP-Evo: Automatic design of MILP solvers using LLMs and program evolution

A research paper on arXiv introduces MILP-Evo, a framework for automatically designing MILP solver logic. The method uses LLMs to iteratively generate candidate programs, tests them directly on MILP instances and selects components such as a cut selector and branching rule. The outputs are explicit, inspectable…

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

When does consensus outperform voting? A critical analysis of label fusion methods in medical image segmentation

A scientific paper compares consensus segmentation methods in medical imaging. It shows that the sophisticated STAPLE method reduces in practice to majority voting with high suboptimality and fails under data imbalance. Simple voting is surprisingly robust; the best results come from deep…

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

OntoBook: Synthetic medical textbooks from ontologies for training medical encoders

A research method converts medical ontologies into a training signal for encoder language models. The team created a dataset of 1.3 million LLM-reformulated textbooks from French ontologies and trained a 149M-parameter ModernCamemBERT with two objectives. On the Distemist-FR benchmark, they achieved an improvement of 8.0…

arXiv cs.AI (Artificial Intelligence) Original source ↗