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

NaviAIS: A dataset and framework for vessel trajectory prediction with structured navigation maps

A paper introduces NaviAIS, a standardized dataset with vessel AIS data for trajectory prediction in maritime environments, and the NaviLane framework. The dataset contains historical and future trajectories in common time windows, rasterized maps and vectorized vessel routes. NaviLane uses a hierarchical…

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

Relative positional encoding for improved routing in Transformers

A research paper applies relative positional encoding (RPE) in Transformers to the Team Orienteering problem. This allows the encoder to better model spatial relationships between graph nodes. Experiments on instances with up to 100 nodes show improvements in achieved scores and optimality compared with standard…

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

Predicting arsenic in groundwater using graph neural networks

A research study combines 74 000+ arsenic samples from the WQP, MRDS and gNATSGO databases and tests machine learning for predicting groundwater contamination. Graph neural networks (GNN) achieved results comparable to or better than gradient boosted trees, showing the benefit of spatially informed learning for…

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

How neural networks solve Sudoku: A study of the Lattice Deduction Transformer

A research paper analyzes the Lattice Deduction Transformer (LDT) on Sudoku. The model solves ~95 % of cells in the first pass rather than using iterative search. Backtracking and constraint propagation increased efficiency 1497× but did not change which Sudoku puzzles were solved. Digit-permutation augmentation achieved…

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

Clinical risk prediction using knowledge graphs and retrieval-augmented generation

A scientific paper introduces TRACER, a method for predicting clinical risks from electronic health records. It combines knowledge graphs enriched with disease severity information, retrieval-augmented generation and clinical note analysis. On the MIMIC-III and MIMIC-IV datasets, it achieved an increase of 28.5…

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

Do Sheaf Neural Networks use holonomy?

A research study examines whether Sheaf Neural Networks actually use the geometric property of holonomy. Neural Sheaf Propagation increases SO(2) rotation from 0.010 to 0.388 radians. Replacing learned mechanisms with identities increases error, confirming their use in the networks.

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

Group travel planned by AI agents through discussion

The research paper AI Tour Meeting introduces a framework in which multiple LLM agents with different personas collaborate on planning a group trip through natural language discussion. The aim is a simulation tool for analyzing the behavior of multiple agents in travel planning.

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

AI models are not optimized for pelicans on bicycles

Developer Dylan tested 7 popular AI models (GPT-5.6 Terra, Claude Sonnet 5, Gemini 3.5 Flash, Grok 4.5, Qwen3.7-Max, GLM-5.2, DeepSeek V4 Pro) on 48 combinations of animals and vehicles. The aim was to check whether the models are optimized for the specific case of a pelican on a bicycle. The analysis found no…

Simon Willison — AI tag (leading independent LLM commentator) Original source ↗
Research only one source so far

A quantum computer learns to correct its errors during computation

Google Quantum AI published a study in Nature on a reinforcement learning agent that optimizes thousands of a quantum computer's control parameters in real time and stabilizes it during computation. It enables simultaneous calibration and computation without interruptions, addressing a fundamental problem of analog…

Google Research Blog Original source ↗
Research only one source so far

An AI system helped Pakistani judges reduce case backlogs with a $38.50 return per dollar

A study by ETH Zurich and Imperial College London tested JudgeGPT, a GPT-4-based AI assistant, with Pakistani judges. Half of the 1 559 judges received access plus targeted training. These judges resolved 1 848 more cases annually (a 6.3 % increase), judgment quality improved, and the return on investment was $38.50 per…

The Decoder (daily AI news) Original source ↗
Research only one source so far

Accelerating text-to-video generation with calibrated sparse attention

Research introduces CalibAtt, a method that accelerates text-to-video generation by identifying and skipping unnecessary attention computations. It achieves speedups of up to 1.58× on Wan 2.1 and Mochi 1 without losing video quality. It also includes papers on video tokenization and generative models with…

Apple Machine Learning Research Original source ↗
Research only one source so far

LVSum: A benchmark for long video summarization with temporal alignment

Apple Machine Learning Research introduces LVSum, a benchmark containing 72 videos (averaging 16 minutes) from 13 domains, with human-written summaries and temporal references. Model evaluation revealed that transcripts contribute substantially more than visual frames; models fail in temporal grounding and modality coherence.

Apple Machine Learning Research Original source ↗
Research only one source so far

Length Value Model: Controlling generation length at token level

The research paper introduces Length Value Model (LenVM) for modeling remaining generation length at token level. On LIFEBench, it improves a 7B model's score from 30.9 to 64.8; on GSM8K, it achieves 63% accuracy with a budget of 200 tokens. It enables control of the trade-off between performance and inference efficiency…

Apple Machine Learning Research Original source ↗
Research only one source so far

RayRoPE: Positional encoding for multi-view attention

Apple introduced RayRoPE, a positional encoding technique in transformers for multi-view vision. It uses predicted points on rays rather than directions and achieves SE(3) invariance. Tests on CO3D demonstrated a 15% improvement in the LPIPS metric.

Apple Machine Learning Research Original source ↗
Research only one source so far

Reducing the cost of removing data from models: Apple research

Apple researchers published a method for more efficient unlearning by identifying low-influence points. The approach reduces computational costs by up to 50% and is part of privacy-preserving research aimed at safer implementation of data protection in AI models.

Apple Machine Learning Research Original source ↗