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

The ForTy map maps forest types with a resolution of 10 meters and an accuracy of 90.2 percent

Google DeepMind and WRI have created ForTy – a global forest map distinguishing 6 types of cover (primary forests, plantations, tree crops, etc.) with 10-meter resolution. It achieves 90.2% accuracy, built on the AlphaEarth Foundations model. The goal is compliance with the EUDR regulation (deforestation-free products).

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

MIT tracks how AI chatbots influence US elections

MIT launched the LLM Election Observatory dashboard to monitor how AI chatbots respond to questions about the 2026 US elections. The project tracks how chatbot responses differ by location, political affiliation, and other factors, in order to understand their influence on election discourse.

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

Microsoft is developing a methodology for testing Microsoft Copilot on synthetic enterprise data

Microsoft has developed a methodology called the Industry Evaluation Data Generation Platform for testing Microsoft Copilot. Instead of real sensitive company data, the system generates synthetic enterprise environments with realistic organizational structures, communication history, and work artifacts. The aim is…

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

Post-training of large language models using LLMON

A research team introduced LLMON, a post-training methodology that separates instructions from data in LLMs. The framework was tested on the Granite-4.1 models (3B and 8B) using a structured corpus of 4.6B tokens. The result is a reduction of prompt injection attacks from 53–100% to 0.02–5.6%.

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

Soft-MTP: Improving Multi-Token Prediction Training Without Additional Overhead

Researchers identified gradient interference in standard MTP training. They proposed Soft-MTP, a one-line change with no additional parameters, which removes the noise. In tests from 167M to 1.2B parameters: lower perplexity, faster speculative decoding, and better zero-shot capabilities. The code is publicly available.

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

SciGame benchmark for testing AI agents on scientific discovery

A research team introduced the SciGame benchmark for testing LLM agents on scientific discovery. The benchmark contains pairs of Original and Counterfactual test environments with canonical laws and laws from papers published after the models' knowledge cutoff. Tests of five frontier models found: traditional knowledge…

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

Rank-aware exponential mechanism for differential privacy of listwise preferences in LLM alignment

The research paper proposes a mechanism for protecting privacy in listwise preferences for LLM alignment. Instead of uniform noise, it concentrates the perturbation on boundary items. Experiments show higher performance than existing local differential privacy methods. The code is publicly available.

arXiv cs.CL (Computation and Language / NLP) Original source ↗
Research only one source so far

Mechanistic interpretability of LLMs in solving ontology engineering tasks

Research using the MIOE benchmark tests how LLMs internally process ontological tasks (entity typing, subsumption prediction). SAE-, DBM-, and DAS-based methods are compared on the Qwen3-8B and Qwen3-4B models. Distributed subspace methods give the best results, but the causal mechanism remains…

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

MGO-SC Optimizer: Sign-Consistency Gate for Stable Training Under Noise

A research preprint on arXiv presents MGO-SC, a modification of Adam-style optimizers with a consistency gate based on sign exponential moving average. On CIFAR-10 with ResNet-18, it achieved a +1.4 point improvement at 40% symmetric label noise without additional second-order information. The authors are releasing the code and raw…

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

Study introduces EVENET model for analyzing particle collisions

The authors introduced EVENET, a model pretrained on 500 million simulated collisions. According to the study, it outperforms comparison models, particularly when data is limited; they validated its transfer to experimental data by rediscovering the ϒ meson in CMS Open Data.

Nature Machine Intelligence Original source ↗
Research only one source so far

Study assesses time savings with AI, including the cost of overlooked errors

The study author proposes assessing time savings alongside overlooked errors. A reanalysis of 246 tasks from the METR study puts the slowdown with AI at 19 % or 24 %, depending on the metric measured. A small experiment showed that overestimating AI reliability led to less checking and a higher error rate.

arXiv cs.AI (Artificial Intelligence) Original source ↗