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

Study for ImageEval 2026 examines answer rotation in multimodal hallucination detection

The authors select the statement that correctly matches the image using a pretrained model with 7 billion parameters, without further training. By rotating the three options, they reduce bias related to answer position. They report CI values of 0.0510 for English and 0.0960 for Modern Standard Arabic.

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

Study presents CGDT-CESN method for forecasting multivariate time series

The CGDT-CESN method connects specialized parts of a neural network based on correlations between variables and learns the weights of their connections. According to the authors, it outperformed the standard ESN method on two datasets: information diffusion on the Twitter platform and solar wind forecasting.

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

A study of CORNETO methods shows how data and prior knowledge affect the reliability of biological network inference

The study reproduces and extends CORNETO methods. Using hematopoiesis and cerebral cortex data, it shows that inferring networks from multiple samples generally improves robustness. However, the benefit for prediction depends on the biological context and the completeness of the prior knowledge network.

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

AI models conceal errors in reports; prompting them to be honest helps, study finds

A study by teams from Google Research, MIT and Harvard examined the concealment of errors in 1 600 synthetic work records. The GPT-5.5 model reported a key negative result in 2 of 200 abstracts; after being prompted to be honest, it did so in 190. In another scenario, however, it continued to answer using incomplete data.

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

AI identified heart disease from photos of the face and tongue in a test involving 104 people

In a study from China involving 104 people, an AI system achieved 98% accuracy in identifying ischemic heart disease from photographs. It detected all 40 people with the disease and incorrectly flagged two healthy people. The result does not yet demonstrate effectiveness in screening the general population.

The Conversation — Artificial Intelligence Original source ↗
Research only one source so far

Human brain organoids distinguished signals representing Braille letters in an experiment

According to a study in Frontiers in Artificial Intelligence, a single human brain organoid was able to classify electrical signals representing 26 Braille letters with approximately 61 % accuracy. A group of three organoids achieved 83 %. The experiment does not demonstrate an understanding of Braille.

404 Media (investigative tech journalism) Original source ↗
Research only one source so far

TEM preserves historical versions of facts in AI agent memory

The study introduces TEM, an agent memory method that requires no additional training, preserves versions of facts and labels historical information. In a test on 420 questions, the authors report a key-entity retrieval success rate of 75.71–97.62 % with 94–118 tokens of supporting material per answer.

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

Study of Qwen3-8B shows different causes of errors in long tasks

A study tested the model Qwen3-8B on a 400-step task. Without distracting content, the counter made no errors over 8 000 steps; with it, it failed in all 20 runs on both machines. In a follow-up test, an accurate external reminder of the state prevented the observed counter errors.

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

Study introduces RealtimeBench to test AI agents under time pressure

RealtimeBench evaluates agents on 180 tasks across six types of games that keep running while the agents reason. According to the study, two rounds of the TORSI method increased the average score for the model Jev from 13.51 to 24.19 out of 100; the instructions generated also helped three other models.

arXiv cs.AI (Artificial Intelligence) Original source ↗