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 introduces LatentGRM for faster evaluation of AI responses

The authors introduce LatentGRM, which evaluates pairs of responses without generating written assessments. According to the study, the LatentGRM-8B variant reduced evaluation time across four test domains by a factor of 6.1–7.0 at vote@5 while maintaining comparable accuracy in assessing preferences.

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

Study introduces Goldsmith for optimizing annotation rules using LLMs

The Goldsmith method revises annotation definitions based on errors on a small set of expert-labeled examples. It accepts changes only when the measured loss decreases. According to the authors, it outperforms the four prompt optimization methods used for comparison and improves subsequent annotation.

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

Study introduces GradSAT for solving SMT problems with gradient normalization

The study introduces GradSAT: it treats individual SMT clauses as Multi-Task Learning tasks and uses GradNorm to balance their gradients. Computation using PyTorch on a GPU is complemented by exact local search. According to the authors, the method stabilizes the search for solutions; the abstract does not report a measured speedup.

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

Study examines false news detection after changes to emotional tone

The authors propose Gated Cross Attention, which links news stories rewritten to change their emotional tone with LLM-generated explanations. They report improved detection on the PolitiFact and LUN datasets and competitive results on GossipCop. The available abstract does not provide numerical results.

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

Study proposes using small language models to simulate conversations with patients with addictions

The authors propose a method that first identifies cognitive components and then generates responses from simulated patients with substance use disorders. According to their evaluation, it improves the consistency of responses with the patient’s history and the counselor’s questions compared with the small models used for comparison.

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

Bias in healthcare AI tools begins as early as data selection and design

The authors of the article, together with collaborators from ten countries, warn that healthcare AI tools may disadvantage underrepresented groups. They use examples to illustrate the risks of unrepresentative data and recommend involving affected communities before development begins.

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

Analysis explains how recommendations and generative AI may keep people watching microdramas for longer

The author uses a research model to describe a loop between watching microdramas, data collection, recommendations and content creation. According to the available excerpt, generative AI speeds up production; a further role in making it harder to stop watching remains a possibility.

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

Cognitive science helped lay the foundations for today’s AI

Authors in the field of cognitive science explain how research into the human mind contributed to the development of neural networks. They highlight the Perceptron system from 1958 and later methods for training multilayer networks; in their view, current language models cannot be understood as mere databases.

The Conversation — Artificial Intelligence Original source ↗
Research confirmed by 2 independent sources

OpenAI has published 372 AI-generated mathematical results

OpenAI has published 372 mathematical results produced by an internal AI model on GitHub. According to the company, they solve open problems or represent significant progress. Many proofs have been formalized in Lean; publication does not replace peer review.

Wired — AI section · and 1 more source Original source ↗
Research only one source so far

WavePrune improves how models handle long contexts, according to a study

WavePrune restricts each RoPE channel to the first rotation period. According to the authors, it improved HELMET scores for four out of five models without further tuning; for Qwen3-8B, the score rose from 35.7 to 40.0. At a context length of 32K, it achieved speedups of 1.15× for prefill and 1.24× for decoding compared with FlashAttention-2.

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

CroissantMiner study: a single pass over a document outperformed four agent architectures in metadata extraction

The CroissantMiner study introduces a benchmark with 602 publications: 102 have human-verified annotations, and 500 have annotations generated using LLMs. When extracting metadata according to the Croissant schema, a single pass over the full context achieved higher accuracy than the four agent architectures tested.

arXiv cs.CL (Computation and Language / NLP) Original source ↗