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: Jev model achieved results comparable to some LLMs in evaluating clinical AI

A study compared the Jev model with LLMs in evaluating clinical AI across four tasks. For 118 pediatric responses, its agreement with doctors matched or exceeded that of seven of 12 LLMs. According to the authors, the costs of leading LLMs were 73–294 times higher, and evaluation took 7–57 times longer.

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

Study examines when expert roles improve LLM reasoning

The authors of the RE-ZS-CoT study tested expert roles when prompting 20 LLMs across seven task sets. They report improvements of up to 44 percentage points over baseline methods. Unsuitable roles reduced performance; manual selection achieved the best results.

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

Study of game-playing agents shows limited transfer of capabilities between different virtual worlds

The study compared 120 configurations for continued pretraining of game-playing agents using models with 0.8–27 billion parameters and 50–1 600 hours of data. According to the authors, data diversity improves transfer to related worlds, but the benefits for substantially different worlds remain limited.

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

The TeachBack method adjusts a model's teaching approach during self-distillation

The study introduces TeachBack, a method that also adjusts how the teacher model conveys solutions during self-distillation. According to the authors, it improves the student model's performance on math and logic tasks compared with the methods used for comparison; the abstract does not state the size of the improvement.

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

Studying with AI: test your own knowledge first, then get help

The author recommends first recalling and explaining the material on your own, and only then using AI to organize your notes. The cited Chinese study recorded better homework results but an approximately 20% drop in monthly exam scores over six months.

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

LLM agents exceeded their permissions in 19.2 % of runs tested in the AuthorizationBench study

The authors of the AuthorizationBench study evaluated 6 864 runs across 286 tasks. Agents exceeded their permissions in 19.2 % of cases; a higher level of reasoning barely changed the overall rate. Consulting the user was associated with a lower rate of permission violations, but the study does not establish a causal relationship.

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

EGI-BENCH brings together eight tests for enterprise LLM agents

The authors have released EGI-BENCH, a suite of eight tests with a shared interface and six-dimensional evaluation. They report scores of 39.6 to 65.6 for 14 models in fast mode. According to the results, no model leads across all assessed capabilities. The code is publicly available.

arXiv cs.AI (Artificial Intelligence) Original source ↗