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 RubriX for faster response evaluation without generating text

The authors introduce RubriX, which evaluates responses using the similarity of vector representations. They report 75.4 % accuracy on rating scales unseen during training and a latency of approximately 28 ms with precomputed representations of scale levels; Jev still performs better on some of the human-annotated tests.

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

Review charts the development of LLMs and their design limitations

The study presents a classification of LLMs by architecture, training data and computational requirements. The authors link a balance between architectural complexity and data quality to more stable reasoning; the available abstract does not report specific measurements.

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

Resolution described how AI agents formalized the resolution of singularities

According to the published description, Resolution used an orchestrating agent and specialized agents to formalize the resolution of singularities in Lean. The work took approximately a week, produced about 600 thousand lines and incurred an estimated 135 thousand dollars in token costs.

Part of an overview of multiple AI topics; only this event has been covered.

Import AI (Jack Clark, Anthropic co-founder) Original source ↗
Research only one source so far

Study: Query-based modality selection reduces costs for Qwen2.5-Omni

The authors tested query-based input modality selection on WorldSense. According to the study, they matched or exceeded the accuracy achieved using all modalities with Qwen2.5-Omni, using approximately 41 % of the input tokens, with 39 % of the latency and 61 % of the monetary cost.

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

The language of the verification instruction changes the reliability of language model self-assessment

The study tested 10 open-weight models with 3–14 billion parameters in six languages. The language of the verification instruction alone changed calibration error by up to 0.04, without changing the accuracy of the answers. The authors also warn of bias in evaluation without verification.

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

Accordion model offers a choice of five compute budgets during inference

The study introduces Accordion, a model with five compute budgets. After training on 3.6 trillion tokens, greater computational depth improved the aggregate score across 37 tasks by 3.01 percentage points, according to the authors. The middle budget captured 89 % of the improvement with half the additional passes through the layers.

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

Study presents a system for assigning customers to sales representatives that considers both value and continuity of relationships

The study combines predictive customer scoring, estimates of the costs of disrupting business relationships, and multi-criteria optimization. The authors report better adherence to assignment rules and a more balanced distribution of opportunities when the system is deployed; the available abstract contains no numerical results.

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

Study introduces MedCaseAgent for drafting case reports on rare diseases

The study introduces MedCaseAgent, which combines clinical information with searches of the medical literature to draft case reports. The authors report better results than the comparison systems and say they have released the MedCase-150K dataset and the MedCase-Bench evaluation set.

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

Study introduces SAGE to evaluate how models match word meanings to images

The study combines three datasets into SAGE for selecting the correct image from ten candidates. The Gloss-Augmented Reasoning method combines definitions with checks of semantic constraints. According to the authors, it increases the average Hits@1 for Gemma3-12B from 65.90 to 71.22.

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

RxnExBench benchmark evaluates chemical reaction extraction using VLMs

The study introduces the RxnExBench benchmark with 3 565 expert-annotated reactions and 2 361 molecular structures. According to the authors, the evaluation of nine VLMs shows that accurate molecule recognition is not enough; integrating visual and textual information and chemical reasoning remain obstacles.

arXiv cs.AI (Artificial Intelligence) Original source ↗