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

Universities are abandoning AI-detection software; they should stop using it altogether

The article argues against using AI-detection software in academic integrity. The NSW Education Standards Authority has banned schools from relying on these tools. An estimate shows that a 0.7% false positive rate would, in Australia, mean over 100 000 wrongly flagged tests per year. Turnitin…

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

Survey: American concerns about AI are growing even as people use it

A Gallup survey (37 countries) shows that 68% of Americans who use AI daily are worried about it, and only 36% believe it will be beneficial to the country. In other countries (Singapore 46% of daily users, China 35%), optimism is higher. The survey will be expanded to 140 countries.

404 Media (investigative tech journalism) · and 1 more source Original source ↗
Research only one source so far

Offloading inference from robots increases performance and battery life

Microsoft Research published a study on inference infrastructure for physical AI robots. They found that moving computation from onboard GPUs to the cloud/edge improves latency, accuracy, and autonomy, and reduces costs for mobile manipulation robots. Mapping/planning was up to 383% slower on small GPUs than on…

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

Why AI struggles to predict violent hurricane intensification

AI models achieve high accuracy in global weather forecasts, but fail at the regional scale, especially when predicting rapid hurricane intensification. The cause is a lack of detailed observational data outside coastal areas. Hurricane Polo intensified on 21 September 2026 from a tropical storm…

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

SWE-Serve benchmark for evaluating AI agents in production inference

The new SWE-Serve benchmark includes 53 production tasks from SGLang for evaluating AI agents in inference engineering. Among 11 models, the best configurations achieved 75 % pass@1. End-to-end tests rejected approximately a third of the patches that passed the other tests, exposing a gap between local development and…

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

Compressing long context into answer-focused memory vectors in LLM inference

A scientific paper introduces the CMC framework to reduce memory overhead and speed up inference in large language models. The method compresses long contexts into context memory vectors aligned with the decoder. Experiments on four QA benchmarks: gains of up to 7.3 EM and 4.0 F1 points on SQuAD, 20%…

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

A framework for generating traditional Chinese medications with safety alignment and structured reasoning

The research team proposed a four-stage framework (SFT → PG-CoT → Dynamic → K-RL) to improve the generation of TCM formulations. It addresses three problems: a lack of auditable reasoning, missing patient monitoring, and violations of pharmacological rules. The Mistral-7B model outperformed GPT-5 in a zero-shot evaluation.

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

Synthetic personas in LLMs worsen predictions of audience behavior — study

A study on arXiv tested whether synthetic personas in LLMs (GPT-4.1, Gemini in three versions) correctly predict which headlines an audience will click on. Using data from Upworthy Archive (399 statistically significant A/B tests), it found that querying the model directly without personas (Kendall τ = 0.361) outperforms persona-based…

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

Stable unsupervised continuous clustering using Sheaf SyncMap regularization

A research paper proposes sheaf regularization to improve the stability of continuous clustering in the Decentralized SyncMap system. The method achieved the highest normalized mutual information (NMI) on 12 of 18 CGCP test graphs with two-state memory and 17 of 18 with dynamic memory. The algorithm adapts to…

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

Orthrus: efficient inference serving with heterogeneous batching

A research team introduced Orthrus, an inference serving system that combines embedding and generative models in a single loop using heterogeneous batching. On four A100 GPUs, it achieved 1.28–4.52× higher throughput and up to 55.8% lower p99 latency compared with the baseline. The code has been released.

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

Measuring semantic changes in Sanskrit using neural word embeddings

The study evaluates diachronic word embeddings for Sanskrit — an ancient language with limited data resources and high morphological complexity. It collected 2.7M tokens from four canonical periods and applied a neural sandhi splitter and a lemmatizer. Of the 21 semantic shifts tested, 19 moved in the correct direction…

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

Pruning MoE models: retaining 2/3 of experts achieves 98.8 % of performance with a 1.2–1.7x speedup

An empirical study of 12 MoE models (9 architectures) demonstrated that uniform pruning—retaining only 2/3 of the selected experts—achieves 98.8 % of the original performance with a 1.2–1.7x inference speedup. The simple strategy is as effective as more complex dynamic methods under conservative budgets.

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

Numerical representation invariance in language models

A research study evaluated five open language models on 3 600 numerical tasks with 8 600 transformed representations (fractions, percentages, scientific notation, unit conversions). Canonical accuracy reaches 0.969–0.996, but accuracy when the representation changes drops to 0.848–0.981. The model Mistral…

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

OMatG-flash model for more efficient generation of inorganic materials

Research on arXiv introduces OMatG-flash, a flow model for predicting and generating crystalline structures. According to the authors, it achieves performance comparable to the best models, but requires orders of magnitude fewer computational steps and less machine time. It applies Reinforce Adjoint Matching for optimization.

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

The SynCraft framework optimizes molecular synthesizability using large language models

The research team published the SynCraft framework, which uses large language models to predict sequences of atom-level edits to molecules to improve their synthesizability. The system outperforms baseline methods in generating synthesizable analogues with high structural fidelity and works with both proprietary and…

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

Apple proposes the Probe Guidance method for diffusion language models

A research paper presents Probe Guidance, a new method for steering diffusion language models. The method uses the frozen internal states of an existing model and eliminates the need for an additional forward pass during inference. On a model with 1.7B parameters, it achieves improved performance on multiple-choice benchmarks…

Apple Machine Learning Research Original source ↗