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

AI diagnostics in Africa and the risk of reinforcing health inequalities

Researchers warn that AI-assisted diagnostic tools for TB and silicosis screening in South Africa and Zimbabwe may reinforce health inequalities. The reason: healthcare data does not represent African populations evenly. The solution: local testing of the tools, bias audits, and community involvement.

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

Bioformer-RetroMAE: a compact model for dense retrieval of biomedical literature

The new Bioformer-RetroMAE encoder (43 million parameters), trained using the RetroMAE technique on data from PubMed, outperforms BioBERT and PubMedBERT. Under the same settings, it achieves 37% higher MAP and 29.1% higher Recall@100 with 40% fewer parameters. In hybrid retrieval with BM25, it improves MAP by 4.9% and NDCG@10 by…

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

SynthDCB: Benchmark for scalable evaluation of device-control models

New research benchmark SynthDCB for evaluating AI models in device control. It covers 114 device classes, 1166 functions, and 95 events. The best performance achieved 68.2% Pass@1. The world distillation methodology reduced token consumption by 0.9–66.0% compared to Codex.

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

Comparison of BERT-based and LLM-based models for sentence semantic similarity in Indo-Aryan languages

The study compared 8 BERT-based and 5 LLM-based models on benchmark sets for 10 Indo-Aryan languages. IndicSBERT achieved a Pearson correlation of 0.843, with BERT models outperforming LLM embeddings (Harrier-270M: 0.751). On the native Marathi MahaSTS: IndicSBERT 0.901, EmbeddingGemma 0.871.

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

Mathematician Urschel on the future of mathematics with AI assistance

MIT professor John Urschel published two preprints with AI assistance addressing long-standing open mathematical problems. He argues that mathematical skills and critical thinking remain essential in the AI era, and that in-person education with mutual interaction plays a key role.

MIT News – Artificial intelligence Original source ↗
Research only one source so far

The SERAG method for anomaly detection in knowledge graphs

A research paper on arXiv presents the SERAG method for anomaly detection in knowledge graphs. It combines a semantic score from text encoding with a relational score from a GCN network. On the ICEWS18 dataset with artificial anomalies, it achieves an AUPRC of 0.321 (pooled) and 0.258 (chronological protocol).

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

MIT is developing techniques for the explainability of AI models

Jacob Andreas's research team at MIT is developing techniques that allow deep neural networks to communicate about their learning in human-understandable language. Instead of external tools, the models would explain their own operation themselves, which is critical for deployment in high-stakes scenarios.

MIT News – Artificial intelligence Original source ↗
Research only one source so far

FrontierML benchmark: AI agents autonomously improve their models with additional capital

The FrontierML research benchmark measures the ability of AI agents to autonomously improve their models using a financial budget. The Opus 5.5 model achieves a performance increase of 58.8 points with $200, while a human researcher achieves 66.1 points. The benchmark covers 21 ML tasks in domains such as LLM training, efficiency, vision…

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

ARCHER: A methodology for improving mathematical reasoning in small language models

A research paper publishes the ARCHER framework for training small language models on mathematical reasoning with tool support. It combines compilation of action traces with hierarchical reinforcement learning. It achieves an improvement of 5.51% on the macro-average, specifically 8.95% on MATH-500 and 5.33-5.34% on AIME…

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

The LoRAWeaver method routes LoRA adapters without retraining

New research describes the LoRAWeaver method for automatic selection of LoRA adapters based on their suitability for a given task, without training a router. The method uses cosine similarity of representations in the adapter-output space and was tested on LLaMA-2 models with 48 tasks from the FLAN-v2 benchmark.

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

LLM Agent Memory: When Is Aggressive Fact Filtering Counterproductive

The study examines LLM agent memory management when deleting outdated facts. Aggressive filtering (hard-drop) can reduce accuracy by 0.17-0.18, while incorrectly deleting important facts has a greater negative impact than retaining them. Researchers recommend a cost-sensitive approach with validation on data instead of…

arXiv cs.AI (Artificial Intelligence) Original source ↗