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: growth in NLP research does not mean broader coverage of underrepresented languages

An analysis of 1 158 papers from 2020–2025 introduces the Linguistic Inclusion Framework. According to the study, growth in research largely increased the representation of languages already being studied. Asian languages account for 64.1 % of speakers in the sample examined, but only 46.7 % of language representation in research.

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

Study introduces Lighthouse assistant for historical research using archival sources

The study introduces Lighthouse, an assistant for researching the American civil rights movement. It combines an LLM with a rule-based system and has access to more than 18 thousand primary sources. The authors evaluate both factual and subjective responses; the excerpt does not provide specific results.

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

Study maps security risks for multimodal AI agents in 6G networks

A review study examines threats to multimodal AI agents in 6G networks. It categorizes attacks targeting input processing, memory, tools and communication between agents, and summarizes defense mechanisms. The available abstract does not report experimental validation of their effectiveness.

arXiv cs.AI (Artificial Intelligence) Original source ↗
Research confirmed by 2 independent sources

Eko agents performed better on subsequent tasks after freely exploring a virtual island

Researchers at MIT let 13 Eko agents using Claude Opus 4.7 independently explore a virtual island for 30 hours. According to their preprint, the agents generally went on to outperform the same agents without this experience. The researchers liken the observed behavior to play.

MIT Technology Review — AI section · and 1 more source Original source ↗
Research confirmed by 2 independent sources

The MAI-DxO system from Microsoft diagnosed complex cases more accurately than general practitioners in a simulation

According to a study published in Science, the MAI-DxO system correctly diagnosed 83 % of complex simulated cases, compared with 36 % for the doctors in the comparison group. Specialists and medical teams were not included in the comparison. For emergency department cases, the system did not deliver a substantial improvement.

The Verge AI · and 1 more source Original source ↗
Research only one source so far

Experiment with Claude did not produce an antivenom, experts say

Citrini Research and Anthropic used Claude to try to design proteins for an antivenom against scorpion venom. The experiment produced 211 designs for laboratory testing. Computation cost nearly 50 000 dollars, and testing approximately 150 000 dollars. According to experts consulted by The Verge, the result was not a finished drug.

Research only one source so far

OnQ improves reasoning in Qwen3 models with low-bit quantization, according to a study

The study introduces OnQ, a method for training quantized models on their own responses under the supervision of the original model. According to the authors, at 2-bit precision it improves the average score of Qwen3-8B across nine tests by 10.26 points compared with the strongest method evaluated.

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

DiffResearch study reports sixfold speedup for research agents using a diffusion language model

The authors of the DiffResearch study report an approximately sixfold speedup in the overall process across 100 Deep Research Bench tasks when replacing an autoregressive model with a diffusion language model. The dFast mode sped up query decomposition itself by a median of 1.99 times without reducing report quality.

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

Study introduces PAWM to predict failures in LLM agents

The study introduces PAWM, which predicts failures several steps ahead and selectively guides LLM agents. The authors report performance improvements for Claude Sonnet 4.5 of 24.3 % in the ScienceWorld environment and 7.7 % in the WebArena-Lite environment.

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

DynaWeave study explores agent training for changing environments

The study introduces DynaWeave, which inserts external events into training tasks and adjusts the affected evaluation checks. The authors used 2 095 verified runs from 36 environments to fine-tune Qwen3.5 models; on CostBench, they report an increase in success rate of 67.34 percentage points.

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