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 WakeKV for reversible KV-cache offloading to CPU memory

The WakeKV method moves the state of less active attention heads to CPU memory and allows it to be restored. According to the authors, it reduces the frequency of missing entries compared with the methods evaluated; an implementation for the Mistral-7B model increased throughput while maintaining quality in LongBench tests.

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

Study introduces Adaptive Mutual Distillation for balanced LLM training

Adaptive Mutual Distillation jointly trains two LLMs and continually adjusts the weights of mutual distillation according to the tasks. According to the authors, combining the two models delivered an average improvement of 2.91 points over multi-task SFT across three base models; the tests included six benchmarks.

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

Study: Reasoning in German matched default English reasoning in German-language RAG

A study of German-language RAG for the world of The Dark Eye found that enforced reasoning in German outperformed reasoning in French, but only matched the default reasoning in English. The benefit increased with richer context that accounted for document structure. The authors released the testing environment and benchmark.

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

Study introduces the CLIMB method for selecting evidence and refining answers in multimodal RAG

The CLIMB method selects complementary pieces of textual evidence and accepts a revised answer only if estimated confidence increases. According to the authors, it improved results on Encyclopedic-VQA and InfoSeek without additional training; the available abstract does not specify the size of the improvement.

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

Study introduces Sigma continuous diffusion language models

The authors introduce Sigma models with 3 and 8 billion parameters. According to the study, they achieve results comparable to those of discrete models in mathematics and programming; continuous trajectories make it possible to control the balance between output quality and diversity and explore inference speedups.

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

Study identifies clinical concept hubs in eleven LLMs

The authors identified clinical concept hubs in all 11 open-weight models examined. According to the abstract, targeted manipulation of these hubs improved performance in simulated clinical settings; hub activation also predicted doctors' preferences in a blinded evaluation.

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

MintFlow enforces constraints with minimal changes to the flow matching trajectory

The authors introduce MintFlow, a method that modifies the intermediate state of generation without additional training so that the output meets the specified constraints. According to the abstract, it preserves the original data distribution better than the methods used for comparison; the excerpt does not provide specific numerical results.

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

HyMLRaman identifies six drugs from Raman spectra

The authors introduced HyMLRaman, a method for identifying six drugs. The combination of EfficientNet-B3 and SVM achieved an accuracy of 96.31 % in stratified ten-fold cross-validation. Generating additional data using DDPM helped only some classifiers.

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

Google researchers propose a method against test memorization in AI agents

Research from Google Cloud AI Research addresses the problem where AI agents that optimize their own harness overfit too much to tests. The new RRSI method regulates optimization with a budget of edit changes and a filter against task-specific tricks. Result: generalization is preserved without increasing compute costs.

The Decoder (daily AI news) Original source ↗
Research confirmed by 2 independent sources

NASA and IBM released an open-source foundation model for lunar science

NASA and IBM released the Lunar Foundation Model, an open-source foundation model for lunar science. Trained on 2 million tile bundles from 17 years of data from the Lunar Reconnaissance Orbiter and other missions. The model specializes in ice and crater detection.

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

Root.cz · and 1 more source Original source ↗
Research only one source so far

Chinese AI models are subject to state doctrine, study by Aleph Alpha finds

Aleph Alpha created a benchmark covering 967 politically sensitive topics. Tests of the Qwen, DeepSeek, and Kimi models showed that only 17–41% of responses were balanced; the rest repeat state doctrine, evade the question, or refuse to answer. The bias also shows up in unrelated questions. Nvidia Nemotron contains approx…

The Decoder (daily AI news) Original source ↗