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

Detection is not reliable control: Decodable empathy directions have only a partial effect

The study tested whether decodable empathy directions in LLMs control model behavior. Across three models (Qwen, Llama, Gemma), it found: the affective facet changes (+0.29 in Qwen, ~26 % of the natural difference), the cognitive facet does not change, and human perception remains unchanged. Conclusion: decodability is not a reliable lever.

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

GlucoFM: Foundation model for continuous glucose monitors

Google Research released GlucoFM, a foundation model for analyzing data from continuous glucose monitors (CGM). The model separates long-term glucose trends from short-term deviations using a dual-stream architecture. It achieves a PR-AUC 5.8 percentage points higher than GluFormer, trained on 109,066 hours of unlabeled data…

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

Neurosymbolic Alignment for physiologically safe clinical LLMs

The research paper presents Neurosymbolic Alignment, a method that improves the safety of a 7B clinical LLM by connecting it to a physiological world model. On Clinical Safety Benchmark, it achieves 90.8% accuracy (compared to 69.5% without the method), outperforms GPT-4 with 10× fewer parameters, and reduces medical errors from 14.1%…

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

Equivariant cellular sheaves in molecular electronic structure

A research paper combines topological deep learning with equivariant neural networks to predict molecular Hamiltonians. The new Equivariant Cellular Sheaf Networks model formalizes the Hamiltonian as a Laplace operator with O(3)-steerable kernels, generalizes existing approaches, and achieves lower…

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

IDEA Prune pipeline for more efficient pruning of language models

The research team at Apple proposed IDEA Prune, a method that combines model expansion, pruning and recovery in a single cycle with iterative pruning. Compressing 2.8B models to 1.3B achieves better token efficiency than training models of the target size from scratch.

Apple Machine Learning Research Original source ↗
Research only one source so far

The MAP framework predicts transcriptomic responses to unknown drugs using knowledge graphs

The team published MAP, a framework for predicting transcriptomic responses to unknown drugs. It combines knowledge graphs (187 089 drugs, 694 246 relationships) with a single-cell foundation model. They achieved an improvement of up to +12.3 % in predicting differentially expressed genes compared with baseline methods.

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

The MCite-RL research method improves multimodal RAG through agentic learning and citations

The study presents MCite-RL, a framework for multimodal RAG focused on precise visual citations. It combines iterative retrieval, agentic decision-making and reinforcement learning to optimize both answer accuracy and source traceability. Tested on the Wiki-VISA, FinRAGBench-V and MMLongBench-Doc benchmarks.

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

Evaluating structured perception under a token budget in multimodal agents

The research presents the ASP method for multimodal agents with a fixed token budget. It tests the method on 7 open-weight models (3B–31B) on the synthetic benchmark SEW-Bench. ASP achieves 75–94 % retrieval accuracy compared with 3–19 % for the baseline without a query; the query-conditional approach emerges as decisive…

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

Wazobia Eval benchmark for evaluating language models in Nigerian

A research team has released the Wazobia Eval benchmark with 550+ manually annotated examples for evaluating language models on their understanding of Nigerian. The dataset includes a 16-category emotion taxonomy, sarcasm detection and evaluation of cultural reasoning. The dataset is publicly available.

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

ADAPT: amortized distillation across post-trained LLM variants

The research paper presents the ADAPT framework for efficiently creating families of LLMs: it combines size-based interpolation with cross-variant optimization, enabling the creation of L × K models (L interpolated sizes, K post-trained variants) in a single distillation run instead of independently training each…

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

LLM translates driver commands into autonomous vehicle parameters

Researchers at TU Delft developed a system that uses an LLM to convert natural language commands (e.g. “I'm in a hurry, drive faster”) into adjustments to autonomous vehicle parameters. The system adjusts the motion planning algorithm with safety in mind, keeps the user involved in the decision-making process, and in simulations…

IEEE Spectrum — Artificial Intelligence Original source ↗
Research only one source so far

A weighted tree structure for memory in long-term agents with large language models

A research paper on arXiv introduces Weighted Memory Tree, a hierarchical memory technique with dynamic retention scores for LLM agents. On the GAIA-Text benchmark with Qwen3-8B, Gemma 4 E4B and Llama-3.1-8B, it achieved an accuracy improvement of 9.97 percentage points and a reduction in token usage of 32.8 % compared with…

arXiv cs.AI (Artificial Intelligence) Original source ↗