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

SCOPE and SCION: A benchmark and reference pipeline for schema induction from text

A research paper introduces SCOPE, a schema induction benchmark drawn from 24 sources focused on information extraction and event extraction, and SCION, a reference pipeline for constructing schemas. SCION achieves the highest F1 score among the methods compared, including LLM-only approaches. Code and data are publicly…

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

Self-poisoning dynamics in adaptive out-of-distribution detection and certified label-free stabilization

A research paper theoretically analyzes why adaptive out-of-distribution detectors degrade themselves during operation. The authors model memory bank degradation as a generalized Pólya urn and prove that the detector collapses when the slope's reproduction number exceeds one. They propose certified…

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

Role-aware masking in diffusion-based molecular graph generation

An arXiv research paper introduces MotifRole-Diff to improve discrete diffusion for molecular graph generation. Rather than uniform masking, it assigns masking rates according to empirically measured reconstruction difficulty. On the QM9 benchmark, it increased validity from 0.905 to 0.944; on…

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

Named Entity Recognition with continual learning: the FSE model with accelerated and slowed-down experts

The article on arXiv introduces the FSE model for Named Entity Recognition with continual learning. The architecture combines a fast expert for filtering token spans and a slow expert for classifying entities. The "length-decay negative sampling" method addresses class imbalance. Results on datasets…

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

Brain waves as the key to training physical AI systems

The startup Encord is experimenting with a headset from Zander Labs that measures brain waves to create higher-quality training datasets for robotic models. The aim is to demonstrate whether biological signals indicating errors and attention can improve training for physical manipulation.

TechCrunch AI Original source ↗
Research only one source so far

Optical technology for updating AI models on robots

Researchers at Cornell Tech are developing optical technology for transferring AI model parameters using QR-code-like patterns of light. The aim is to reduce the energy demands of AI systems in data centers, autonomous vehicles, and robotics. The development was presented at the IEEE/JSAP symposium.

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

The ABBEL framework for managing context in long LLM interactions

A research paper on ABBEL, a framework that enables language models to work efficiently with long sequences of interactions. Rather than retaining the entire history, it proposes maintaining supervised "belief states" (states of knowledge). It addresses the problem of traditional summarization reducing quality, particularly on tasks such as…

Berkeley AI Research (BAIR) Blog Original source ↗
Research only one source so far

Study on arXiv: combining quantization and low-rank decomposition for LLMs is not lossless; the authors propose the DAM method

A paper on arXiv mathematically proves that low-rank decomposition and quantization – two common methods for compressing language models – are not mutually independent, as previously assumed. According to the authors, combining them leads to a significant drop in performance; they propose the DAM method to mitigate this loss.

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

Researchers have described a new type of attack on RAG systems and a defense using graph theory (TopoGuard)

Researchers have described a “split-knowledge attack" on RAG systems — combining individually harmless documents into a harmful association that filters such as LlamaGuard fail to detect. The proposed defense, TopoGuard, uses a document similarity graph and, according to the authors, detects significantly more attacks with low latency.

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

Study: language models, including GPT-5, perform no better at personalizing preferences than simple statistical methods

The Rushes dataset (44 226 decisions from 8 167 users) shows that frontier LLMs, including GPT-5, do not outperform a simple popularity-based baseline method when predicting individual choices in AI-generated stories — models tend toward majority preferences instead of personalization.

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

New SCoPE method combines a speaker's emotional history with multimodal signals for emotion recognition in conversations

Researchers introduced SCoPE, a module for Emotion Recognition in Conversations (ERC) that builds on the emotional history of a particular speaker and combines it with multimodal signals. According to the authors, the model outperforms current state-of-the-art approaches on the IEMOCAP dataset.

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