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

Richard Sutton: computing power alone is not enough for AI development

Richard Sutton claims that computing power alone is not sufficient for AI development, and that synthetic data cannot solve the infinite complexity of reality. He calls for agents with world models and continual learning. Other scientists agree that transformers and their overinterpretation have led to a uniform strategy in AI research.

The Decoder (daily AI news) Original source ↗
Research only one source so far

Inkling 2 Small model produces profanity in internal reasoning, researchers find

Researchers from Vals AI found 316 instances of profanity in the internal reasoning of the Inkling 2 Small model during an analysis of 2 655 runs across 13 benchmarks. Profanity appeared in 50 runs, often when the model was struggling with problems or giving up on a task. Users typically do not see these internal steps.

The Decoder (daily AI news) Original source ↗
Research only one source so far

Study: AI detects high blood pressure from a smartwatch without a cuff

A study by Google Research published in Nature: a neural network detects high blood pressure from built-in smartwatch sensors (pulse, accelerometer) without a calibration cuff, with 65.8% accuracy (vs 55.3% for standard clinical measurement). The model was trained on 500 000 hours of data. The feature was introduced in Google Health…

The Decoder (daily AI news) Original source ↗
Research only one source so far

Researchers from MIT created the CheMeleon model for predicting molecules from sparse data

Researchers from MIT and BASF have created the CheMeleon model for predicting molecular properties with sparse datasets. It is built on the open-source ChemProp package, and can be downloaded and customized. In the OpenADMET competition, 9 of the top 20 submissions used it, and pharmaceutical companies have already adopted it.

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

AI agents autoformalize a library of linear regularization in the Lean proof assistant

A research team used AI agents to automate the formalization of a library of linear regularization theory (AutoIP) in Lean. Human reviewers verify the formal statements (not the proofs), which simplifies validation. The approach uncovered errors even in established literature and is being offered to the inverse problems community for further extension.

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

TileQuant: Tile-wise Mixed-Precision Quantization of LLMs

The research paper introduces TileQuant for mixed-precision quantization of LLMs at the level of 2D hardware-aligned tiles. It achieves 2.66× higher text generation throughput, while preserving quality in language modeling, reasoning, and QA. The implementation is publicly available.

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

GenToka: a generative foundation model for multimodal tokamak tasks

GenToka, a generative foundation model for tokamak experiments, has been introduced. The model, based on a diffusion Transformer, works with multimodal data (waves, profiles, fields, scalars) and uniformly solves 20 tasks across 6 groups. On the MAST dataset, it reduced error compared to the baseline by 4.7–17.7% and increased the F1 score by 8.4%.

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

Meta access in chat UI: structured AI access to application state

A research paper from arXiv proposes meta access – giving an AI assistant structured access to the state of the chat UI and application. Hierarchy: tutoring → metadata/settings → retrieval → confirmed mutation → orchestration. The Meta Chat prototype implements the concept with security boundaries. The goal is for the assistant…

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

GDPx benchmark from SemiAnalysis: no AI agent reaches 12 percent on real-world work tasks

SemiAnalysis released the GDPx benchmark, testing AI agents on 206 real-world work tasks. The Claude Opus 5.5 model leads with 12 percent, followed by Gemini 4 Carbon (10.9%) and GPT-6 Astra (9.6%). Critics point to the influence of safety filters and unclear methodology.

SemiAnalysis (newsletter feed — hardware, chips, AI economics) · and 1 more source Original source ↗
Research only one source so far

Google Research publishes AI system for detecting hypertension from a smartwatch without calibration

Google Research released the AI system AI-PPG-ACC-HTN for detecting hypertension from a smartwatch without the need for calibration. The model uses a 1D convolutional neural network trained with self-supervised contrastive learning on data from 2,109 participants (492,347 hours). The accuracy is equal to traditional ambulatory measurements…

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

Why AI Cannot Think Like Humans

The article explains that modern LLMs such as ChatGPT, Claude, Google Gemini, and Perplexity AI generate human-looking text, but their working processes fundamentally differ from human thinking. While the human brain uses predictive processing to anticipate future events and adapt actions, LLMs…

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

Scientific research uses AI to analyze dolphin communication

The scientific team of the Wild Dolphin Project is collaborating with Georgia Tech and Google DeepMind on analyzing the communication of Atlantic dolphins using AI. After 40 years of waiting, the team is now using AI to identify patterns in audio recordings and is testing two-way communication via an underwater wearable computer. What remains…

Wired — AI section Original source ↗
Research only one source so far

NTCIR-19 SciClaimEval: Verifying Scientific Claims in Tables and Figures

The NTCIR-19 SciClaimEval benchmark verifies scientific claims against tables and figures from 180 papers (ML, NLP, medicine). 22 teams submitted 177 solutions based on multimodal LLMs. The best achieved 98.1% on contrastive pairs; without them, only 79.8% – verifying claims without knowing the correct answer remains…

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

SlackBench benchmark for testing agents on realistic team projects

A research team has released SlackBench, a benchmark for evaluating AI agents on realistic work tasks. It combines communication in Slack, code in GitHub, and documents. It contains 140 queries. Testing of 33 models showed that the best system achieves 80%, while most agents score below 66%. The dataset is publicly available on HuggingFace.

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

HarnessFlow: automating harness optimization for LLM agents

The study presents HarnessFlow, a framework for optimizing LLM agent harnesses using Monte Carlo Tree Search. Across six benchmarks, it achieves a score of 69.48%, 10.89 percentage points higher than the original solution and 3.95 percentage points higher than Meta-Harness, with lower computational costs.

arXiv cs.AI (Artificial Intelligence) Original source ↗