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 confirmed by 4 independent sources

Mathematician Levent Alpöge solved the Riemann Hypothesis with the help of the Claude model

Mathematician Levent Alpöge from Anthropic claims to have solved the Riemann Hypothesis (a mathematical problem dating back to 1859) with the help of the Claude model. The proof is formalized in the Lean language; the mathematical community is verifying it. Alpöge also published, together with a colleague, a proof concerning the Navier-Stokes equations.

Simon Willison — AI tag (leading independent LLM commentator) · and 4 more sources Original source ↗
Research only one source so far

Bytedance: multi-phase AI training is unnecessary, three phases are enough

A paper from Bytedance and American universities claims that training AI models for autonomous tasks does not need demanding nested phases — three steps are enough. The authors criticize the complexity of current methods and the high cost of reproduction, supporting their claims with a competitive open model.

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

Research on multimodal foundation models: MIT-IBM develops benchmarks and methods to address failures without large annotations

The MIT-IBM Watson AI Lab is developing benchmarks and data-efficient methods to address failures of multimodal foundation models. The team, led by Jim Glass, focuses on the relationship between perception and language so that models can better understand phenomena without needing a large annotated training dataset. They will develop public…

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

Meta expanded the Digital Brain Project and released UMA Nitro at a science summit

At a science summit, Meta announced the expansion of the Digital Brain Project (15K → 25K hours of brain recordings), released UMA Nitro for molecular simulations, and made UMA Playground available to chemistry teachers. Partners presented assistive robotics with DINOv3 and SAM 3 and the Lean tool for verifying mathematical proofs.

Meta Newsroom — AI Original source ↗
Research only one source so far

METR organization updates AI metric – subjective tasks remain a weakness

METR has updated its time horizon plot – a metric for measuring AI progress. The new framework includes fuzzy tasks relevant to AI research. Finding: AI is significantly better at verifiable tasks (coding, mathematics) than at subjective ones without clear criteria, because reinforcement learning only works…

MIT Technology Review — AI section Original source ↗
Research only one source so far

Bloomberg analysis: AI-exposed jobs in China see a bigger decline

Bloomberg analysis: 10 million Chinese job postings (2021–2025) show that AI-exposed roles are declining ten percent more. Chinese studies confirm: AI caused a 5.2% loss in new hires. 70 million Chinese workers are at risk. Roles are merging – one person with AI does the work of up to six…

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

Mathematicians address the future of the field after a qualitative leap by AI agents

A panel discussion at MIT brought together leading mathematicians from MIT, Harvard, Northwestern, and NYU. They addressed the impact of AI agents, which in 2026 achieved a qualitative leap in solving long-standing open mathematical problems. The panelists discussed the quality of AI solutions, changes in research methods, and future…

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

{"text_1": "Special models EMBER and ORBIT outperform GPT and Gemini in predicting behavior in Counter-Strike 2"}

{"text_0": "Czech company Mesocort AI trained an AI on data from 32 205 Counter-Strike 2 matches (271 thousand high-stakes situations from 120 thousand players). The special model EMBER achieved 73.5% accuracy in predicting player behavior in crisis situations, outperforming the products GPT-5.6 Terra and Gemini 3.7 Flash (61–63%) and…"}

Research only one source so far

Safety of Medical LLMs: From Hallucinations to Neuro-Symbolic Verification

A critical review of the safety of medical LLMs. The authors analyze how hallucinations arise, how they propagate in agentic systems, and how neuro-symbolic techniques (formal logic, ontologies, knowledge graphs) can add verifiability. They point out that regulation should audit reasoning procedures, not just…

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

Argumentation Frameworks for LLM Decision-Making: Task-Specific Methods and Evaluation

The research examines integrating formal argumentation frameworks into LLMs. Existing methods do not show consistent improvement. The authors proposed Qnt-Arg (aggregates arguments by speakers under gradual semantics) and Qlt-Arg (uses credibility estimates and grounded semantics to select evidence). Qnt-Arg showed…

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

MKA Benchmark: Testing Multimodal Agents on Professional Tasks

The MKA benchmark evaluates the ability of autonomous agents to work with visually dense documents (CAD, diagrams, maps) and produce outputs in Excel, PowerPoint, PDF, and other formats. It contains 200 tasks across 10 sectors. Claude Fable 5.2 leads with 73.3% Overall WSR and 56.6% Climb-WSR, Gemini 4 Carbon achieves…

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

LLM Complexity: A New Metric for Measuring the Inference Complexity of Large Language Models

Researchers introduced the LLM Complexity metric, which measures task complexity from the LLM's perspective by characterizing the area under the accuracy-versus-token-count curve. The metric provides a consistent ranking of tasks across different models and enables the creation of a curriculum for fine-tuning that increases accuracy…

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

Neurosymbolic framework for predicting diabetic readmissions: integrating healthcare data and clinical models

A research study on 101 766 patients from arXiv proposes a neurosymbolic framework for predicting diabetic readmissions, combining healthcare data, Bio_ClinicalBERT embeddings, and RAG extraction. The model achieves higher sensitivity (66.67%) and F2-score (0.3824), reducing hallucinations in Hindi-English text from…

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

Depth Pro 2: sharp and temporally consistent metric depth estimation

Apple's research team has released the Depth Pro 2 model for depth estimation from images and video. The model combines the strengths of single-frame and multi-frame approaches: it produces metrically accurate, sharp, and temporally consistent depth maps. The code and weights are publicly available.

Apple Machine Learning Research Original source ↗