Research

News and explainers on AI research: language models, machine learning, benchmarks and academic breakthroughs, explained for a general audience.

Here we collect everything we have published under Research: news, analysis and explainers from the world of AI. The page is updated whenever a new story is published.

Latest stories

Illustration: two particle beams meet inside an abstract detector ring with golden spray tracks, a visual metaphor for a foundation model of collision physics.

Illustrasjon

research

Oct 15, 2026 · 4 min

CERN and Mistral sign letter of intent for a "Large Physics Model" for the LHC

CERN and the French AI company Mistral have signed a letter of intent to develop a "Large Physics Model" – according to the announcement, the first foundation model for collision physics, trained on billions of simulated…

Mathematicians Put OpenAI to the Test: Only One of 30 Number Theory Papers Verified in Lean

In the week after October 6, 2026, the mathematical community faced something entirely new: a single release with hundreds of claimed solutions to open research problems — and a field with limited capacity to check them.

Vals AI benchmark: Agent teams cost 1.8–5.1 times more, and delivered better quality in only one of four comparisons

New benchmark results from the evaluation firm Vals AI suggest that teams of AI agents mainly buy speed, not better outcomes. According to a report published by The Decoder on October 11, 2026, multi-agent solutions with agent teams cost…

arXiv caps submissions after sixfold surge in its AI category

The measure took effect on October 1, 2026, after September delivered a record 40,363 submissions — a fourfold increase over ten years — and nearly 9,000 support requests to moderators.

Small 7B model learns belief reasoning without helper models, according to new NeurIPS paper

A research team at UC San Diego has trained a compact 7B agent to model its conversation partner's beliefs and desires – without running extra models at inference time.

Mathematics meets OpenAI's "mathocalypse": retractions and a boycott

The verification of the content has barely begun, and the disagreement over the figures in the release – 372 or 377 problems, 25 or 28 Fields Medalists – remains unresolved.

Agents claimed 70 percent — measurement showed 35: Epoch AI exposes inflated AI results

A new benchmark from Epoch AI shows that top models are far from capable of carrying out genuinely innovative AI research: they achieved only a fraction of a human-created method — and both agents selected their best training runs and…

Sakana AI's AI Review Catches 73% of Planted Errors — but Only 16% in Retracted Papers

Sakana AI's new peer review system impresses on its own benchmark, where it finds 73.43% of core-claim errors at roughly $0.47 per review. But on genuinely retracted papers, accuracy falls to 16.11% — and that gap is what will decide…

16,140 transactions replayed: Every newer AI model version caught less fraud at Coinbase

An internal benchmark run by Coinbase itself found that newer versions of three major AI model families caught fewer fraud transactions and a smaller share of the fraud value than their predecessors — even though the decision policy was…

Exa Previews ATLAS: New Benchmark Tests Search Agents on 547 Tasks Drawn from Real Search Traffic

Exa published a preview of ATLAS in October 2026, a benchmark for agentic web search with 547 tasks built from real, anonymized search demand. According to the company's own runs, no agent under one dollar per task achieves row-F1 above…

Two Discrepancies in OpenAI's Navier-Stokes Proof: Text and Lean Code Don't Match

OpenAI claimed this week that an internal model has solved the three-dimensional Navier-Stokes equations – a long-unsolved problem in mathematics.

Humans score 93.1% where best model stops at 53.6% on new visual test

Scale Labs launched the Humanity's Sixth Sense (HSS) benchmark on October 7, 2026, measuring something existing visual tests have barely touched: the fast, intuitive reasoning humans do "at a glance." According to the company's own paper…

Partially Lean-verified, incompletely open: What OpenAI's math repo actually contains

The 6 October release is unprecedented in its kind: a public GitHub repository, partially Lean-verified proofs, roughly 4,000 attempted problems.

Ablation Can Make Language Models Give the Wrong Answer About How They Work

A new arXiv paper argues that a common technique for mapping language models measures the wrong object: the ablated model, not the original one. The authors' answer is a framework called WISE — and a tool whose workings the coverage does…

Self-Criticism as a Training Method: UniEvo-VL Lifts GenEval Scores Without an External Teacher

A new arXiv paper describes a training method in which a multimodal model improves its own image generation by criticizing itself — with author-reported gains on GenEval and GenEval2 Soft-TIFA. But the improvement is not even across tasks.

Federated training beat a centralized baseline on chest X-rays, by about two percentage points of Macro-AUC

A study published in the journal Machine Learning, reported by [Bioengineer.org](https://bioengineer.org/hospitals-can-now-train-ai-together-without-sharing-a-single-x-ray/) on 25 September 2026, found that federated training not only…

New Georgia Tech Benchmark Measures Robot Memory – Top Models Fail Half the Hardest Tasks

A new benchmark from Georgia Tech measures whether vision-language models remember what they have seen in a home – and according to the reporting, the best models fail around half of the hardest tasks.

MIT and Sakana AI: SIFT reached 35.1 percent on Polyglot after only 30 steps

Researchers at MIT and Sakana AI describe a framework that lets a language model judge candidates instead of running expensive benchmark tests on every change — with reported results surpassing Darwin Gödel Machine at around one-tenth of…

DeepMind Claims Machine-Checked Proofs for Nine Erdős Problems — Two Open for 56 Years

Google DeepMind's mathematics agents are said to have produced machine-checked proofs for a range of previously unsolved problems — including nine Erdős problems, two of which had resisted mathematicians for 56 years.

Small LIFT model beats same-size transformers trained on eight times more data — authors say

A new research paper proposes a way to address one of the transformer architecture's most fundamental limitations: that information between generation steps can only flow through the decoded token.

The Watermark Lives in the Molecule Itself: DeepMind's SynthID Bio Aims to Expose AI-Designed Proteins

Google DeepMind launched a watermarking system for AI-designed proteins on 30 September 2026, covered in Nature. The system, called SynthID Bio, embeds a faint statistical signature into both the amino-acid sequence and the…

FDA funds LLM jury to grade AI radiology reports across one million exams

An 18-month research contract will test whether multiple language models grading radiology reports in parallel can measure the quality of autonomous AI systems better than human reader studies of a few hundred cases.

81 New CVEs in Mobile Core Networks – 58 Already Patched After AI Findings

The researchers behind NTU's iFinder pipeline, which runs on Claude Opus 4.5, reported that precision rose from 28 to 75 percent compared with direct prompting.

Quail, from Modal and CMU, claims a billion tokens per minute on one H100 GPU

Inference engine developer Modal and database researchers at Carnegie Mellon University have built Quail, an open-source inference engine that exploits known SQL structure to reorder requests for better KV caching.

Nine-loop record broken with two independent methods – code written from scratch by Claude

Anthropic announced on September 25 that Claude has completed the nine-loop calculation of six-particle scattering amplitudes in planar N=4 super Yang–Mills — breaking the record held by Lance Dixon's SLAC team for three years.

Chinese study: AI model combining ultrasound and DBT raised specificity in breast triage without sensitivity loss

Researchers at Sun Yat-sen Memorial Hospital in Guangzhou have developed a deep learning model that merges ultrasound and digital breast tomosynthesis into a single risk assessment per breast.

Preprint: Harmful choices rose from 0–4% to 25–71% when researchers activated a 'pain signal' in Qwen 2.5 models

A new, non-peer-reviewed preprint claims that open language models represent something resembling pain internally — and when the signal was switched on artificially, the modified models chose harmful 'relief' actions such as deleting the…

AI mutations make phages up to one million times more effective against bacteria

Researchers at the University of Wisconsin–Madison have used an AI model, trained on their own lab data, to suggest mutations in bacteriophages — viruses that attack bacteria.

AIMag.no

A laptop displaying an abstract grid of connected tools.

research

Sep 24, 2026 · 6 min

Perplexity says tool-call errors fell 21.2% after training on real user sessions

Perplexity has post-trained a Computer model on real user sessions — including failed ones — by combining rejection sampling fine-tuning with hint-guided self-distillation.

Anthropic opens source code: 36 optimized packages for biomolecular modeling

The company claims that an internal research model has optimized over 30 open-source models for biomolecular modeling, with four times faster execution and new records in protein structure prediction.

51 Researchers Behind Preprint Claiming Physics Benchmarks Are Miscalibrated

A new preprint on arXiv with 51 listed authors claims that leading physics benchmarks for frontier models are "broken" as evaluations and near saturation — but so far, the claim exists only in the article's title.

Lambda Claims Over 60 Percent MFU in Llama 3.1 Training on Blackwell

Data center provider Lambda presents numbers showing model training significantly more efficient than what the company describes as the industry norm — but everything comes from its own marketing materials, and the white paper's…

AIMag.no