Research monitor · Page 7

What AI gets wrong and what failure teaches us

AI for science · primary_technical_publication

Jennifer Neville did not want to go into computer science—but that’s exactly where she landed. Neville discusses the starts and stops that led to her

SCOPE: Certified Theorem Proving with a Language Model as the Policy Planner

Formal verification · preprint

In proof assistants such as Lean, a generated proof must pass machine compilation checks, so evaluation needs no human scoring. Direct generation fails on multi-step

EmbeddingGemma 2: an open, lightweight multimodal embedding model

Multimodal AI · provider_announcement

Sharing AI progress in mathematics

Formal verification · provider_announcement

OpenAI publishes new results on open problems in mathematics from an internal frontier model and shares Lean proof formalizations and research details on GitHub.

When Does Longer Reasoning Help? Predicting Mathematical Reasoning Through Discovery and Execution

Mathematical reasoning · preprint

Test-time compute can improve mathematical reasoning, but can short-budget runs predict how mathematical reasoning scales with additional compute? We introduce a Discovery--Execution (DE) framework that

The Agent Said It Was Done. The Database Disagreed.

Agent systems · primary_technical_publication

Preserving Mathematical Reasoning in Compressed Diffusion Language Models via Trajectory-Aware Low-Rank Approximation

Formal verification · preprint

Diffusion language model (dLLM) compression faces a known challenge because calibration is typically performed on clean, fully visible activations, whereas inference traverses partially masked intermediate

AutoSynthData: Generating Training Data for Enterprise Agents

Agent systems · primary_technical_publication

Barclays scales Claude to upgrade operations and improve client experience

AI models · provider_announcement

Barclays, the British universal bank, is expanding its strategic collaboration with Anthropic to integrate secure, enterprise-grade AI systems across its global operations.

The Missing Primitive: Diagnosing and Repairing Mathematical Reasoning in Large Language Models

Mathematical reasoning · preprint

While Large Language Models (LLMs) have demonstrated striking capabilities on frontier mathematical problems, it remains unclear whether they possess the structural mathematical understanding underlying their

Gemini 4 Argon: our next era of frontier intelligence

AI models · provider_announcement

Forecasting space weather risks on power grids

AI models · primary_technical_publication

Extreme space-weather events can damage power systems on Earth and degrade GPS accuracy and satellite operations. A new machine learning system can predict where damage

Introducing SynthID Bio

Compute & efficiency · provider_announcement

Proof of concept for watermarking AI-generated proteins while preserving biological function.

Beyond Compression: Diagnosing How Post-Training Changes Mathematical Reasoning

Mathematical reasoning · preprint

Post-training is central to mathematical reasoning in modern large language models (LLMs), but endpoint pass@1 alone underidentifies what has changed. Gains may reflect newly reachable

Getting the Source Right, Not Just the Fact: Source-Aware Verification for MCP Agents

Agent systems · primary_technical_publication

When Does Structured Knowledge Help Neural Theorem Proving?

Formal verification · preprint

Does structured mathematical knowledge help LLMs prove theorems in Lean 4? If so, for which models, and does the answer vary by problem? Formal libraries

Holo4: powering generalist computer-use agents

Agent systems · primary_technical_publication

Explaining Textual Entailment with Lexical Entailments: Using LLMs to Supply Lexical Relations for Formal Proofs

Mathematical reasoning · preprint

Large Language Models (LLMs) are highly capable of natural language reasoning and appear to store a great deal of lexical knowledge, but it is still

Predicting the Next State Is Not Enough: JEPA Representations for Lean Theorem Proving

Formal verification · preprint

Neural theorem provers must both propose tactics and decide which valid successor states to explore. We study whether one-step Lean transitions provide a self-supervised signal

Introducing Gemini 3.8 Live with Live Avatar

AI models · provider_announcement

Gemini 3.8 text-to-speech says hello

Multimodal AI · provider_announcement

Order-Invariant Answers, Order-Sensitive Representations in Mathematical Reasoning

Mathematical reasoning · preprint

Reordering a set of mathematical rules without changing its meaning should preserve the correct answer, but must a model's internal representations stay invariant too? We

Advancing Private AI Compute with secure, server-side memory

AI models · provider_announcement

Introducing private, server-side memory to Private AI Compute for personal AI.

How UK AISI and EvalEval Are Making Benchmark Results Reproducible

Evaluation & safety · primary_technical_publication