#machine learning

Google deployed an agentic AI peer-reviewer at ICML and STOC that reviewed approximately 10,000 papers with a 30-minute turnaround. A new research paper shows it detects 34% more mathematical errors than zero-shot prompting. This demonstrates a scalable, automated approach to scientific peer review, potentially reducing review bottlenecks and improving error detection. It sets a precedent for AI-assisted review at conference scale, which could impact how research is evaluated. The system uses an agentic framework that iteratively reasons about papers, combining retrieval and verification steps. It achieved the 34% improvement over baseline zero-shot prompting, which asks the model to review without additional guidance.

IMGNet introduces a face verification model that replaces cosine similarity with sliding window sign pattern matching, achieving 96.27% accuracy on LFW with a 10.58 MB model trained on CASIA-WebFace. This novel approach could enable more efficient and compact face verification systems, and the finding that sign pattern matching improves ArcFace embeddings without retraining suggests a fundamental property of well-trained face embeddings. The model uses a SW Block replacing standard convolution with multi-scale relational operations, an IMG Sign MSE Loss defined purely over sign pattern agreement, and a voting system combining three metrics. Applied to ArcFace (buffalol) without retraining, IMG Sign Score achieves 99.58% on LFW, only 0.24% below ArcFace+Cosine.

Apple announced Trust Insights, a new on-device anti-fraud feature in iOS 27 that analyzes user behavior patterns to detect scams without reading personal content like messages or photos. This feature enhances user security against phone scams while maintaining strong privacy by processing data entirely on-device and only sending a single anonymized output to servers. Trust Insights uses device sensor data and contextual analysis to identify suspicious behavior like being guided by a scammer to transfer money or change accounts; it can trigger warnings, delays, or extra authentication before transactions.

Researchers introduce Contrastive Decoding Diffing (CDD), a method that recovers verbatim fine-tuning data from large language models using only logit differences between base and finetuned models, without requiring weight or activation access. This technique significantly improves over prior white-box methods, achieving high verbatim recovery scores on the SDF benchmark, and has important implications for ML safety and interpretability by enabling data leakage detection without full model access. CDD uses a single default configuration across model families ranging from 1B to 32B parameters, and on 19 out of 20 model pairs it achieves a verbatim recovery score of 4 or higher out of 5, whereas the prior Activation Difference Lens (ADL) method never exceeds 3 out of 5 despite requiring full weight access.

A free interactive tool called The Global Research Space now visualizes 11 million academic papers using SPECTER2 embeddings and UMAP projection, with temporal slicing and daily updates. This helps researchers quickly grasp macroscopic trends across vast scientific literature, making it easier to stay current with daily paper floods. The map uses voronoi partitioning around high-density peaks at multiple depth levels, supports keyword and semantic queries, and includes an analytics layer for ranking institutions, authors, and topics.

MathFormer, a 4-million parameter seq2seq model, achieves 98.6% accuracy on symbolic math expansion tasks without any prior mathematical knowledge, suggesting it learns structural token transformations rather than true reasoning. This finding challenges the common assumption that large language models (LLMs) 'reason' mathematically, implying their performance may stem from sophisticated pattern completion. Understanding this distinction is crucial for developing models with genuine reasoning capabilities. The model uses a GPT-style transformer architecture and is trained solely on token-level sequence mapping from factorized to expanded polynomial expressions, without any encoding of mathematical operators or variable semantics.