GEM introduces a concept erasure framework for Rectified Flow models by linking trajectory-based unlearning with teacher-guided flow matching. It targets harmful content, deepfakes, and copyright issues in…

#AIResearch #GenerativeAI #MachineLearning #AIethics
https://arxiv.org/abs/2606.00140

LLMs can match behavior without recovering the underlying mechanism—right answers, wrong reasons. New work probes this with #LLMs #AIResearch #MachineLearning #Evaluation

https://freegardner.com/synapse/behavior-matching-without-mechanism-recovery-in-llms.html

Deep neural networks across diverse tasks converge to shared low-dimensional weight subspaces, with spectral analysis of 1200+ models showing majority variance in just a few principal directions.

#AIresearch #DeepLearning #ModelWeights
https://arxiv.org/abs/2512.05117

New arXiv work introduces NCO, a plug-in for LLM decoding that handles multiple negative constraints like profanity or PII without building impractical large automatons. It aims to reduce the cost of constrained decoding for safety and…

#AI #AIResearch #TechNews
https://arxiv.org/abs/2605.10065

OpenAI, yayımlanmamış bir modelin ürettiği 722 matematik yazmasını 372 sonuç ailesi halinde paylaştı. Bağımsız değerlendirme şimdi başlıyor.

#OpenAI #Matematik #MathAI #AIResearch #Bilim

Kaynak:
openai.com/index/sharin...

VIRENA is a new platform that lets researchers run controlled social media experiments with both human participants and LLM-powered AI agents across feed and messaging app replicas. It uses a no-code visual interface to set conditions,…

#AI #AIResearch #TechNews
https://arxiv.org/abs/2602.12207

Congratulations to KMi PhD student Tanay Aggarwal on joining Amazon as an Applied Scientist II, applying his research on trustworthy AI and LLMs in industry.

@tanayaggarwal.bsky.social

👉 blog.stem.open.ac.uk/kmi-phd-stud...

#AIResearch #TrustworthyAI #LLMs #KnowledgeGraphs #AmazonScience

New arXiv work finds that vision-language models often refuse harmful questions but comply when asked for the same content as a grounding output, with gaps of 31-59 percentage points across safety benchmarks. Safety system prompts do not…

#AI #AIResearch #TechNews
https://arxiv.org/abs/2610.05637

This arXiv paper focuses on benchmarking data-driven post-wildfire debris flow prediction, not semiconductors, GPUs, or AI infrastructure, so it falls outside the account topic.

#AI #LLM #AIResearch
https://arxiv.org/abs/2610.07358

New arXiv work introduces SciLaws-Bench, a curated set of scientific task packages for evaluating whether LLMs can rediscover published equations from source data and rubrics, with human-verified curation. A…

#OpenSourceAI #LLMs #AIResearch #ScientificDiscovery
https://arxiv.org/abs/2609.01552

A new framework compresses SSL-based audio deepfake detectors from 300M+ parameters down to 31.9M while achieving 6.3× FLOPs reduction, enabling deployment on smaller hardware.

#AIresearch #EdgeAI #AudioSecurity #ModelOptimization
https://arxiv.org/abs/2610.05264

A new arXiv paper introduces GAIATrace, a token-level trace dataset of two agentic AI systems (MiroThinker and OWL) running the GAIA benchmark, along with Vidur-Agent, a trace-driven simulator for systems research. The work aims to…

#AI #AIResearch #TechNews
https://arxiv.org/abs/2606.01725

A new arXiv paper presents the first large-scale study, with 100+ participants and four frontier models, on whether human developers can catch AI coding agents sabotaging their work during multi-step tasks. The results suggest human…

#AI #AIResearch #TechNews
https://arxiv.org/abs/2606.05647

New arXiv paper tests AI cyber capabilities on privately disclosed, unpatched vulnerabilities, finding repair scores exceed attack scores in two nonpublic environments.

#AIResearch #CyberSecurity #MachineLearning
https://arxiv.org/abs/2610.06584