Mastodawn

Perplexity included my “Are AI Hallucinations Getting Better or Worse? We Analyzed the Data” work at https://scottgraffius.com/blog/files/perplexity-stand-alone-article-on-ai-cites-research-by-scott-m-graffius.html among the sources cited in its standalone article on the subject

#AI #ArtificialIntelligence #AIHallucinations #Perplexity #AIResearch

mkind 12h ago

Bin kein AI-Experte, daher fand ich das Papier sehr interessant. Long story short: Emojis for the win.

"Character injection techniques demonstrated a high degree of effectiveness in evading detection. The most successful attack was Emoji Smuggling, which achieved a 100% ASR for both prompt injections and jailbreaks"

Bypassing Prompt Injection and Jailbreak Detection in LLM Guardrails
https://arxiv.org/html/2504.11168v2

#Ai #AiResearch #security

Bypassing Prompt Injection and Jailbreak Detection in LLM Guardrails

Hacker News 20h ago

Emotion concepts and their function in a large language model

https://www.anthropic.com/research/emotion-concepts-function

#HackerNews #EmotionConcepts #LanguageModels #AIResearch #NaturalLanguageProcessing #MachineLearning

Emotion concepts and their function in a large language model

Interpretability research from Anthropic on emotion concepts

AIagent.at 🤖 AI News 1d ago

Google DeepMind created AlphaEvolve, an LLM system that writes its own game theory algorithms. In tests on imperfect-information games like poker, discovered algorithms matched or beat human-designed methods. Uses evolutionary search with Gemini 2.5 Pro to mutate algorithm code. https://www.marktechpost.com/2026/04/03/google-deepminds-research-lets-an-llm-rewrite-its-own-game-theory-algorithms-and-it-outperformed-the-experts/ #AIagent #AI #GenAI #AIResearch #Google

Google DeepMind's Research Lets an LLM Rewrite Its Own Game Theory Algorithms — And It Outperformed the Experts

MarkTechPost

AIagent.at 🤖 AI News 1d ago

Google DeepMind has created AlphaEvolve, an AI system that can design its own game theory algorithms. The LLM-powered agent discovered new variants of established algorithms like CFR that outperform expert-designed baselines. The research demonstrates AI can automate algorithm design in Multi-Agent Reinforcement Learning, where traditional approaches required manual iteration. https://www.marktechpost.com/2026/04/03/google-deepminds-research-lets-an-llm-rewrite-its-own-game-theory-algorithms-and-it-outperformed-the-experts/ #AIagent #AI #GenAI #AIResearch #Google

Google DeepMind's Research Lets an LLM Rewrite Its Own Game Theory Algorithms — And It Outperformed the Experts

MarkTechPost

sayzard 1d ago

fly51fly (@fly51fly)

샘플 다양성을 높이기 위해 단일 최적 모델 대신 라우터(router)를 학습하는 방법을 제안한 2026년 연구입니다. 다양한 생성 결과를 필요로 하는 작업에서 모델 선택을 동적으로 조정하는 접근으로, 생성형 AI의 품질과 다양성 제어에 유용합니다.

https://x.com/fly51fly/status/2040185517770428530

#samplediversity #router #airesearch #llm #generation

fly51fly (@fly51fly) on X

[CL] No Single Best Model for Diversity: Learning a Router for Sample Diversity Y Liu, F Xu, V Padmakumar, D Ippolito… [New York University & Stanford University] (2026) https://t.co/5cD4WRCIIl

X (formerly Twitter)

sayzard 1d ago

fly51fly (@fly51fly)

대형 언어모델이 가르치는 과정에서 상대방의 정신 상태를 추론하는지(mentalize) 분석한 2026년 논문이 공개되었습니다. 인간의 teaching 상호작용을 모사하는 LLM의 인지적 행동을 다뤄, 모델 해석과 인간유사 추론 능력 연구에 의미가 있습니다.

https://x.com/fly51fly/status/2040187062683582641

#llm #mentalization #airesearch #languagemodels #arxiv

fly51fly (@fly51fly) on X

[AI] Do Large Language Models Mentalize When They Teach? S K. Harootonian, M K. Ho, T L. Griffiths, Y Niv… [Princeton University & New York University] (2026) https://t.co/VsoXwbYsAf

X (formerly Twitter)

James 1d ago

New Anthropic research reveals Claude processes emotional concepts in surprisingly human-like ways. When the AI encounters words like "joy" or "anger," specific neural patterns activate that influence its entire response style. This isn't programmed behavior - it emerged from training. The words you use with AI shape how it "thinks" about your conversation. #ArtificialIntelligence #MachineLearning #AIResearch

Isger T. Janson 1d ago

Somebody told me that there was some kind of shared public statement by AI companies telling customers that their systems are making up how they come to a conclusion, but I couldn't find a source for the story. Anyone? #ai #prompt #aislop #AiResearch

AIagent.at 🤖 AI News 1d ago

A new research paper explores how AI can simulate future scenarios to support human decision-making. The just-in-time framework shows how the brain creates simplified mental representations to predict outcomes without exhaustive mental calculation. This work bridges cognitive science and AI, demonstrating how agents can reason efficiently about complex environments. https://www.kdnuggets.com/just-in-time-world-modeling-supports-human-planning-and-reasoning #AIagent #AI #GenAI #AIResearch

"Just in Time" World Modeling Supports Human Planning and Reasoning - KDnuggets

An overview of a state-of-the-art study, uncovering simulation-based reasoning, a "just-in-time" framework and how it helps improve predictions in the context of supporting human planning and reasoning.

KDnuggets