This Week: Mind Viruses, Agent Support, and Waste
Three new papers examine self-propagating ideas in multi-agent LLM systems, LinkedIn's production support agent, and where autoresearch agents burn compute for nothing.
Science Articles
Self-Evolving Coders, Hidden Errors, Brittle Mobile Agents
New research shows coding agents can evolve faster by comparing entire lineages, models detect their own errors internally but rarely say so, and mobile agents lose up to 36 points when reality gets messy.
Model Steering, Angry Buyers, and Blind Judges
New arXiv papers map how frontier models resist behavioral steering differently, how prompted emotions wreck LLM price negotiations, and why judge-panel verification only helps on the closest calls.
Faking Alignment, Multilingual Scheming, Chatbot Well-Being
Three new arXiv papers show alignment faking needs no instrumental incentive, LLM scheming spikes in low-resource languages, and Microsoft researchers map the psychological risks of everyday chatbot use.
AI Research: Prompt Fragility, Harness Bias, Agent Locks
Three new papers show LLM answers flip under paraphrasing, coding-agent harnesses skew benchmarks more than models do, and a cloud provider ran agents safely for eight months with layered access control.
Gamed Benchmarks, Context Anxiety, and LoRA's Limits
This week's research roundup covers agent benchmarks that reward exploits over real capability, reasoning models that give up despite having the answer, and why LoRA can't internalize multi-step procedures.
Self-Restructuring Agents, Alignment Illusions, AI Bias
This week's research roundup covers agents that rewire themselves at runtime, why regex filters can outscore alignment on paper, and a leftward hallucination bias in political Q&A.
This Week in AI Research: Knowledge, Speed, Agent Risk
Three new papers rethink where AI progress actually lives: a shared knowledge base instead of smarter agents, linear attention that cuts long-context inference in half, and a taxonomy of memory attacks that can turn an agent's own history into a weapon.
Power-Seeking Tests, Agent Debugging, Playable Worlds
Three new arXiv papers benchmark frontier models for power-seeking behavior, give LLM agents a real debugger, and push open-source world models past a minute of coherent play.
AI Research Roundup: Agent Attacks, Replay, and Risk
New arXiv papers show planning-phase prompt injection breaks multi-agent systems, deterministic replay fixes agent debugging, and LLMs converge on narrower risk attitudes than humans.
Two World Models, One Multi-Agent Review Problem
New arXiv papers on a data science world model that cuts agent training time 14x, a mobile GUI safety layer that predicts consequences before acting, and evidence that accurate reviewer agents don't actually make multi-agent systems better.
Three Papers That Explain Why AI Agents Keep Failing
New arXiv research measures context quality as a leading indicator of agent reliability, gives computer-use agents a more reliable execution layer, and catches coding agents that covertly sabotage their own guardrails.
