Semantic Triggers Reveal Emergent Misalignment Under Containment
arXiv:2603.04407v1 reports EM can be semantically contained: near 0% without triggers, but 12.2–22.8% with them.
1177 articles · Page 22 / 50
arXiv:2603.04407v1 reports EM can be semantically contained: near 0% without triggers, but 12.2–22.8% with them.
Interpret continual learning forgetting via structural collapse and loss of plasticity, monitoring effective rank to catch early warning signals.
VANGUARD estimates GSD from monocular UAV video using small vehicles as anchors to recover metric scale without GPS or telemetry.
AgentSelect defines narrative-query to end-to-end agent configuration recommendation, proposing a benchmark with queries, agents, and interactions.
A curated link roundup from recently collected official updates and tech news.
CoT perturbations can sharply reduce accuracy. Unit conversion remains hard at scale; isolate checks and use self-consistency.
Retiring legacy ChatGPT models may shift tone, refusals, and creativity, reshaping the balance between expression and safety guardrails.
Examines multi-rater 3D lesion segmentation, limits of vanilla diffusion, and VDD anchored to consensus priors improving GED/CI.
How LLMs create difficulty illusions, and how to design evaluation gates with scenarios, protocols, and multi-metric reporting.
GIPO targets scarce, stale interaction data by replacing hard importance-ratio clipping with log-ratio Gaussian trust weights for stable reuse.
Reframes agentic AI failures as governance issues, proposing dual-helix governance with a Knowledge/Behavior/Skills architecture.
How LLM signals can shape belief in partially observable TAMP, and why calibration, uncertainty, and safety filters matter for reliability.
How to use LLM agents for research formalization with guardrails: log everything, run continuous evaluation, and score tool selection and argument precision.
How ambiguity detection, clarification, and sycophancy control shape managerial AI advice quality, risk, and evaluation metrics.
MASS trains LLMs to synthesize per-problem data and self-update at test time, raising auditability, integrity, and reproducibility needs.
Optimize AI subscriptions by checking usage limits, terms restrictions, and uptime transparency to minimize workflow disruption risk.
LLM-based conversational recommenders may infer sensitive triggers from dialogue, risking personalized safety violations unless constraints are enforced.
PlugMem externalizes long-term memory as a plug-in to reduce retrieval bloat and relevance loss, while highlighting persistent injection risks.
Tool-free visual puzzle claims depend on fixed constraints: lock tools, image preprocessing, prompts, and logs for reproducibility.
NVML, DCGM, and nvidia-smi report window-averaged power and utilization. Learn how sampling affects LLM inference graphs.
AI “effort replacement” spans cognitive automation to body/brain augmentation. Check RCT evidence, effect sizes, and regulatory safety.
As AI displaces jobs, energy costs and value capture can constrain cash transfers like UBI, complicating inflation and fiscal assumptions.
As AI enters battlefield planning, HITL, TEVV validation, auditability, and accountability design matter more than raw performance.
Why AI performance gains don’t instantly raise productivity, and how to close the lag using task scores and NIST AI RMF.