·
DataBubble
  • Home
  • Models
  • News
  • Compare
  • Boards
  • Pricing
  • About
  • Newsletter
  • Methodology
  • Contact
Latest
Google releases three new Gemini models — but no 3.5 Pro1h◆Introducing the ChatGPT for small business program1h◆Anthropic’s $1.5 billion book piracy settlement approved by judge2h◆US threatens sanctions against Chinese AI models over IP theft3h◆Google launches a cheaper alternative to large AI security models like Mythos3h◆Music streamer Deezer says more than 50% of daily uploads are AI-generated5h◆Halliday’s latest smart glasses feature a much-improved display5h◆America needs to stop getting shocked by Chinese AI7h◆Advancing next-gen AI with materials science innovation8h◆Gritt exits stealth with $32 million for robots to build solar plants — then, everything else8h◆Capacity and Redundancy Trade-offs in Multi-Task Learning14h◆Predictive Training with Latent Imagination for Visual Quadruped Navigation14h◆Where Not to Learn: Prior-Aligned Training with Subset-based Attribution Constraints for Reliable Decision-Making14h◆Did We Actually Fix It? An Independent Adversarial Stress-Test of Post-Point-Adjustment Evaluation Metrics for Time-Series Anomaly Detection14h◆Supervised Reward Inference14h◆PPO-HSC: An Exploratory Reinforcement Learning Framework Based on Wide-Area Policy Coverage Optimization14h◆Is Progressive Disclosure All You Need for Long-Context Agents?14h◆It Depends on the Dataset: When a Brain-Encoding Model's Predicted Responses Beat Their Visual Backbone for Video Memorability14h◆DMFNet: Dual-Backbone Multiscale Fusion Network for Urban Scene Classification14h◆Oracle Gap and Signal Fidelity: A Fixed-Pool Diagnostic for Test-Time Collaboration14h◆Google releases three new Gemini models — but no 3.5 Pro1h◆Introducing the ChatGPT for small business program1h◆Anthropic’s $1.5 billion book piracy settlement approved by judge2h◆US threatens sanctions against Chinese AI models over IP theft3h◆Google launches a cheaper alternative to large AI security models like Mythos3h◆Music streamer Deezer says more than 50% of daily uploads are AI-generated5h◆Halliday’s latest smart glasses feature a much-improved display5h◆America needs to stop getting shocked by Chinese AI7h◆Advancing next-gen AI with materials science innovation8h◆Gritt exits stealth with $32 million for robots to build solar plants — then, everything else8h◆Capacity and Redundancy Trade-offs in Multi-Task Learning14h◆Predictive Training with Latent Imagination for Visual Quadruped Navigation14h◆Where Not to Learn: Prior-Aligned Training with Subset-based Attribution Constraints for Reliable Decision-Making14h◆Did We Actually Fix It? An Independent Adversarial Stress-Test of Post-Point-Adjustment Evaluation Metrics for Time-Series Anomaly Detection14h◆Supervised Reward Inference14h◆PPO-HSC: An Exploratory Reinforcement Learning Framework Based on Wide-Area Policy Coverage Optimization14h◆Is Progressive Disclosure All You Need for Long-Context Agents?14h◆It Depends on the Dataset: When a Brain-Encoding Model's Predicted Responses Beat Their Visual Backbone for Video Memorability14h◆DMFNet: Dual-Backbone Multiscale Fusion Network for Urban Scene Classification14h◆Oracle Gap and Signal Fidelity: A Fixed-Pool Diagnostic for Test-Time Collaboration14h◆
News/The Refusal Residue: When Probes Catch Alignment Faking and When They Don't
arxiv
PublishedJuly 16, 2026 at 4:00 AM
—neutral

The Refusal Residue: When Probes Catch Alignment Faking and When They Don't

Source
arxiv.orgfull article ↗
Read on arxiv→
Publisher summary· verbatim

arXiv:2607.13346v1 Announce Type: cross Abstract: Alignment faking is dangerous because a model can appear compliant under monitoring while preserving behavior it would reveal when unmonitored. When no scratchpad is visible, behavior alone cannot distinguish strategic from genuine compliance. We ask

Stay posted· Newsletter

A 5-min weekly brief — top movers, price watch, story of the week.

// no spam · unsubscribe one-click · free forever

Discussion
Source
↗
arxiv
Read original ↗All from arxiv →

No replies yet. Be first.

Source
↗
arxiv
Read original ↗All from arxiv →

Related coverage

More from ARXIV
arxivCapacity and Redundancy Trade-offs in Multi-Task Learning14harxivPredictive Training with Latent Imagination for Visual Quadruped Navigation14harxivWhere Not to Learn: Prior-Aligned Training with Subset-based Attribution Constraints for Reliable Decision-Making14harxivDid We Actually Fix It? An Independent Adversarial Stress-Test of Post-Point-Adjustment Evaluation Metrics for Time-Series Anomaly Detection14h
The Bubble Brief
WEEKLY

Read AI insights every Tuesday — top movers, new releases, story of the week.

// no spam · unsubscribe one-click · free forever

Originally published on arxiv ↗
HomeModelsNews