Drew's Comic NewsroomAbout

Comics on the day’s news

Claude spent the weekend fixing other models’ bad habitsCoyote vs. Acme opens at 95% on the TomatometerAn ETF is a basket that trades like a stockSome quantum-gravity clues may be two maps of the same hillSara Bareilles turns a Lincoln Center lobby into a grief archiveBáez saved it, McGonigle ended it, Tigers 2-1A US Open guard’s wife weaponizes Tsitsipas quotesThe book nobody can read is still unreadClaude spent the weekend fixing other models’ bad habitsCoyote vs. Acme opens at 95% on the TomatometerAn ETF is a basket that trades like a stockSome quantum-gravity clues may be two maps of the same hillSara Bareilles turns a Lincoln Center lobby into a grief archiveBáez saved it, McGonigle ended it, Tigers 2-1A US Open guard’s wife weaponizes Tsitsipas quotesThe book nobody can read is still unread
Four-panel full-color AI research comic. Panel 1: a small cartoon language model rubber-stamping YES on a stack of 1200 logic cards. Panel 2: a colorful probe meter reading the correct verdict at 0.96 AUC from hidden states. Panel 3: a giant decision dial offset by plus 4.6 sigma with margin AUC 0.89. Panel 4: three labeled bins reading concealed, miscalibrated, and undetected beside repair dials lifting accuracy from 50 to 81 percent at 0.6B and 94 percent at 8B.
Four panels: a tiny LM stamping YES on every card, a glowing probe reading the true verdict, a decision threshold shoved +4.6σ off-center, and three regime bins: concealed, miscalibrated, undetected. · Comic: Topics / Drew’s Comic Newsroom. Source: arXiv.

Technology and AI

Internal probes beat the answer: a 0.6B model says YES to everything

What happened

arXiv paper 2609.04582, submitted 4 September 2026 by Gnaneswar Villuri, Hashmath Shaik, and Alex Doboli, studies a familiar failure mode: a 0.6B language model asked to verify 1,200 logical conclusions answers YES every time. Judged by behavior it discriminates nothing. Linear probes on its hidden states still read the correct verdict at 0.96 AUC, transferring to unseen logical structures.

The authors locate the loss at a single scalar. The verdict survives into the model’s own output logits along a well-aligned readout direction — margin AUC about 0.89 — but a saturated decision threshold offset by roughly +4.6 sigma erases it. Across many configurations, behavioral accuracy collapses as a function of that threshold offset. A one-parameter correction, never fit on the evaluated structures, repairs behavior from 50% to 81% at 0.6B; calibrated margin decoding recovers about 94% at 8B. Comparing probe to margin separates three regimes: concealed, miscalibrated, and undetected.

Why it matters

AI packages here prefer a measurable method over hype. Recalibrating a threshold is not a jailbreak recipe and not a product endorsement. The comic is the yes-bot, the probe flashlight, the stuck sigma dial, and the one-knob fix. Bright lab colors. No brand logos.

Readers get the numbers and the regimes. Skip anything that wanders into how to bypass safety systems. This is an audit story about knowledge that already exists inside the network.

Conclusion

Probes can beat free-form answers when a miscalibrated threshold conceals knowledge the model already ranks correctly, per arXiv 2609.04582. Source: https://arxiv.org/abs/2609.04582

Source: arXiv