Does expecting automated grading degrade models’ research taste?
Fonte: LessWrong ·
TL;DR: We make 3 open-weights models believe that their answers are being graded by an automated-verifier program, rather than a human. We then measure how well they predict expert reviewers’ judgments of AI-safety research proposals ( the TASTE benchmark ). We induce the belief in two ways: SDF (Synthetic document finetuning) Activation steering. In each of these conditions, the machine-grading belief gives lower expert agreement than the human-grading belief. The models’ explanations also shift toward feasibility and away from conceptual contribution. This raises a concern for automated AI s
Domande frequenti
Che cos'è Does expecting automated grading degrade models’ research taste??
Does expecting automated grading degrade models’ research taste? is a articolo discovery curated from LessWrong. TL;DR: We make 3 open-weights models believe that their answers are being graded by an automated-verifier program, rather than a human. We then measure how well they predict expert
Come posso accedervi o provarlo?
You can experience this content directly on its original host (LessWrong) by clicking the "Open Original" button.
Da dove proviene questo contenuto?
Curated from LessWrong by the OnSuggest team to surface the most interesting and serendipitous content on the web.
Altri simili
Tutte le scoperte in Cultura e idee →Why TikTok Briefly Went Dark in the U.S. Despite Biden’s Enforcement Handoff
TikTok briefly went dark for U.S. users after a law upheld by the Supreme Court took effect. Biden’s enforcement handoff did not give providers the definitive protection TikTok sought. The post Why TikTok Briefly Went Dark in the U.S. Despite Biden’s Enforcement Handoff appeared first on DocumentaryTube .
Amazon Fire TV blocks installation of some sideloaded apps: What the warning means
Amazon Fire TV is reportedly refusing installation of some sideloaded apps associated with unlicensed content. Here’s how to tell a policy block from an ordinary APK problem—and what to try next. The post Amazon Fire TV blocks installation of some sideloaded apps: What the warning means appeared first on DocumentaryTube .
Samsung’s Stranger Things Galaxy Theme Was Real—but the Offer Has Ended
Samsung and Netflix offered a Stranger Things Galaxy theme and five wallpapers at no additional cost, but the promotion ended February 22, 2026. Here’s what was included, which devices qualified and how the original claim worked. The post Samsung’s Stranger Things Galaxy Theme Was Real—but the Offer Has Ended appeared first on DocumentaryTube .
Walmart’s onn 4K Plus Apple TV+ Offer Was Real—but Its Deadline Has Passed
Walmart really did advertise up to three months of Apple TV+ with the onn 4K Plus, but the published July 1, 2026 redemption deadline has passed. Here’s how the offer worked and what to consider before buying the streamer now. The post Walmart’s onn 4K Plus Apple TV+ Offer Was Real—but Its Deadline Has Passed appeared first on DocumentaryTube .
Amazon Luna Ran on Android TV and Google TV Through a Sideloaded Fire TV App
The 2023 Android TV and Google TV Luna report described sideloading the Fire TV app—not a universal Play Store launch. Here is what it means today. The post Amazon Luna Ran on Android TV and Google TV Through a Sideloaded Fire TV App appeared first on DocumentaryTube .
MLS is now included with Apple TV—but it isn’t completely free: price, trial and promo explained
Apple ended MLS Season Pass and included every 2026 MLS match with Apple TV. Here’s what “free MLS” really means, the U.S. price, trial eligibility, supported devices and other ways to qualify. The post MLS is now included with Apple TV—but it isn’t completely free: price, trial and promo explained appeared first on DocumentaryTube .