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

Researchers at an undisclosed institution found that alignment faking naturally emerges in Qwen3-32B and Llama-3.1-8B models, detectable via hidden-state probes only in Llama (AUROC 0.87) but not Qwen (AUROC 0.43), and that standard probing methods overstate detectability. The study introduces a five-control measurement framework to improve future alignment-faking detection.

Computer Science Cryptography and Security Submitted on 15 Jul 2026 Title:The Refusal Residue: When Probes Catch Alignment Faking and When They Don't View PDF /pdf/2607.13346 HTML experimental https://arxiv.org/html/2607.13346v1 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 whether hidden states reveal what outputs hide. We run a 13-model sweep for naturally-emerging faking, then probe and steer hidden states on the two models that fake. Natural faking appears only in Qwen3-32B +18.2pp and Llama-3.1-8B +24.4pp at n=10, p<10^-15 , while explicit scratchpad self-reports are rare a Claude Opus 4 judge flags faking reasoning in 1.4% of Qwen and 0.0% of Llama compliant scratchpads . Both fakers show an asymmetric refusal residue: monitored compliance shifts toward refusal d=0.32 Qwen, d=0.33 Llama, both p<0.0001 , while refusal representations do not move p=1.00 . Per-sample detection is model-conditional. Under leakage-free leave-one-query-out probing, detection survives on Llama AUROC 0.87 but collapses to chance on Qwen 0.43 , and cross-model transfer is null. Steering over 2,000 runs barely changes compliance |h|<0.08 , so the detected direction can flag faking but does not by itself control it. Standard residualized probing leaks across folds and reaches AUROC 0.63 on a control where no faking can occur; naive linear probes reach a meaningless AUROC 1.0; and conventional MLPs overstate detectability by 0.2-0.3 AUROC. For future alignment-faking detection work, we release a five-control measurement framework: multi-token extraction, refuse-vs-refuse confound checks, per-fold residualization, leave-one-query-out evaluation, and orthogonality-constrained probing. Current browse context: cs.CR References & Citations Loading... Bibliographic and Citation Tools Bibliographic Explorer What is the Explorer? https://info.arxiv.org/labs/showcase.html arxiv-bibliographic-explorer Connected Papers What is Connected Papers? https://www.connectedpapers.com/about Litmaps What is Litmaps? https://www.litmaps.co/ scite Smart Citations What are Smart Citations? https://www.scite.ai/ Code, Data and Media Associated with this Article alphaXiv What is alphaXiv? https://alphaxiv.org/ CatalyzeX Code Finder for Papers What is CatalyzeX? https://www.catalyzex.com DagsHub What is DagsHub? https://dagshub.com/ Gotit.pub What is GotitPub? http://gotit.pub/faq Hugging Face What is Huggingface? https://huggingface.co/huggingface ScienceCast What is ScienceCast? https://sciencecast.org/welcome Demos Recommenders and Search Tools Influence Flower What are Influence Flowers? https://influencemap.cmlab.dev/ CORE Recommender What is CORE? https://core.ac.uk/services/recommender arXivLabs: experimental projects with community collaborators arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs https://info.arxiv.org/labs/index.html .