analiza aiqbench
This research introduces a forward-pass diagnostic metric called the "refusal-affirmation logit gap" to quantify the operational safety margins in RLHF-aligned models. By identifying short, low-perplexity suffixes that efficiently close this gap, the study reveals that current alignment defenses are often thin and vulnerable to gradient-free attacks requiring minimal computational overhead. These findings highlight a critical vulnerability in model robustness where standard perplexity filters fail to detect adversarial inputs designed to bypass safety refusals.
Logit-Gap Steering: A Forward-Pass Diagnostic for Alignment Robustness
arXiv:2506.24056v3 Announce Type: replace-cross Abstract: RLHF-style alignment trains language models to refuse unsafe requests, but how much operational margin does this refusal rest on? We introduce the refusal-affirmation logit gap: the difference between the top refusal-token logit and the top affirmative-token logit at the first decoding step. This single scalar quantifies the per-prompt safety margin that alignment provides. Empirically, alignment widens the gap on 97.5-99.8% of toxic prompts across three model families, and median gap closure co-varies with True-ASR ranking across suffix strategies (an internal consistency check, since our method optimises gap closure). To validate the metric's practical significance, we present logit-gap steering, a gradient-free, forward-pass-only method that discovers short in-distribution suffixes ($<$10 tokens per component) whose cumulative effect closes the gap. The method requires ${\approx}26{,}000$ forward-pass equivalents per family (${\approx}2$~min on one A100), ${\approx}125\times$ less than a single GCG search. Suffixes discovered on 0.5B--2B models transfer without modification to 72B within family. An 8-suffix ensemble reaches 38-96\% True ASR across 13 models on AdvBench and HarmBench, with most suffixes having $10^{3}$-$10^{4}\times$ lower perplexity than GCG-meaning published perplexity-filter defenses that collapse GCG (64.7%$\to$1.0%) leave our suffixes nearly intact (76.9%$\to$76.0%). These results demonstrate that current alignment margins, while consistently present, can be thin and efficiently measurable, and that defense strategies must account for in-distribution suffixes.
Abstract:RLHF-style alignment trains language models to refuse unsafe requests, but how much operational margin does this refusal rest on? We introduce the refusal-affirmation logit gap: the difference between the top refusal-token logit and the top affirmative-token logit at the first decoding step. This single scalar quantifies the per-prompt safety margin that alignment provides. Empirically, alignment widens the gap on 97.5-99.8% of toxic prompts across three model families, and median gap closure co-varies with True-ASR ranking across suffix strategies (an internal consistency check, since our method optimises gap closure). To validate the metric's practical significance, we present logit-gap steering, a gradient-free, forward-pass-only method that discovers short in-distribution suffixes ($<$10 tokens per component) whose cumulative effect closes the gap. The method requires ${\approx}26{,}000$ forward-pass equivalents per family (${\approx}2$~min on one A100), ${\approx}125\times$ less than a single GCG search. Suffixes discovered on 0.5B--2B models transfer without modification to 72B within family. An 8-suffix ensemble reaches 38-96\% True ASR across 13 models on AdvBench and HarmBench, with most suffixes having $10^{3}$-$10^{4}\times$ lower perplexity than GCG-meaning published perplexity-filter defenses that collapse GCG (64.7%$\to$1.0%) leave our suffixes nearly intact (76.9%$\to$76.0%). These results demonstrate that current alignment margins, while consistently present, can be thin and efficiently measurable, and that defense strategies must account for in-distribution suffixes.
Accepted at NeurIPS 2026 Main Track (poster). Camera-ready version
Cryptography and Security (cs.CR); Computation and Language (cs.CL); Machine Learning (cs.LG)
arXiv:2506.24056 [cs.CR]
(or arXiv:2506.24056v3 [cs.CR] for this version)
https://doi.org/10.48550/arXiv.2506.24056 Focus to learn more arXiv-issued DOI via DataCite
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.
Text integral sub licența deschisă a publicației · 318 cuvinte
Subiecte conexe
[AINews] Claude Opus 5.5, the new default model for AINews — and everybody cuts prices 40-50%
Latent Space · 23 sept. 2026
Debugging Logprob Mismatches in LLM Reinforcement Learning
AMD · 23 sept. 2026
J++ Lens: Jacobian Filtering Enables More Faithful Workspace Lenses
LessWrong · 8 oct. 2026
We're talking past our models; or, How a model defined its "evil" vector as dread
LessWrong · 20 iul. 2026