We have 3 of 7 papers
We haven’t gathered this author’s papers yet. Follow them and we’ll fetch their work.
Not the right person? Other researchers publish under this name.
This work proposes ECHO (Evaluation-Conditioned Harmfulness Optimization), which replaces the compliance proxy with response-level, continuous harmfulness scores from HarmSig, a 1.7B regressor grounded in actual harmful capability.
The key idea is to induce the adversary to solve a misspecified inverse problem, in which no plausible label sequence in the sequence space can explain the observed gradients, by inducing inconsistency across three dimensions: objective, direction, and scale.
DPA is grounded in a critical insight: regardless of fine-tuning tactics to evade provenance, the practical necessity of maintaining utility constrains the model to preserve the fundamental intersection of semantic substance and lexical form, and captures this persistent lexical-semantic intersection as intrinsic distr...
We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.
