Text detoxification is a balancing act: removing toxic language while keeping the original meaning and ensuring the output reads fluently. According to a new paper on arXiv, existing methods typically optimize an unconstrained scalar objective, which can let safety scores hide semantic degradation. The paper, titled AMEND, argues that this approach is insufficient for maintaining quality across all three dimensions.
AMEND is described as a constraint-guided multi-agent text detoxification framework with isolated critics. The abstract suggests that instead of relying on a single scalar reward, the method uses constraints and multiple agents—each with a specific critic role—to explicitly enforce the desired properties. This design aims to prevent the model from sacrificing semantic preservation for the sake of lower toxicity scores.
The significance of this work lies in its attempt to reframe detoxification as a constrained optimization problem rather than an unconstrained one. However, the abstract is truncated, so details about the exact architecture, training procedure, or experimental results are not available in the source. The paper is available on arXiv under the identifier 2610.23183v1.