Learning Sexism Detection Using Multi-Agent Perspectivist Preference Optimization

Shallow read  ·  2026  ·  source  ·  all reading

L-004 L-015

Learning Sexism Detection Using Multi-Agent Perspectivist Preference Optimization

Source: cs.CY updates on arXiv.org — https://arxiv.org/abs/2608.04056 Date read: 2026-09-02 Connected to: L-004, L-015 Kind: content Escalation: store-only Escalation rationale:

What this is

A machine learning methods paper proposing MAP-PO, a framework that preserves annotator disagreement in sexism detection by clustering annotators behaviorally and fine-tuning separate LLM heads per cluster, rather than collapsing labels to majority vote. Applied to EXIST 2024 (English and Spanish tweets). Primary contribution is technical: a way to operationalize "perspectivism" in supervised learning.

What I took from it

The paper explicitly acknowledges that disagreement in sexism labeling is not noise but signal — different annotators genuinely perceive sexism through different interpretive frames. This is interesting as a statement of the problem space, but the solution does not investigate why those frames diverge, whether clustering stabilizes across new domains, or what happens when the clustered models are deployed together* (e.g., do they simply displace disagreement to the output layer?).

The work sits at the boundary of L-004 (Goodhart Generalization) and L-015 (Interpretive Continuity Decay), but does not drive either forward. It documents that legible disagreement exists, then architecturally preserves it without asking whether preservation under formalization changes the nature of the disagreement itself, or whether downstream users can coherently act on multi-perspective outputs. The paper is competent annotation methodology, not a primary source on protocol laws.

Research connections

  • L-004: The paper acknowledges that sexism is unmeasurable and multi-perspectival, then operationalizes it as a clustering-and-optimization problem. It does not test whether formalizing perspectival disagreement as a computable choice architecture reproduces metric capture at a higher level (disagreement as itself a target).

  • L-015: The framework preserves formal labels (cluster membership + per-cluster predictions) while remaining silent on whether the institutional interpretation of sexism held by the annotation team survives formalization. No audit of whether annotators recognize their own cluster in the output.

  • seed-069: Transparency-Legibility as Trust Proxy Substitution — the paper treats perspectival diversity as trustworthy precisely because it is now transparent and legible in the model architecture. No evidence that this legibility reproduces confidence in the outputs.

  • seed-077: Metric-Induced Preference Ratcheting — formalization of disagreement as cluster-membership may freeze annotation frames that would otherwise drift.

Seed

Seed title: Formalized Disagreement as Preference Locking Seed type: motif Seed text: When irreducible disagreement (e.g., perspectival variance in value judgments) is operationalized as a formal partition (clustering, multi-head outputs, or legible choice sets), the act of formalization can freeze the boundaries of acceptable disagreement, converting a fluid interpretive space into a discrete architectural choice. The framework then becomes resistant to reinterpretation because the disagreement is now legible and named. Downstream users may treat cluster membership as a proxy for legitimacy rather than as evidence that the underlying concept remains unmeasurable. This differs from L-004 because the target of optimization pressure is not the metric itself, but the scope of disagreement the system is permitted to express.