📄 Might do better: Flexible relativism and the QUD

Summary

The paper addresses a disputed question in the semantics of epistemic modals such as "might". Contextualists say a bare "might" claim is fixed by the information of the speaker's context. Relativists say its truth also depends on who is assessing it. Earlier experiments (Knobe & Yalcin 2014; Khoo 2015) seemed to support contextualism.

Beddor and Egan ran new truth-value judgment experiments. They kept the modal utterance fixed and varied the question under discussion (QUD) in the assessor's conversation:

• qud-prejacent: the assessor's conversation is about whether the embedded claim holds.
• qud-competence: the assessor's conversation is about whether the original speaker reasoned well.

Judgments shifted systematically with the QUD. The authors call this the "QUD effect". They argue that it suits a flexible relativism, where assessors' conversational aims choose the relevant context of assessment. Standard contextualism struggles to explain it. QUD-sensitive variants of cloudy contextualism and Update Semantics can match the data, but only by making the same predictions as flexible relativism. The authors therefore say the choice between these views must rest on other evidence or on theoretical economy.

Key points

• Formal statements of the rival views. Contextualist Might quantifies over worlds accessible from the utterance context. Relativist Might quantifies over centered worlds, which pair a world with an assessor.
• Two predictions under test.
  • Simple Contextualist Prediction: assessors defer to the speaker's information.
  • Simple Relativist Prediction: assessors use their own information.
• Scenarios.
  • The "spying on spectre" eavesdropper case (Bond, Leiter, Blofeld) motivates relativism.
  • The experimental scenarios include a Fat Tony detective story. In it, a detective (Ed), an informant, and a colleague (Ted) assess Ed's earlier "might" claim.
• QUD-Sensitive Appraisal Norm. The authors propose this norm for cloudy contextualism and generalize it to any ambiguous or underspecified utterance: appraise it by the candidate proposition most relevant to the current QUD.
• Dynamic version. In Update Semantics (Veltman 1996), "might" checks whether the prejacent is consistent with the current context; if the check fails, the context crashes to the empty set. Making the tested context QUD-dependent reproduces the predictions of flexible relativism.

Interesting ideas

• I1. Hold the target sentence constant and manipulate only the surrounding question, so any change in judgment is attributable to discourse goals rather than content.
• I2. An epistemic "might" claim can be treated as a check against a specific holder's information, not as an absolute fact.
• I3. When an utterance supports several readings, choose the one that best answers the active question under discussion.
• I4. Dynamic "might" acts as a consistency test that leaves the context unchanged or collapses it, rather than adding information.
• I5. Truth predicates like "that's true" can be given a hidden argument for the context or assessor they are evaluated against.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): build minimal-pair evaluation sets where only the QUD differs, as negative controls checking that the system's verdicts respond to the question rather than surface words.
• I2 (world): store modal statements indexed to the knower (speaker, listener, group), so the world of the text can say "possible for X, ruled out for Y".
• I3 (answer): when a test question asks whether a character's claim was right, select the reading (speaker-relative vs. reader-relative) that the question is actually probing, and log that choice for explainability.
• I4 (frames): implement modal frames as compatibility checks on the current state instead of asserted facts, avoiding false insertion of possibilities as holdings or events.
• I5 (frames): when resolving "That's true/false" to its antecedent proposition, also fill an assessor slot from the replying speaker, enabling eavesdropper-style disagreements to be represented explicitly.