Documentation

Linglib.Phenomena.Clarification.Studies.DongEtAl2026

Dong, Hu, Hui, Zhang, Vulić, Bobu & Collier (2026) #

@cite{dong-etal-2026}

Value of Information: A Framework for Human–Agent Communication.

Overview #

LLM agents face a clarify-or-commit dilemma: ask the user for clarification (incurring cognitive cost) or act on incomplete information (risking error). This paper proposes a Value of Information (VoI) framework: agents ask question q only when VoI(q) exceeds communication cost c.

RSA framing #

The paper explicitly adopts an RSA perspective (@cite{frank-goodman-2012}; @cite{goodman-frank-2016}), viewing dialogue as rational action. VoI(q) (eq. 4: V_post(b,q) − V(b)) is structurally the same as questionUtility (@cite{van-rooy-2003}): the expected gain in decision value from asking q.

Connection to @cite{hawkins-etal-2025} #

VoI captures WHETHER to ask (the questioner's decision), while the commit action (argmax over expected utility) captures WHAT to do. The VoI framework is complementary to @cite{hawkins-etal-2025}'s respondent model.

Risk-Sensitivity and Action Utility #

The key qualitative finding (Appendix A, Figure 4): in the Mixed 20Q task, the VoI agent asks more questions for medical diagnosis (U = 10) than animal guessing (U = 1), because higher stakes increase expected regret.

Action-utility scoring (@cite{hawkins-etal-2025}'s β > 0) encodes stakes into s1Score: exp(α · U(target, guess)) scales with reward k. In richer models (multiple targets with partial matches), this creates δ-sensitive S1 preferences — the mechanism behind the @cite{tsvilodub-etal-2026} formalization's cross-config comparisons.

In this binary identification task, however, the action-utility effect is degenerate: the L0 gate zeros the wrong guess, so S1 assigns probability 1 to the correct guess regardless of k. The task is too simple for action utility to produce qualitative differences — both k = 1 and k = 10 yield identical S1/L1 predictions after normalization.

source
inductive Phenomena.Clarification.Studies.DongEtAl2026.Target :
Type

Binary identification task (simplified from the paper's Mixed-Stakes 20 Questions, which uses 100 animals / 15 diseases).

Instances For
    source
    instance Phenomena.Clarification.Studies.DongEtAl2026.instDecidableEqTarget :
    DecidableEq Target
    Equations
    source
    def Phenomena.Clarification.Studies.DongEtAl2026.instReprTarget.repr :
    Target → ℕ → Std.Format
    Equations
    • One or more equations did not get rendered due to their size.
    Instances For
      source
      instance Phenomena.Clarification.Studies.DongEtAl2026.instReprTarget :
      Repr Target
      Equations
      source
      instance Phenomena.Clarification.Studies.DongEtAl2026.instFintypeTarget :
      Fintype Target
      Equations
      • One or more equations did not get rendered due to their size.
      source
      def Phenomena.Clarification.Studies.DongEtAl2026.targetMatches :
      Target → Target → Bool

      Boolean match: does guess match target?

      Equations
      Instances For
        source
        noncomputable def Phenomena.Clarification.Studies.DongEtAl2026.identCfg (k : ℝ) (hk : 0 ≤ k) :
        RSA.RSAConfig Target Target

        Binary identification as RSA reference game with action-utility scoring.

        s1Score(L0, α, target, guess) = if L0(target|guess) = 0 then 0 else exp(α · k)

        At k = 1 and k = 10, S1(correct|target) = 1 after normalization — the binary task is degenerate. Action-utility scoring IS the right mechanism (@cite{hawkins-etal-2025}'s β = 1), but the task is too simple for it to produce qualitative differences.

        Equations
        • One or more equations did not get rendered due to their size.
        Instances For
          source
          noncomputable def Phenomena.Clarification.Studies.DongEtAl2026.animalCfg :
          RSA.RSAConfig Target Target

          Animal guessing (k = 1).

          Equations
          Instances For
            source
            noncomputable def Phenomena.Clarification.Studies.DongEtAl2026.medicalCfg :
            RSA.RSAConfig Target Target

            Medical diagnosis (k = 10).

            Equations
            Instances For

              Both configs produce identical S1/L1 predictions — the binary task is degenerate. The L0 gate zeros wrong guesses, so S1 assigns probability 1 to the correct guess at any k.

              source
              theorem Phenomena.Clarification.Studies.DongEtAl2026.S1_animal_prefers_correct :
              animalCfg.S1 () Target.t₁ Target.t₁ > animalCfg.S1 () Target.t₁ Target.t₂

              S1 prefers correct guess (animal guessing, k = 1).

              source
              theorem Phenomena.Clarification.Studies.DongEtAl2026.S1_medical_prefers_correct :
              medicalCfg.S1 () Target.t₁ Target.t₁ > medicalCfg.S1 () Target.t₁ Target.t₂

              S1 prefers correct guess (medical diagnosis, k = 10). Same qualitative prediction despite 10× stakes.

              source
              theorem Phenomena.Clarification.Studies.DongEtAl2026.L1_animal_identifies :
              animalCfg.L1 Target.t₁ Target.t₁ > animalCfg.L1 Target.t₁ Target.t₂

              L1 correctly infers target from guess (animal).

              source
              theorem Phenomena.Clarification.Studies.DongEtAl2026.L1_medical_identifies :
              medicalCfg.L1 Target.t₁ Target.t₁ > medicalCfg.L1 Target.t₁ Target.t₂

              L1 correctly infers target from guess (medical). Same qualitative prediction despite 10× stakes.