Documentation

Linglib.Phenomena.Processing.Studies.GiulianelliEtAl2026

Giulianelli, Wallbridge, Cotterell & Fernández (2026) #

@cite{giulianelli-etal-2026}

Incremental alternative sampling as a lens into the temporal and representational resolution of linguistic prediction. Journal of Memory and Language, 148, 104715.

Key contributions #

  1. Defines incremental information value — the expected representational distance between alternatives sampled before vs. after observing a word
  2. Shows that different psycholinguistic measures are best predicted by different (forecast horizon, representational level) configurations
  3. The full IAS model (all h × all l) outperforms standard surprisal for most measures
  4. Surprisal implicitly integrates across multiple temporal and representational resolutions — its predictability is closest to a best-case (d^{min}) notion rather than average discrepancy (d^{mean})

Datasets #

Language models #

GPT-2 in four sizes (Small 12L / Medium 24L / Large 36L / XL 48L), forecast horizons h ∈ 1..10, k = 50 sampled alternatives per timestep.

Connection to ProcessingModel #

Editorial note: The following bridge is our interpretation connecting IAS's findings to linglib's ProcessingModel.ProcessingProfile. The paper itself does not reference ProcessingProfile.

IAS's layer-wise decomposition reveals that different representational levels predict different psycholinguistic measures. This parallels ProcessingProfile's separation of locality, boundaries, referentialLoad, and ease into orthogonal cognitive dimensions — both architectures recognize that processing cost is irreducibly multi-dimensional.

source
inductive Phenomena.Processing.Studies.GiulianelliEtAl2026.Dataset :
Type

The two psycholinguistic datasets used in this study.

  • aligned : Dataset

    205 English sentences, sentence-level stimuli, multiple response types

  • naturalStories : Dataset

    10 English narratives (~1000 words each), multi-sentence, self-paced RT

Instances For
    source
    instance Phenomena.Processing.Studies.GiulianelliEtAl2026.instDecidableEqDataset :
    DecidableEq Dataset
    Equations
    source
    def Phenomena.Processing.Studies.GiulianelliEtAl2026.instBEqDataset.beq :
    Dataset → Dataset → Bool
    Equations
    Instances For
      source
      instance Phenomena.Processing.Studies.GiulianelliEtAl2026.instBEqDataset :
      BEq Dataset
      Equations
      source
      def Phenomena.Processing.Studies.GiulianelliEtAl2026.instReprDataset.repr :
      Dataset → Nat → Std.Format
      Equations
      • One or more equations did not get rendered due to their size.
      Instances For
        source
        instance Phenomena.Processing.Studies.GiulianelliEtAl2026.instReprDataset :
        Repr Dataset
        Equations
        source
        def Phenomena.Processing.Studies.GiulianelliEtAl2026.peakHorizon :
        Core.GeneralisedSurprisal.PsychMeasure → Nat

        For each psycholinguistic measure, the forecast horizon at which incremental information value is most predictive (highest Δ_{Adj.R²}) in the sentence-level Aligned dataset. N400 and P600 peak at h = 2 (two-word lookahead); all other measures peak at h = 1 (next word).

        Equations
        Instances For
          source
          def Phenomena.Processing.Studies.GiulianelliEtAl2026.peakLevel :
          Core.GeneralisedSurprisal.PsychMeasure → Core.GeneralisedSurprisal.RepLevel

          For each psycholinguistic measure, the representational level at which incremental information value is most predictive (primary peak).

          • Explicit predictability (cloze, ratings): embedding layer (lexical identity)
          • N400: embedding layer — shallow lexical-semantic prediction error
          • Eye-tracked RT: early-to-intermediate layers — structural processing
          • Self-paced RT and P600: intermediate layers — syntactic integration

          Note: most measures show bimodal layer patterns; see layerPattern for the full picture.

          Equations
          Instances For
            source
            inductive Phenomena.Processing.Studies.GiulianelliEtAl2026.LayerPattern :
            Type

            Layer-activation pattern across representational depth.

            The paper's central finding is that predictive power as a function of layer depth has distinct shapes for different measure classes.

            • unimodal : LayerPattern

              Single peak at one depth region (N400: embedding; P600: intermediate)

            • uShaped : LayerPattern

              Peaks at both boundary layers (embedding + final); characteristic of explicit predictability measures (cloze, ratings)

            • sShaped : LayerPattern

              Peaks at early-to-intermediate layers with secondary rise at final layer; characteristic of reading time measures

            Instances For
              source
              instance Phenomena.Processing.Studies.GiulianelliEtAl2026.instDecidableEqLayerPattern :
              DecidableEq LayerPattern
              Equations
              source
              def Phenomena.Processing.Studies.GiulianelliEtAl2026.instBEqLayerPattern.beq :
              LayerPattern → LayerPattern → Bool
              Equations
              Instances For
                source
                instance Phenomena.Processing.Studies.GiulianelliEtAl2026.instBEqLayerPattern :
                BEq LayerPattern
                Equations
                source
                instance Phenomena.Processing.Studies.GiulianelliEtAl2026.instReprLayerPattern :
                Repr LayerPattern
                Equations
                source
                def Phenomena.Processing.Studies.GiulianelliEtAl2026.instReprLayerPattern.repr :
                LayerPattern → Nat → Std.Format
                Equations
                • One or more equations did not get rendered due to their size.
                Instances For
                  source
                  def Phenomena.Processing.Studies.GiulianelliEtAl2026.layerPattern :
                  Core.GeneralisedSurprisal.PsychMeasure → LayerPattern

                  Layer-activation pattern for each psycholinguistic measure.

                  • Explicit measures (cloze, ratings): U-shaped — peak at embedding and final layer, the two layers closest to lexical identity
                  • Reading times: S-shaped — peak at early-to-intermediate layers with secondary rise at the final layer
                  • N400: unimodal at embedding layer
                  • P600: unimodal at intermediate (syntactic) layers
                  Equations
                  Instances For
                    source
                    def Phenomena.Processing.Studies.GiulianelliEtAl2026.naturalStoriesSPRT_peakHorizon :
                    Nat

                    In Natural Stories (multi-sentence stimuli), self-paced reading time predictive power increases with forecast horizon up to h = 7, unlike sentence-level stimuli where h = 1 is optimal. The extended context and dashed moving-window display promote longer-range predictions.

                    Equations
                    Instances For
                      source
                      def Phenomena.Processing.Studies.GiulianelliEtAl2026.iasOutperformsSurprisal :
                      Core.GeneralisedSurprisal.PsychMeasure → Dataset → Bool

                      The full IAS model (all h × all l simultaneously) outperforms standard surprisal in predictive power for most psycholinguistic measures.

                      Exceptions:

                      • Predictability ratings: surprisal slightly better (not significant)
                      • Self-paced RT in Aligned: difference not significant
                      Equations
                      Instances For
                        source
                        def Phenomena.Processing.Studies.GiulianelliEtAl2026.observedSign :
                        Core.GeneralisedSurprisal.PsychMeasure → Int

                        Observed sign of the relationship between IAS information value and each psycholinguistic measure, as actually found in the data.

                        Differs from PsychMeasure.expectedSign for P600: the paper predicted a positive relationship but found a negative one ("for P600, however, our directional hypothesis is refuted").

                        Equations
                        Instances For
                          source
                          def Phenomena.Processing.Studies.GiulianelliEtAl2026.highComplementarity :
                          Core.GeneralisedSurprisal.PsychMeasure → Dataset → Bool

                          Whether IAS and surprisal show high complementarity (joint model substantially exceeds either alone) for a given measure and dataset.

                          Complementarity is highest for predictability measures and self-paced RT in multi-sentence stimuli. For other measures, the joint model ≈ IAS alone, meaning IAS subsumes surprisal's predictive contribution.

                          Equations
                          Instances For
                            source
                            def Phenomena.Processing.Studies.GiulianelliEtAl2026.surprisalBestMatchesSummary :
                            Core.GeneralisedSurprisal.DistanceSummary

                            Under which distance summary statistic surprisal's implicit predictions are best captured.

                            Under d^{mean}, surprisal correlates most with the final layer at h = 1 (lexical expectations for the next word). Under d^{min}, surprisal correlates most with intermediate layers at h = 3–5 (closest-hypothesis tracking), with correlation coefficients up to 0.81.

                            The d^{min} finding reveals that surprisal's predictability is closest to a best-case (closest-hypothesis) notion rather than average discrepancy, which may explain its strong predictive power despite lacking explicit temporal-representational resolution.

                            Equations
                            Instances For
                              source
                              theorem Phenomena.Processing.Studies.GiulianelliEtAl2026.explicit_measures_peak_at_h1 (m : Core.GeneralisedSurprisal.PsychMeasure) (h : m.isExplicit = true) :
                              peakHorizon m = 1

                              Explicit predictability measures all peak at horizon 1 (next word).

                              source
                              theorem Phenomena.Processing.Studies.GiulianelliEtAl2026.erp_peaks_at_h2 :
                              peakHorizon Core.GeneralisedSurprisal.PsychMeasure.n400 = 2 ∧ peakHorizon Core.GeneralisedSurprisal.PsychMeasure.p600 = 2

                              N400 and P600 both peak at horizon 2, not horizon 1 — prediction of ERP components benefits from two-word lookahead.

                              source
                              theorem Phenomena.Processing.Studies.GiulianelliEtAl2026.n400_p600_level_dissociation :
                              peakLevel Core.GeneralisedSurprisal.PsychMeasure.n400 = Core.GeneralisedSurprisal.RepLevel.lexical ∧ peakLevel Core.GeneralisedSurprisal.PsychMeasure.p600 = Core.GeneralisedSurprisal.RepLevel.syntactic

                              N400 and P600 are predicted by different representational levels: N400 by embedding (lexical-semantic), P600 by intermediate (syntactic). This dissociation mirrors the established functional distinction between these components.

                              source
                              theorem Phenomena.Processing.Studies.GiulianelliEtAl2026.discourse_extends_horizon :
                              naturalStoriesSPRT_peakHorizon > peakHorizon Core.GeneralisedSurprisal.PsychMeasure.selfPacedRT

                              Self-paced RT in multi-sentence stimuli benefits from substantially longer forecast horizons than sentence-level stimuli.

                              source
                              theorem Phenomena.Processing.Studies.GiulianelliEtAl2026.ias_outperforms_for_erps :
                              iasOutperformsSurprisal Core.GeneralisedSurprisal.PsychMeasure.n400 Dataset.aligned = true ∧ iasOutperformsSurprisal Core.GeneralisedSurprisal.PsychMeasure.p600 Dataset.aligned = true

                              IAS outperforms surprisal for both ERP components (Aligned dataset).

                              source
                              theorem Phenomena.Processing.Studies.GiulianelliEtAl2026.ias_outperforms_for_eyetracking :
                              iasOutperformsSurprisal Core.GeneralisedSurprisal.PsychMeasure.firstFixationRT Dataset.aligned = true ∧ iasOutperformsSurprisal Core.GeneralisedSurprisal.PsychMeasure.firstPassRT Dataset.aligned = true ∧ iasOutperformsSurprisal Core.GeneralisedSurprisal.PsychMeasure.rightBoundedRT Dataset.aligned = true ∧ iasOutperformsSurprisal Core.GeneralisedSurprisal.PsychMeasure.goPastRT Dataset.aligned = true

                              IAS outperforms surprisal for all eye-tracking measures (Aligned).

                              source
                              theorem Phenomena.Processing.Studies.GiulianelliEtAl2026.sprt_ias_advantage_dataset_dependent :
                              iasOutperformsSurprisal Core.GeneralisedSurprisal.PsychMeasure.selfPacedRT Dataset.naturalStories = true ∧ ¬iasOutperformsSurprisal Core.GeneralisedSurprisal.PsychMeasure.selfPacedRT Dataset.aligned = true

                              IAS outperforms surprisal for self-paced RT in Natural Stories but not in the sentence-level Aligned dataset.

                              source
                              theorem Phenomena.Processing.Studies.GiulianelliEtAl2026.explicit_implicit_level_dissociation :
                              peakLevel Core.GeneralisedSurprisal.PsychMeasure.clozeProbability ≠ peakLevel Core.GeneralisedSurprisal.PsychMeasure.firstPassRT

                              Explicit measures and implicit measures are predicted by different representational levels: explicit at lexical, implicit at syntactic or shallow-syntactic.

                              source
                              theorem Phenomena.Processing.Studies.GiulianelliEtAl2026.explicit_rt_pattern_dissociation :
                              layerPattern Core.GeneralisedSurprisal.PsychMeasure.clozeProbability = LayerPattern.uShaped ∧ layerPattern Core.GeneralisedSurprisal.PsychMeasure.firstPassRT = LayerPattern.sShaped

                              Explicit measures show U-shaped layer patterns while reading times show S-shaped patterns — different measure classes engage different representational geometries.

                              source
                              theorem Phenomena.Processing.Studies.GiulianelliEtAl2026.p600_sign_reversal :
                              Core.GeneralisedSurprisal.PsychMeasure.p600.expectedSign = 1 ∧ observedSign Core.GeneralisedSurprisal.PsychMeasure.p600 = -1

                              P600's observed sign is negative, contradicting the predicted positive relationship. This is the only measure where the directional hypothesis is refuted.

                              source
                              theorem Phenomena.Processing.Studies.GiulianelliEtAl2026.other_signs_confirmed :
                              observedSign Core.GeneralisedSurprisal.PsychMeasure.n400 = Core.GeneralisedSurprisal.PsychMeasure.n400.expectedSign ∧ observedSign Core.GeneralisedSurprisal.PsychMeasure.selfPacedRT = Core.GeneralisedSurprisal.PsychMeasure.selfPacedRT.expectedSign ∧ observedSign Core.GeneralisedSurprisal.PsychMeasure.firstPassRT = Core.GeneralisedSurprisal.PsychMeasure.firstPassRT.expectedSign

                              All other measures confirm their predicted signs.

                              source
                              theorem Phenomena.Processing.Studies.GiulianelliEtAl2026.complementarity_dissociation :
                              highComplementarity Core.GeneralisedSurprisal.PsychMeasure.clozeProbability Dataset.aligned = true ∧ ¬highComplementarity Core.GeneralisedSurprisal.PsychMeasure.firstPassRT Dataset.aligned = true

                              For most measures, IAS subsumes surprisal (low complementarity), but for explicit predictability measures, they complement each other.

                              source
                              theorem Phenomena.Processing.Studies.GiulianelliEtAl2026.surprisal_is_best_case :
                              surprisalBestMatchesSummary = Core.GeneralisedSurprisal.DistanceSummary.min

                              Surprisal's implicit predictions are best characterized by the minimum distance statistic, not mean distance.