Skip to content

Dual Alignment Between Language Model Layers and Human Sentence Processing

Conference: ACL 2026
arXiv: 2604.18563
Code: https://github.com/kuribayashi4/internal_surprisal_targeted_assessment (Available)
Area: Interpretability / Cognition / Psycholinguistics
Keywords: surprisal, Logit-Lens, syntactic ambiguity, reading time, dual alignment

TL;DR

The authors used logit-lens to extract "internal surprisal" from each layer of 19 LMs (GPT-2/Pythia/OPT) and discovered a counter-intuitive "dual alignment": while surprisal from shallow layers aligns best with humans on naturalistic corpora, it is the deep layers that align better on syntactic challenge sentences (e.g., garden-path, NPS, NPZ, RC, Attachment). This corresponds to human dual-mechanism reading models—"shallow by default + switch to deep reanalysis when difficult"—leading to the proposal of using inter-layer surprisal differences (KL/JS) as "inter-layer prediction updates" to serve as supplementary reading-time features.

Background & Motivation

Background: Computational psycholinguistics has long utilized LM surprisal \(S_t = -\log P(w_t \mid w_{<t})\) as a predictor for reading-time (RT), as experiments show RT is approximately linearly correlated with surprisal (Smith & Levy 2013). Recently, Kuribayashi et al. (2025) extended logit-lens to hierarchical levels, finding that surprisal from early layers is more human-like than the final layer on naturalistic corpora, partially addressing the "holistic misalignment" issue.

Limitations of Prior Work: However, the "targeted misalignment" remains unresolved—on syntactic challenge sentences such as garden-path (e.g., MVRR "The girl fed the lamb remained..."), NPS, and NPZ, humans slow down significantly at the disambiguating point, but surprisal from the final layer of all LMs severely underestimates the magnitude of this slowdown. A natural question arises: Are early layers also better on syntactic challenge sentences?

Key Challenge: The authors' experiments provide a negative answer—early layers show almost no difference on syntactic challenge sentences (surprisal for D+ and D− is nearly identical because they only capture local co-occurrence and are insensitive to long-range dependencies). This implies that the "best layer" for alignment is not a global constant but depends on task difficulty.

Goal: (i) Clarify where the "best layer" resides under different syntactic difficulties; (ii) Provide a unified perspective explaining why this dual alignment exists; (iii) Explicitly formulate this "inter-layer difference" into a new reading-time predictor.

Key Insight: The authors analogize the hierarchical forward computation of LMs to the two-stage processing in human reading—shallow layers ≈ default shallow processing (fast, surface-level, local) / deep layers ≈ reanalysis / deep integration (slow, requiring full context). If this metaphor holds, garden-path sentences should necessitate a "shift to deep layers."

Core Idea: Layer-wise surprisal is not a monotonic curve—shallow layers align best for naturalistic text, while deep layers align best for syntactically challenging text. Furthermore, the prediction update from shallow to deep (surprisal update / KL / JS) can itself serve as a proxy for processing cost.

Method

Overall Architecture

The method consists of three steps, all centered around "hierarchical surprisal" in psycholinguistic regressions:

  1. Extract Internal Surprisal: For each LM, token, and layer \(l\), the internal distribution is calculated via logit-lens \(P^{(l)}(W = w_t \mid w_{<t}) = \text{softmax}(W_U \text{LN}(h_{t-1}^{(l)}))_{\text{id}(w_t)}\), yielding \(S^{(l)}_t = -\log P^{(l)}(w_t \mid w_{<t})\). To address potential unreliability of logit-lens in early layers, the authors supplement the analysis with tuned-lens to verify the stability of the conclusions.
  2. Layer-by-layer Slowdown Regression on Syntactic Ambiguity Data: Using five syntactic challenge constructions (MVRR / NPS / NPZ / RC / Attachment) from Huang et al. (2024), 120 pairs of D+ / D− sentences are used alongside 1.2M token-level RTs from 2K human self-paced reading experiments. Linear regressions \(\hat{y} = \beta_0 + \beta_1 \cdot \text{Surprisal} + \beta_2 \cdot \text{Length} + \beta_3 \cdot \text{LogFreq} + \ldots\) (including spillover) are fitted independently for each layer. Slowdowns estimated by the model at the disambiguating points \(t^*\) and \(t^*+1\) are compared against human data.
  3. Inter-layer Prediction Update as RT Feature: Surprisal update is defined as \(\text{SU}(w_t \mid w_{<t}) = S^{\text{shallow}}_t - S^{\text{deep}}_t = \log \frac{Q_t(w_t)}{P_t(w_t)}\), and extended to full-distribution \(\text{KL}(Q_t \| P_t)\) and the symmetric \(\text{JS}(Q_t \| P_t)\). These are integrated with surprisal to observe gains in Predictive Power (PPP).

Key Designs

  1. Layer-wise Surprisal Extraction via Logit-Lens + Tuned-Lens Robustness:

    • Function: Explicitly extracts "which layer is predicting the next word," transforming surprisal from a single scalar into a layer-wise curve.
    • Mechanism: The model's own unembedding matrix \(W_U\) (with LayerNorm) is applied to the hidden state \(h^{(l)}_{i}\) of the \(i\)-th token at each layer to obtain a prediction distribution. Subwords are handled via joint probability. Tuned-Lens (Belrose 2023) is used to ensure conclusions are not artifacts of logit-lens bias in early layers (Appendix B.1).
    • Design Motivation: Unfolding "model prediction" into a hierarchical sequence allows the investigation of which LM layers correspond to human fast vs. slow processing.

  2. D+/D− × ROI/¬ROI Four-Quadrant PPP Analysis:

    • Function: Precisely isolates the conditions under which the "layer depth → PPP improvement" trend occurs.
    • Mechanism: Tokens are categorized into 2×2 quadrants: ambiguous vs. unambiguous sentences (D+ vs D−) and inside vs. outside the disambiguating window (ROI: \(t^*-2\) to \(t^*+2\) vs ¬ROI). Pearson correlation between layer depth and PPP (\(\Delta\text{LL}\)) is calculated for each quadrant.
    • Design Motivation: This validates the dual alignment hypothesis—if humans switch to deep processing only during ambiguity resolution, the deep-layer advantage should only appear in the D+ ∩ ROI quadrant.

  3. Probability-Update Metrics (SU / KL / JS) as New RT Features:

    • Function: Uses the "prediction difference between shallow and deep layers" as an independent proxy for cognitive cost.
    • Mechanism: Three metrics are defined: (i) SU looks at \(\log Q_t(w_t)/P_t(w_t)\) only at the target word; (ii) KL calculates \(\mathbb{E}_{w \sim Q_t}[\text{SU}(w)]\) over the entire vocabulary; (iii) JS is the symmetric version. Respective per-layer surprisals are z-score normalized. These are then added to regressions to evaluate PPP gains.
    • Design Motivation: Based on the "shallow predicts first → deep revises" model, the magnitude of revision should represent effort. JS performs best due to its symmetry and inclusion of full-distribution information.

Loss & Training

  • Probe-only, No Training: The authors do not fine-tune LMs; they extract layer outputs via logit-lens/tuned-lens and run linear regressions on reading-time data.
  • Regression Model: \(\text{RT}(w_t) = \beta_0 + \beta_1 \text{Surprisal}(w_t) + \beta_2 \text{Length}(w_t) + \beta_3 \text{LogFreq}(w_t) + \text{spillover}(w_{t-1}, w_{t-2}) + \epsilon\).
  • PPP Metric: \(\Delta\text{LL} = \text{LL}_{\text{full}} - \text{LL}_{\text{baseline}}\), where the full model includes surprisal.
  • Filler Training / Target Testing: Regressions are trained on filler sentences and tested on target (D+/D−) sentences to avoid overfitting to garden-path structures.

Key Experimental Results

Main Results

Exp.1 (Fig.2): Estimated slowdown vs. human baseline (red line) for 19 LMs:

Construction Human Slowdown (ms) LM Estimated (Across Layers) Best Layer Position
MVRR ~100 Max ~50 (Late GPT2-xl) Late Layers
NPS ~45 Max ~25 Late Layers
NPZ ~100 Max ~50 Late Layers
RC ~25 Max ~15 Late Layers
Attachment ~10 ~5-10 Late Layers

Conclusion: All layers underestimate human slowdown, but later layers are significantly closer than early ones, contradicting the "early layer best" finding in naturalistic reading (Kuribayashi 2025).

Exp.2 (Tab.2): Pearson correlation (depth vs. PPP) for D+ ∩ ROI:

Model MVRR D+∩RoI NPS D+∩RoI NPZ D+∩RoI RC D+∩RoI Attachment D+∩RoI
GPT2-xl +0.88 -0.07 +0.88 +0.96 -0.32
OPT-13b +0.09 +0.71 +0.81 +0.88 +0.26
Pythia-12b +0.88 +0.93 +0.79 +0.97 +0.80

Pythia-12B shows strong positive correlations (+0.79 to +0.97) across all constructions in D+ ∩ ROI, whereas D− ∩ ROI shows negative correlations (-0.41 to -0.89).

Ablation Study

Exp.3 (Fig.4): Replacing or adding SU / KL / JS to surprisal regressions (average PPP across 19 LMs):

Feature Full Avg PPP RoI Avg PPP Note
Surprisal (last) Baseline Baseline Standard Approach
Surprisal Up (SU) > Baseline Marginal Target word only
KL(Q‖P) Sig. for most Partially Sig. Asymmetric
JS Best among three Partially Sig. Symmetric
Surprisal + JS Best overall Best overall Complementary

LR tests confirm that Surprisal+JS is significantly superior to Surprisal alone for MVR, RC, and Attachment.

Key Findings

  • Early Layers Fail on Syntactic Challenges: In MVRR ("fed the lamb remained"), early layers only perceive local co-occurrence, resulting in nearly identical surprisal for D+ and D−, indicating a lack of long-range syntactic sensitivity.
  • Dual Alignment is Scale-Dependent: In Pythia models (70M to 12B), the depth-PPP correlation in D+ ∩ ROI increases from 0 to +0.97. Scaling allows models to diverge into "shallow vs. deep" mechanisms, mirroring human dual-processing.
  • Metric Ranking: JS > KL > SU. JS provides extra explanatory power in ROI regions beyond simple surprisal.
  • Persistent Underestimation: Even with deep layers and JS, LMs only capture ~50% of human slowdown, suggesting that LM "reanalysis" efforts do not perfectly match human cognitive effort.

Highlights & Insights

  • Context-Dependent Alignment: The finding that the "best layer" depends on task difficulty shifts hierarchical probing from a static search to a dynamic perspective.
  • Precision via 2×2 Design: Using D+/D− and ROI/non-ROI allows for clean causal isolation of the deep-layer advantage.
  • Universal Framework for Cost: JS/KL/SU metrics provide a way to quantify "reanalysis effort" in any cognitive modeling task.
  • Honest Limitations: The authors report the persistent underestimation of slowdown rather than over-engineering features to fit the data.

Limitations & Future Work

  • Residual Underestimation: LM-based metrics still lack about 50% of the effort required for human garden-path processing.
  • Scope: Limited to English and written reading data.
  • Instruction-Tuning: Excludes SFT/RLHF models, which are known to distort cognitive alignment.
  • Theoretical Gap: The mapping between LM layers (discrete) and human brain dynamics (temporal) still requires a more robust theoretical bridge.
  • vs. Kuribayashi et al. (2025): Extends the "early layer best" theory to a "dual alignment" theory by introducing syntactic challenges.
  • vs. Huang et al. (2024): While prior work noted slowdown underestimation in final layers, this paper shows that while final layers are better than early ones for these cases, they are still insufficient.
  • vs. Tenney et al. (2019): aligns with the "BERT rediscovers the NLP pipeline" idea, showing a progression from shallow/POS to deep/syntactic-semantic processing.
  • vs. Li & Futrell (2024): Quantifies their theoretical shallow/deep processing models using hierarchical surprisal metrics.

Rating

  • Novelty: ⭐⭐⭐⭐ The discovery of "deep layer advantage specifically in garden-paths" is a counter-intuitive finding for the field.
  • Experimental Thoroughness: ⭐⭐⭐⭐ Extensive coverage across LMs and phenomena; robust validation with Tuned-Lens.
  • Writing Quality: ⭐⭐⭐⭐⭐ Clear framing and excellent visual representation of the core "dual alignment" concept.
  • Value: ⭐⭐⭐⭐ Provides new features (JS update) for the psycholinguistics community and dynamic interpretability insights for NLP.