Continuous Interpretive Steering for Scalar Diversity

1. Problem Statement

Current evaluations of pragmatic inference in Large Language Models (LLMs) rely heavily on prompt-based manipulations (e.g., varying instructions or few-shot examples). These methods are often brittle, sensitive to surface forms, and fail to probe the model's internal representations directly.

Furthermore, existing approaches often treat pragmatic enrichment as a binary target (pragmatic vs. literal), ignoring a critical linguistic phenomenon known as Scalar Diversity. Scalar diversity refers to the fact that different scalar items (e.g., some vs. warm) trigger pragmatic inferences (scalar implicatures) to different degrees. For instance, "some" strongly implies "not all," whereas "warm" rarely implies "not hot." Current LLM evaluation frameworks often collapse this graded variation, failing to assess whether models can modulate pragmatic strength at the item level.

Core Research Question: Can LLMs exhibit graded, item-sensitive pragmatic behavior, and can this be systematically probed and recovered through internal interventions rather than surface-level prompting?

2. Methodology

A. Dataset: GraSD (Graded Scalar Diversity)

To support the analysis of graded pragmatic behavior, the authors constructed a new dataset, GraSD.

• Source: Aggregated 121 unique <weak, strong> scalar item pairs from four prior psycholinguistic studies (van Tiel et al., 2016; Ronai and Xiang, 2021/2024; Pankratz and Van Tiel, 2021).
• Structure: Each pair generates three sentence variants:
  1. Anchor: Contains the weak scalar term (e.g., "Some pets...").
  2. Logical: Replaces the weak term with the strong alternative (e.g., "All pets...").
  3. Pragmatic: Encodes the scalar implicature (e.g., "Not all pets...").
• Scale: Using a constraint-based augmentation strategy with GPT-4o, the dataset was expanded to 121,000 contextualized sentence instances (100 per pair), ensuring linguistic diversity while preserving the intended pragmatic conditions.

B. Models

Experiments were conducted on four open-weight transformer-based LLMs to ensure architectural diversity:

• LLaMA3
• Qwen2
• Gemma2
• OLMo

C. Continuous Interpretive Steering (CIS)

The paper introduces CIS, a method that intervenes directly in the model's internal activation space during inference without updating model weights.

• Mechanism:
  1. Representation Extraction: Hidden states ($h_l$) are extracted from a fixed set of $k$ transformer layers and concatenated to form a multi-layer representation vector $h(x)$.
  2. Direction Definition: A "pragmatic direction" vector ($v$) is defined as the difference between the Pragmatic and Logical representations of a sample subset.
  3. Steering: The anchor representation is modified via the equation:
     $$h'(x) = h(x) + \alpha v$$
     Where $\alpha$ is a scalar coefficient controlling the magnitude of the intervention.
• Experimental Conditions:
  1. Uniform Activation Steering: Applies a fixed $\alpha$ to all items. This tests if a global shift induces pragmatic behavior but risks collapsing item-level variation.
  2. Graded Activation Steering: Varies $\alpha$ continuously across items based on their pre-defined "grade" (A–E) of scalar diversity strength. This tests if the model's internal geometry supports item-sensitive modulation.

D. Analysis Strategy

• Metric: Interpretive preference is measured via cosine similarity. An item is considered "pragmatic" if the steered anchor representation is closer to the Pragmatic variant than the Logical variant.
• Statistics:
  • Wilcoxon signed-rank tests: To determine if steering induces a significant shift from the baseline.
  • Spearman rank correlations: To assess whether the relative ordering of items (based on their baseline pragmatic sensitivity) is preserved under steering.

3. Key Results

A. Uniform vs. Graded Steering Effects

• Baseline: Without steering, models rarely produce pragmatic interpretations, showing a bias toward literal meanings.
• Uniform Steering:
  • Effect: Significantly increased the global proportion of pragmatic interpretations across all models ($p < .001$).
  • Limitation: It acted as a "coarse-grained" intervention. The Spearman correlation between baseline item preferences and uniform steering results was weak/non-significant (e.g., $\rho \approx 0.02$ for LLaMA3). This indicates that uniform steering forces all items toward pragmatics indiscriminately, collapsing scalar diversity.
• Graded Steering:
  • Effect: Also produced a significant global shift toward pragmatics ($p < .001$).
  • Advantage: It acted as a "fine-grained" intervention. The Spearman correlation was moderate and significant (e.g., $\rho = 0.38$ for LLaMA3, $p < .001$).
  • Finding: Graded steering preserved the item-level structure. Items with naturally higher scalar diversity (Grade A) showed larger interpretive shifts than those with lower diversity (Grade E), aligning with human linguistic patterns.

B. Distribution of Changes

• Uniform: The distribution of changes in pragmatic interpretation rates was sharply skewed, with most items showing similar, large increases.
• Graded: The distribution was broad and heterogeneous. The magnitude of change varied systematically according to the item's theoretical grade, demonstrating that the model's representation space encodes graded sensitivity.

4. Key Contributions

1. Continuous Interpretive Steering (CIS): A novel framework that treats steering strength as a continuous experimental variable to probe the geometry of pragmatic inference, moving beyond binary "success/failure" metrics.
2. GraSD Dataset: A large-scale, theoretically grounded dataset (121k instances) specifically designed to encode and evaluate graded scalar diversity.
3. Evidence of Internal Gradedness: Demonstrated that LLMs encode graded pragmatic sensitivity in their internal representation space. While standard prompting or uniform steering obscures this, controlled, graded intervention can systematically recover item-level variations.
4. Methodological Shift: Argues against relying solely on prompt-based evaluations, showing that internal activation steering provides a more stable and mechanistic view of model reasoning.

5. Significance and Implications

• Beyond Binary Evaluation: The study challenges the field to move away from binary assessments of pragmatic ability. It shows that "human-like" pragmatic behavior requires the capacity to modulate inference strength based on lexical items, not just a general ability to infer.
• Mechanistic Interpretability: CIS provides a tool for "steering" models to reveal latent structures. The fact that graded steering aligns with human scalar diversity suggests that LLMs have learned the graded nature of language, even if standard prompting fails to elicit it.
• Generalizability: While focused on scalar implicature, the CIS framework is applicable to other pragmatic phenomena that exhibit graded interpretive profiles, offering a new paradigm for investigating the internal organization of LLMs.

6. Limitations

• Scope: Focused exclusively on scalar implicature; generalization to other pragmatic phenomena is untested.
• Direction Estimation: The steering direction is estimated from a limited subset of data, which may not capture full contextual variability.
• Operationalization: Pragmatic preference is measured via representational similarity rather than explicit behavioral output, which may not perfectly map to end-user interpretations.