France or Spain or Germany or France: A Neural Account of Non-Redundant Redundant Disjunctions

1. Problem Statement

The paper investigates a linguistic phenomenon where sentences containing formally redundant disjunctions (e.g., "She will go to France or Spain, or perhaps to Germany or France") are perceived as acceptable and non-redundant when embedded in specific contexts (e.g., "Mary will go to a philosophy program in France or Spain, or a mathematics program in Germany or France").

• The Paradox: Logically, repeating "France" ($A \lor B \lor C \lor A$) is redundant. However, in context, the repetition is licensed because the two occurrences of "France" refer to distinct possibilities (a philosophy program vs. a mathematics program).
• Existing Analysis: Previous work (Mandelkern, 2024) explains this using symbolic formal semantics, positing that "or" introduces possibility conjuncts ($\Diamond p \land \Diamond q$) resolved against distinct modal domains.
• The Gap: The authors aim to provide a neural account of this phenomenon. Instead of implementing symbolic logic directly, they seek to explain how Large Language Models (LLMs) achieve the same functional outcome (non-redundant repetition) through neural mechanisms like contextual binding and attention patterns.

2. Methodology

The study employs a multi-stage approach combining behavioral experiments with mechanistic interpretability techniques.

A. Behavioral Experiments (Experiments 1a, 1b, 1c)

• Stimuli Design: The authors created a 2x2x2 factorial design involving three entities (X, Y, Z) from various semantic domains (e.g., countries, jobs).
  • Critical Condition: Sentence 1 (S1) establishes a structured disjunction where X appears in two distinct contexts (e.g., "Philosophy in X or Y" vs. "Math in Z or X"). Sentence 2 (S2) asks for a completion, expecting the repetition of X.
  • Control Condition: S1 is a simple list without disjunctive structure (e.g., "X, Y, and Z, but especially X"). Repeating X in S2 here is genuinely redundant.
• Human Subjects: 236 participants completed sentence completion tasks via Prolific.
• Language Models: 12 autoregressive models were tested (Pythia, GPT-2, LLaMa2) ranging from 70M to 7B parameters. The metric was the "generation rate" (probability) of the correct repeated token (X) versus other tokens.
• Ordering Analysis: The study manipulated the order of disjuncts (e.g., X or Y vs. Y or X) to test sensitivity to surface patterns.

B. Mechanistic Interpretability (Experiments 2 & 3)

• Activation Patching (Experiment 2): A causal intervention method used to determine if repeated tokens encode distinct contextual information.
  • Procedure: The model processed a base sentence. Activations from a "source" sentence (where the repeated token was bound to a specific context, e.g., "France for physics") were patched into the base sentence at the position of the repeated token.
  • Goal: To see if patching the contextually bound representation changes the probability of the subsequent context (e.g., "for physics" vs. "for math").
• Induction Head Analysis (Experiment 3):
  • Procedure: The authors analyzed "induction heads" (attention heads known for pattern repetition) in LLaMa2-7B.
  • Goal: To measure attention weights at the position of the repeated token in Critical vs. Control conditions to understand how the model selects the correct instance to copy.

3. Key Contributions

1. Neural Mechanism for Semantic Non-Redundancy: The paper demonstrates that LMs do not rely on symbolic logic to handle non-redundant disjunctions. Instead, they use contextual binding (where token representations absorb surrounding context) and selective attention (induction heads) to differentiate repeated lexical items.
2. Scaling Laws for Contextual Binding: The ability to correctly handle non-redundant disjunctions emerges only in models above a certain parameter threshold (~400M parameters), suggesting this is a capacity that scales with model size.
3. Divergence Between Humans and Models: While both humans and large LMs accept the repetition, they differ in ordering sensitivity. LMs show strong sensitivity to the surface order of the second disjunction (matching S1 and S2), whereas humans showed null effects in this specific experimental setup.
4. Complementarity of Symbolic and Neural Accounts: The authors argue that the neural account (contextual representation + induction heads) achieves the same functional result as the symbolic account (modal domains), validating the claim that neural networks can replicate complex semantic phenomena without explicit symbolic rules.

4. Results

Behavioral Results

• Humans: Produced the repeated item (X) in 72.3% of critical trials vs. 28.2% in control trials ($p < .001$), confirming humans treat the repetition as non-redundant in structured contexts.
• LLMs: Large models (>400M params) replicated human behavior, generating X in critical conditions and suppressing it in controls. Small models (<400M) failed, often exhibiting "dumb copying" (repeating the most recent noun regardless of context).
• Ordering Effects: LMs showed a strong preference for the second disjunction's order to match between S1 and S2. This effect was most pronounced in smaller models and evolved into a preference for total matching in larger models. Humans showed no significant ordering effects.

Mechanistic Results

• Activation Patching: When activations from a contextually bound "France" (e.g., associated with "physics") were patched into the model, the probability of the corresponding suffix ("for physics") increased significantly. This proves that the model causally encodes distinct contextual information into the representation of the repeated token.
• Induction Heads: Attention analysis revealed that induction heads are selectively activated in the Critical condition (specifically at the second occurrence of X) but suppressed in the Control condition. The strength of this activation correlates with the surface order match of the second disjunction, providing a mechanistic explanation for the ordering effects observed in Experiment 1.

5. Significance

• Linguistic Theory: The paper bridges the gap between formal semantics and neural network interpretability. It suggests that the "possibility modal" required by symbolic theories can be realized neurally through the contextual embedding of tokens and the selective deployment of induction heads.
• Model Interpretability: It provides a concrete example of how LMs solve complex semantic tasks not by "reasoning" in a human-like way, but by learning to bind surface patterns to context and using attention mechanisms to retrieve the correct pattern.
• Future Directions: The findings highlight that while LMs can replicate human linguistic judgments, the underlying mechanisms (surface-order sensitivity) may differ from human cognition. This suggests that future linguistic theories should account for how neural architectures process structure, potentially requiring new frameworks that integrate both symbolic and neural perspectives.

In conclusion, the paper successfully argues that the phenomenon of "non-redundant redundant disjunctions" can be fully explained by neural mechanisms: contextual binding creates distinct representations for repeated words, and induction heads selectively attend to these representations to enable copying, effectively solving the problem without explicit symbolic logic.