Sentences like "She will go to France or Spain, or perhaps to Germany or France." appear formally redundant, yet become acceptable in contexts such as "Mary will go to a philosophy program in France or Spain, or a mathematics program in Germany or France." While this phenomenon has typically been analyzed using symbolic formal representations, we aim to provide an account grounded in artificial neural mechanisms. We first present new behavioral evidence from humans and large language models demonstrating the robustness of this apparent non-redundancy across contexts. We then show that, in language models, redundancy avoidance arises from two interacting mechanisms: models learn to bind contextually relevant information to repeated lexical items, and Transformer induction heads selectively attend to these context-licensed representations. We argue that this neural explanation sheds light on the mechanisms underlying context-sensitive semantic interpretation, and that it complements existing symbolic analyses.
Figures & tables
Figure 1: In context, our critical sentences become non-redundant and large LMs reliably produce the repeated item; in the control condition, they suppress the copy mechanism.
Figure 2: Left: We broadly see LM competency, with high rate of the repeated answer (X) as top choice. Right: We also see order preferences in LMs. Namely in small models we see a strong preference towards a match in the ordering of the second disjunction, but as models get larger this preference gets superseded by a preference for total matching between S1 and S2.
Figure 3: We perform activation patching across models and find disjunctive elements to be bound in different contexts.
Figure 4: Average attention patterns of induction heads on the control condition and each of the critical conditions (with different entity orders). Rows = induction heads (top-9 by causal importance), columns = entity positions in S1, strength of color = average attention from the critical token to each position.
Pragmatic reasoning in language models studies how models infer and communicate meanings shaped by context, intent, shared knowledge, and social norms.
How do language models reconcile conflicting contextual evidence within a prompt? We study this question using lexical redefinition, where an explicit instruction specifies how a word should be interpreted and a subsequent sentence provides additional evidence through its usage. Across 870 word pairs, we find that lexical redefinitions are not applied as fixed rules. Instead, the behaviour depends systematically on the relationship between the explicit instruction and the intervening sentence: the contextualised usage can preserve, weaken, or completely reverse the instruction. These results suggest that the interpretation of an explicit instruction remains sensitive to subsequent contextual evidence. Lexical redefinition therefore provides a controlled behavioural paradigm for studying how language models integrate multiple sources of information within a prompt.
Roksana Goworek, Haim Dubossarsky
Queen Mary University of London · University of Cambridge
If the same neuron activates for both "lender" and "riverside," standard metrics attribute the overlap to superposition--the neuron must be compressing two unrelated concepts. This work explores how much of the overlap is due a lexical confound: neurons fire for a shared word form (such as "bank") rather than for two compressed concepts. A 2x2 factorial decomposition reveals that the lexical-only condition (same word, different meaning) consistently exceeds the semantic-only condition (different word, same meaning) across models spanning 110M-70B parameters. The confound carries into sparse autoencoders (18-36% of features blend senses), sits in <=1% of activation dimensions, and hurts downstream tasks: filtering it out improves word sense disambiguation and makes knowledge edits more selective (p = 0.002).
Iyad Ait Hou, Rebecca Hwa
Department of Computer Science George Washington University Washington, D.C, USA
Recent benchmarks reveal that despite strong reasoning capabilities, large language models (LLMs) still struggle to faithfully apply complex contextual knowledge. These failures are often not wholesale reasoning collapses: in context-rich tasks, models may follow the central reasoning path while missing peripheral, persistent, or format-sensitive requirements.