The paper investigates whether attention heads in transformer language models that specialize in syntactic relations causally influence the models’ syntactic predictions. Experiments show mixed to negative evidence: even when these heads are forced to attend to a distractor noun, the model’s predictions still favor the correct subject. Interventions on syntax-specialized heads affect predictions more than random heads, but only for one of two proposed measures.
arXiv:2610.16837v1 Announce Type: cross Abstract: Syntactic language models (SLMs) that generate sentences along with their syntactic tree structures have shown promise in both syntactic generalizati...
The study evaluates eleven autoregressive transformer models on English agreement attraction scenarios using a surprisal-based approach. Results show that while transformers match human reading times for prepositional phrase configurations, they perform poorly on object‑extracted relative clauses, with predictions diverging across models and failing to capture human interference patterns. The authors argue that current transformers cannot adequately model human morphosyntactic processing and call for more rigorous, comprehensive testing to avoid misleading conclusions from limited syntactic setups.
The paper investigates the reliability of attention‑head ablation as a causal inference tool in language models. Using GPT‑2 small, the authors find that a natural post‑projection zeroing method is almost uncorrelated with a corrected pre‑projection ablation and yields a completely different set of top‑5 important heads. They also show that binary accuracy can mask effects near performance floors or ceilings, whereas gold‑token log‑probability provides a graded signal. By employing a discovery/held‑out split and 1,000 matched random‑head and layer‑matched‑head controls, the corrected per‑head effect ranking remains highly stable (Spearman ρ = 0.974) and the top‑5 heads significantly outperform both control distributions (Monte Carlo p = 0.001). However, evidence for task specificity is weak on GPT‑2, and replication on DistilGPT‑2 confirms the intervention‑semantic and matched‑control findings.
arXiv:2610.33299v1 Announce Type: new Abstract: To understand the emergence of grammatical behavior in language models, we track circuits for subject-verb number agreement in pythia-1b and ten seeds...
arXiv:2608.28924v1 Announce Type: new Abstract: Linguistic theory has long recognized cross-linguistic syntactic regularities, leading to claims that these similar structures are processed by similar...
The paper introduces Controlled Evidence Aggregation (CEA), a method that manipulates the counts and order of two symbols in fixed prompts to probe a frozen Transformer’s preference between two answer tokens. Using Balanced Utility Causal Selection (BUC), the authors identify small sets of attention heads whose ablation removes a large portion of the count effect, and show that these same heads also influence performance on a natural language task. The study demonstrates that symbolic probes can uncover compact head sets with causal impact on both synthetic and real-world language decisions.
arXiv:2601.19926v3 Announce Type: replace-cross Abstract: We present a systematic review of 337 articles evaluating the syntactic abilities of Transformer-based language models (TLMs), reporting on o...
arXiv:2608. 12387v1 Announce Type: cross Abstract: Positional biases such as recency and primacy effects have been documented in large language models (LLMs), yet the underlying mechanism by which these models make their evaluations remains poorly understood.
The study examines how language models (LMs) perform on semantic focus-sensitivity tasks using two evaluation methods: explicit prompting and token-probability-based scoring. Across models of varying sizes, token probabilities consistently achieve above-chance performance, while prompting generally aligns with prior findings but can surpass token probabilities in larger models. The authors attribute this to LMs’ limited use of focus markers, which can distort token-probability results, and demonstrate that prompting remains robust even when focus markers are randomized.
The paper investigates indirect evidence in language models by applying influence functions to identify training sentences that affect predictions on 15 grammatical phenomena from BLiMP. After filtering out direct evidence, the authors select the 60,000 most influential remaining sentences and show that removing them significantly degrades held‑out accuracy, more so than random removal. They further analyze the characteristics of this indirect evidence and demonstrate that clustering phenomena based on shared indirect evidence recovers known linguistic categories.
The study investigates how task context reshapes the roles of shared attention heads in large language models. By applying identical graph facts and measuring interventions on attention heads and downstream activations, the authors show that task changes redistribute influence across five open‑weight models, causing the same heads to respond to different facts and affect later components differently. Linear response maps before and after activation edits reveal significant reductions in prediction error, indicating that task context alters both the facts processed and the downstream impact of shared components.
Fine‑tuning reshapes internal representations of large language models, affecting attention patterns and layer‑wise activations. The study shows that components identified by EAP as important for task performance cluster in specific layers, yet these layers do not align with those undergoing the largest representational changes. Additionally, overlapping EAP components across different tasks do not guarantee cross‑task transfer and can even degrade performance when tasks differ in nature.