arXiv:2606. 05378v1 Announce Type: new Abstract: We test whether a single screen-and-ablate recipe -- identify attention-head circuits by task-pattern selectivity, then verify by causal ablation against a matched-random null -- produces consistent mechanistic claims across model families.
arXiv:2605. 24059v2 Announce Type: replace Abstract: We present a three-step recipe for identifying attention-head circuits in pretrained transformers.
The study investigates the consistency and specificity of language model circuits across six tasks and five models, focusing on component-level (attention heads and MLP blocks) and neuron-level circuits. Component-level circuits are highly consistent and causally important but lack task specificity, as ablating a circuit for one task similarly harms performance on other tasks. Neuron-level circuits show higher task specificity but lower consistency, with overlap mainly between closely related tasks. The analysis of Llama‑3.2‑3B reveals that shared components are predominantly MLP blocks, while attention heads act as generic attention‑sink heads.
arXiv:2606. 02378v1 Announce Type: cross Abstract: We track the developmental trajectory of attention-head circuit formation across three 1B-class language models spanning two architecture families (dense transformer, mixture-of-experts) and two pretraining corpora (The Pile, DCLM): Pythia 1B, OLMo 1B-0724-hf, and OLMoE 1B-7B-0924.
arXiv:2606. 00930v1 Announce Type: cross Abstract: Mechanistic interpretability often assumes that probes identifying a representational signature also identify the circuit executing the corresponding computation.
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:2608. 11797v1 Announce Type: new Abstract: Model merging by task arithmetic works until it doesn't, and the field diagnoses why with magnitudes: layerwise representation bias, deviations from cross-task linearity, parameter overlap.
arXiv:2606. 08292v2 Announce Type: replace Abstract: Mechanistic studies often assign a component a role when removing it damages a behavior, its activation linearly encodes task information, and restoring that activation repairs the damage.
The paper investigates whether structural differences in circuits discovered by circuit discovery methods reflect distinct mechanisms. By varying input-token frequency while keeping the task constant, the authors find that although circuits appear specialized by frequency structurally, functional and representational analyses reveal no reliable differences, a phenomenon they call phantom specialization. Across multiple models and tasks, structurally distinct circuits implement the same computation, with core shared subgraphs recovering most of the performance and interchangeable internal representations confirmed by causal interventions.
arXiv:2607. 21692v1 Announce Type: new Abstract: Sparse attention reduces the cost of long contexts by allowing each query to read only selected parts of the input.
arXiv:2605. 07284v2 Announce Type: replace Abstract: A late-layer change learned during post-training may work on the base model's earlier state, or it may depend on earlier computation learned with it.
The study investigates how post‑training of large autoregressive language models (ARMs) into masked diffusion models (MDMs) affects their internal computation. Across two 7B ARM‑MDM families and four diagnostic tasks, the authors find that MDMs retain much of the ARM’s high‑attribution pathways on prefix‑dominant tasks, but reorganize computation toward earlier layers on globally constrained tasks. Component‑level probes reveal that ARMs depend on sharply specialized components, whereas MDMs show weaker specialization and more diffuse output‑space alignment.
arXiv:2607. 01002v1 Announce Type: cross Abstract: In long-context use, large language models frequently synthesize answers from the meaning of a relevant context span rather than literally copy-pasting them.