arXiv:2606. 12818v1 Announce Type: cross Abstract: Irrelevant numbers in a prompt can shift language model judgments, producing anchoring effects in numerical reasoning.
arXiv:2608. 14320v1 Announce Type: new Abstract: The anchoring effect is a cognitive bias in which an initial reference value shifts a later judgment toward itself.
arXiv:2609.25602v1 Announce Type: new Abstract: In language models, the choice between believing the prompt and believing the weights is made by a handful of identifiable attention heads. Instruction...
The paper introduces FTB Graph, a method for mapping the causal circuitry that determines the first-token language identity in multilingual language models. Using Edge Attribution Patching and exact activation patching across six architectures (GPT‑2, BLOOM‑560M, Pythia‑1B/2.8B, Qwen2.5‑1.5B Base/Instruct), the authors extract directed acyclic graphs that reveal deep or mid‑to‑deep broadcasting hubs, with notable differences among models. The study finds that first‑token routing is largely established during pretraining and largely preserved by instruction tuning, while linear gradient approximations can diverge from causal interventions, underscoring the need for exact‑patching verification.
arXiv:2604. 04385v5 Announce Type: replace-cross Abstract: We localize the policy routing mechanism in alignment-trained language models.
The paper introduces a three-level evaluation framework—behavioral deployment, LM-head readout, and probe recoverability—to distinguish whether a language model fails a syntactic test by not encoding structure or by failing to use it. Using a trilingual control-dependency benchmark, the authors find that probe recoverability consistently exceeds LM-head readout, which in turn exceeds behavioral deployment across seven models and three languages, with the largest gap observed in Qwen3-0.6B Instruct. Layer-localized activation patching shows that instruction tuning shifts the decoded layer later, suggesting decoding favors surface shortcuts and that behavioral evaluation understates what models encode while probing alone overstates what they deploy.
arXiv:2606. 01202v1 Announce Type: new Abstract: Language models do not simply choose an answer at the output layer.
arXiv:2601. 09624v2 Announce Type: replace-cross Abstract: Machine unlearning is becoming essential for building trustworthy and compliant language models.
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.
The paper introduces a sequential activation patching framework to study how Chain-of-Thought (CoT) prompting influences large language models over multiple generated tokens. By tracking CoT-conditioned attention-head activations across token positions and aggregating them with Part-of-Speech guidance, the authors identify distributed head sets that jointly contribute to answer generation. Targeted zero-ablation experiments confirm that these heads are functionally important, affecting mechanisms such as reasoning-trajectory maintenance, answer anchoring, exemplar-target separation, and numerical generation.
arXiv:2509. 24808v2 Announce Type: replace Abstract: Explaining why a language model produces a particular output requires local, input-level explanations.
arXiv:2609.15545v1 Announce Type: new Abstract: Hybrid language models can improve capability as well as efficiency, raising the question of how architectural complementarity becomes learned computat...
The paper investigates how steering vectors influence large language models (LLMs) by conducting a mechanistic case study on refusal behavior. Using a multi-token activation patching framework, the authors find that steering methods primarily target the OV circuit of the attention mechanism, largely ignoring the QK circuit, and that these circuits are functionally interchangeable across different steering approaches. The study also shows that steering vectors can be sparsified by 85–96% with minimal performance loss and that key dimensions are consistently identified across methods.