The paper investigates K/V-cache interventions—transplanting a target-conditioned key/value trajectory into a source-persona generation—as a method for controlling persona in decoder-only language models. Experiments on Llama‑3.1‑8B across 13 configurations reveal that strong representation alignment (measured by V‑gap) does not guarantee behavioral persona expression, with only mid‑layer replacements achieving both alignment and lexical diversity. Position perturbations uniformly suppress persona expression, highlighting that representation similarity alone is insufficient to predict downstream behavior.
Fine-tuning enables a source model to acquire desired capabilities and behaviors in a target domain while retaining much of its general-purpose competence. However, this adaptation process can also degrade alignment properties that were present in the source model.
arXiv:2607. 22676v1 Announce Type: new Abstract: Post-training is a key mechanism for adapting large language models to downstream tasks.
arXiv:2608. 04347v1 Announce Type: new Abstract: Fine-tuning enables a source model to acquire desired capabilities and behaviors in a target domain while retaining much of its general-purpose competence.
arXiv:2608. 13482v1 Announce Type: cross Abstract: As language-model-based AI is increasingly deployed in autonomous settings, aligning its goals and values with those of humans becomes critical.
arXiv:2608.11025v2 Announce Type: replace Abstract: Emergent misalignment (EM) is the phenomenon where fine-tuning a language model on a narrow task leads to harmful behavior in unrelated domains. A...
arXiv:2606. 23700v1 Announce Type: cross Abstract: Emergent misalignment (EM) has been linked to the activation of misaligned persona vectors and evil character traits, suggesting that EM operates through disruption of the model's aligned character rather than direct learning of harmful content.
arXiv:2606. 27242v1 Announce Type: new Abstract: Training-free source selection for LLM families with shared vocabularies arises in scientific string domains such as SMILES, protein, and genomic sequences, where candidate corpora share a tokenizer but differ in prediction targets.
arXiv:2605. 24583v3 Announce Type: replace Abstract: Comparing a model's internal activations before and after alignment is a natural way to ask what safety training changes: one forms the matrix of paired aligned-minus-base activations on safety-relevant inputs and reads off its effective rank or top direction.
arXiv:2606. 01060v1 Announce Type: cross Abstract: Preference alignment has substantially improved the observable behavior of large language models, yet it remains unclear what alignment changes internally.
Emergent misalignment (EM) is the phenomenon where fine-tuning a language model on a narrow task leads to harmful behavior in unrelated domains. A leading mechanistic account attributes EM to persona features: latent directions acquired during pre-training that misaligned fine-tuning amplifies.
The study investigates how layer‑wise intervention responses in language models change over the course of pretraining, using single‑block identity bypass across multiple checkpoints and model‑domain combinations. It finds that while depth ordering of responses persists, their magnitudes shift, with nearby checkpoints showing stronger rank correspondence than distant ones and large changes occurring at positions that recur across samples and transfer across evaluation domains. Controlled experiments reveal that these longitudinal changes cannot be explained by a single downstream sensitivity and depend on perturbation strength and direction, indicating that layer sensitivity is structured but dynamic.
arXiv:2607. 01239v1 Announce Type: cross Abstract: Character-level perturbations bypass safety alignment in modern LLMs despite leaving prompts human-readable.