Derry Wijaya

Lucky Susanto, Anasta Pranawijayana, Cortino Sukotjo et al.

Large language models (LLMs) are increasingly adopted in high-stakes domains such as healthcare and medical education, where the risk of generating factually incorrect (i.e., hallucinated) information is a major concern. While significant efforts have been made to detect and mitigate such hallucinations, predicting whether an LLM's response is correct remains a critical yet underexplored problem. This study investigates the feasibility of predicting correctness by analyzing a general-purpose model (GPT-4o) and a reasoning-centric model (OSS-120B) on a multiple-choice prosthodontics exam. We utilize metadata and hallucination signals across three distinct prompting strategies to build a correctness predictor for each (model, prompting) pair. Our findings demonstrate that this metadata-based approach can improve accuracy by up to +7.14% and achieve a precision of 83.12% over a baseline that assumes all answers are correct. We further show that while actual hallucination is a strong indicator of incorrectness, metadata signals alone are not reliable predictors of hallucination. Finally, we reveal that prompting strategies, despite not affecting overall accuracy, significantly alter the models' internal behaviors and the predictive utility of their metadata. These results present a promising direction for developing reliability signals in LLMs but also highlight that the methods explored in this paper are not yet robust enough for critical, high-stakes deployment.

Khumaisa Nur'aini, Ayu Purwarianti, Alham Fikri Aji et al.

Adapting LLMs to low-resource languages is difficult: labeled data is scarce, full-model fine-tuning is unstable, and continued cross-lingual tuning can cause catastrophic forgetting. We propose Circuit-Targeted Supervised Fine-Tuning (CT-SFT): a counterfactual-free adaptation of CD-T (Contextual Decomposition Transformer) that uses a label-balanced mean baseline and task-directional relevance scoring to identify a sparse set of task-relevant attention heads in a proxy-language checkpoint, then transfer learns to a target language by updating only those heads (plus LayerNorm) via head-level gradient masking. Across NusaX-Senti and XNLI, CT-SFT improves cross-lingual accuracy over continued full fine-tuning while updating only a small subset of model parameters. We find an editing-preserving trade-off: harder transfers favor editing circuit heads, while easier transfers often favor near-zero (i.e., low-relevance heads) updates, preserving the source mechanism. CT-SFT also substantially reduces catastrophic forgetting, preserving proxy/source-language competence during transfer.