DistillLens: Symmetric Knowledge Distillation Through Logit Lens

Standard Knowledge Distillation (KD) compresses Large Language Models (LLMs) by optimizing final outputs, yet it typically treats the teacher's intermediate layer's thought process as a black box. While feature-based distillation attempts to bridge this gap, existing methods (e.g., MSE and asymmetric KL divergence) ignore the rich uncertainty profiles required for the final output. In this paper, we introduce DistillLens, a framework that symmetrically aligns the evolving thought processes of student and teacher models. By projecting intermediate hidden states into the vocabulary space via the Logit Lens, we enforce structural alignment using a symmetric divergence objective. Our analysis proves that this constraint imposes a dual-sided penalty, preventing both overconfidence and underconfidence while preserving the high-entropy information conduits essential for final deduction. Extensive experiments on GPT-2 and Llama architectures demonstrate that DistillLens consistently outperforms standard KD and feature-transfer baselines on diverse instruction-following benchmarks. The code is available at https://github.com/manishdhakal/DistillLens.

Loss Modeling for Multi-Annotator Datasets

Accounting for the opinions of all annotators of a dataset is critical for fairness. However, when annotating large datasets, individual annotators will frequently provide thousands of ratings which can lead to fatigue. Additionally, these annotation processes can occur over multiple days which can lead to an inaccurate representation of an annotator's opinion over time. To combat this, we propose to learn a more accurate representation of diverse opinions by utilizing multitask learning in conjunction with loss-based label correction. We show that using our novel formulation, we can cleanly separate agreeing and disagreeing annotations. Furthermore, we demonstrate that this modification can improve prediction performance in a single or multi-annotator setting. Lastly, we show that this method remains robust to additional label noise that is applied to subjective data.