Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2602.13567 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866910022122864640 |
|---|---|
| author | Dhakal, Manish Jinadu, Uthman Budathoki, Anjila Sunderraman, Rajshekhar Ding, Yi |
| author_facet | Dhakal, Manish Jinadu, Uthman Budathoki, Anjila Sunderraman, Rajshekhar Ding, Yi |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2602_13567 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | DistillLens: Symmetric Knowledge Distillation Through Logit Lens Dhakal, Manish Jinadu, Uthman Budathoki, Anjila Sunderraman, Rajshekhar Ding, Yi Computation and Language 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. |
| title | DistillLens: Symmetric Knowledge Distillation Through Logit Lens |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2602.13567 |