When Perplexity Lies: Generation-Focused Distillation of Hybrid Sequence Models

• URL: http://arxiv.org/abs/2603.26556v1
• Date: Fri, 27 Mar 2026 16:16:23 GMT
• Title: When Perplexity Lies: Generation-Focused Distillation of Hybrid Sequence Models
• Abstract summary: We propose a Hybrid Kimi Delta Attention (Hybrid-KDA) architecture paired with GenDistill.<n>We find that log-likelihood-based evaluation consistently underestimates the gap between teacher and student.
• Score: 18.966595547209824
• License: http://creativecommons.org/licenses/by-nc-nd/4.0/
• Abstract: Converting a pretrained Transformer into a more efficient hybrid model through distillation offers a promising approach to reducing inference costs. However, achieving high-quality generation in distilled models requires careful joint design of both the student architecture and the distillation process. Many prior distillation works evaluate downstream multiple-choice benchmarks by ranking candidate answers with log-likelihood rather than requiring autoregressive generation, which can obscure important differences in model quality. For example, we show that a 7B parameter distilled model that nearly matches its teacher to within 0.2\,pp under log-likelihood scoring actually falls behind by 20.8\,pp when the model must generate answers autoregressively. We propose a Hybrid Kimi Delta Attention (Hybrid-KDA) architecture paired with GenDistill, a multi-stage distillation pipeline, and use generation-based evaluation throughout to guide design decisions. Applying this approach to Qwen3-0.6B, we systematically ablate six design axes: training objective, loss masking, training duration, dataset selection, parameter freezing, and architecture choice. We find that log-likelihood-based evaluation consistently underestimates the gap between teacher and student, and can in some cases reverse the ranking of design choices, meaning that conclusions drawn from perplexity-only evaluation may be misleading. Among the factors we study, dataset selection, completion-only masking, and freezing attention layers during post-training have the largest impact on generation quality. Our best Hybrid-KDA model retains 86--90\% of teacher accuracy on knowledge benchmarks while reducing KV cache memory by up to 75\% and improving time-to-first-token by 2--4$\times$ at 128K-token contexts.

Related papers

• Cognitive-Uncertainty Guided Knowledge Distillation for Accurate Classification of Student Misconceptions [11.509375725817195]
  We propose a two-stage knowledge distillation framework that mines high-value samples from existing data.<n>With augmented training on only 10.30% of filtered samples, we achieve MAP@3 of 0.9585 (+17.8%) on the MAP-Charting dataset.<n>Using only a 4B parameter model, we attain 84.38% accuracy on cross-topic tests of middle school algebra misconception benchmarks.
  arXiv  Detail & Related papers  (2026-05-14T12:17:38Z)
• Post-Training Probability Manifold Correction via Structured SVD Pruning and Self-Referential Distillation [0.0]
  We introduce Sparse Knowledge Distillation (SparseKD), a method that compresses transformer models by combining structured SVD pruning with self-referential knowledge distillation.<n>SparseKD requires no external super-teacher, no architectural changes, and no custom inference kernels, making it immediately deployable with existing infrastructure.
  arXiv  Detail & Related papers  (2026-01-30T22:42:02Z)
• Learning from Stochastic Teacher Representations Using Student-Guided Knowledge Distillation [64.15918654558816]
  Self-distillation (SSD) training strategy is introduced for filtering and weighting teacher representation to distill from task-relevant representations only.<n> Experimental results on real-world affective computing, wearable/biosignal datasets from the UCR Archive, the HAR dataset, and image classification datasets show that the proposed SSD method can outperform state-of-the-art methods.
  arXiv  Detail & Related papers  (2025-04-19T14:08:56Z)
• Warmup-Distill: Bridge the Distribution Mismatch between Teacher and Student before Knowledge Distillation [84.38105530043741]
  We propose Warmup-Distill, which aligns the distillation of the student to that of the teacher in advance of distillation.<n>Experiments on the seven benchmarks demonstrate that Warmup-Distill could provide a warmup student more suitable for distillation.
  arXiv  Detail & Related papers  (2025-02-17T12:58:12Z)
• One-for-All: Bridge the Gap Between Heterogeneous Architectures in Knowledge Distillation [69.65734716679925]
  Knowledge distillation has proven to be a highly effective approach for enhancing model performance through a teacher-student training scheme. Most existing distillation methods are designed under the assumption that the teacher and student models belong to the same model family. We propose a simple yet effective one-for-all KD framework called OFA-KD, which significantly improves the distillation performance between heterogeneous architectures.
  arXiv  Detail & Related papers  (2023-10-30T11:13:02Z)
• DisWOT: Student Architecture Search for Distillation WithOut Training [0.0]
  We explore a novel training-free framework to search for the best student architectures for a given teacher. Our work first empirically show that the optimal model under vanilla training cannot be the winner in distillation. Our experiments on CIFAR, ImageNet and NAS-Bench-201 demonstrate that our technique achieves state-of-the-art results on different search spaces.
  arXiv  Detail & Related papers  (2023-03-28T01:58:45Z)
• EmbedDistill: A Geometric Knowledge Distillation for Information Retrieval [83.79667141681418]
  Large neural models (such as Transformers) achieve state-of-the-art performance for information retrieval (IR) We propose a novel distillation approach that leverages the relative geometry among queries and documents learned by the large teacher model. We show that our approach successfully distills from both dual-encoder (DE) and cross-encoder (CE) teacher models to 1/10th size asymmetric students that can retain 95-97% of the teacher performance.
  arXiv  Detail & Related papers  (2023-01-27T22:04:37Z)
• A Self-Paced Mixed Distillation Method for Non-Autoregressive Generation [135.84684279852098]
  Non-Autoregressive (NAR) models significantly under-perform Auto-regressive (AR) models on various language generation tasks. Among the NAR models, BANG is the first large-scale pre-training model on English un-labeled raw text corpus. We propose a novel self-paced mixed distillation method to further improve the generation quality of BANG.
  arXiv  Detail & Related papers  (2022-05-23T09:54:53Z)
• Sparse Distillation: Speeding Up Text Classification by Using Bigger Models [49.8019791766848]
  Distilling state-of-the-art transformer models into lightweight student models is an effective way to reduce computation cost at inference time. In this paper, we aim to further push the limit of inference speed by exploring a new area in the design space of the student model. Our experiments show that the student models retain 97% of the RoBERTa-Large teacher performance on a collection of six text classification tasks.
  arXiv  Detail & Related papers  (2021-10-16T10:04:14Z)
• Distill2Vec: Dynamic Graph Representation Learning with Knowledge Distillation [4.568777157687959]
  We propose Distill2Vec, a knowledge distillation strategy to train a compact model with a low number of trainable parameters. Our experiments with publicly available datasets show the superiority of our proposed model over several state-of-the-art approaches.
  arXiv  Detail & Related papers  (2020-11-11T09:49:24Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.