When Perplexity Lies: Generation-Focused Distillation of Hybrid Sequence Models
-cross Abstract: Converting a pretrained Transformer into a more efficient hybrid model through distillation offers a promising approach to reducing…
Arxiv 2 versions
-
1 new sentence, beginning: “Converting a pretrained Transformer into a more efficient hybrid model through distillation offers a promising approach to…”
When Perplexity Lies: Generation-Focused Distillation of Hybrid Sequence Models
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, on overlapping benchmarks, we show that a 7B distilled model that nearly matches its teacher to within 0.2 pp under log-likelihood scoring falls behind by 20.8 pp when it must generate answers autoregressively. We investigate this phenomenon with GenDistill, a multi-stage pipeline we designed for distilling a pretrained Transformer into an efficient Hybrid Kimi Delta Attention (Hybrid-KDA) student. Using it as a controlled testbed on Qwen3-0.6B, we systematically ablate six design axes (training objective, loss masking, training duration, dataset selection, parameter freezing, and architecture choice) and evaluate every choice under both log-likelihood and generation-based protocols. 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, so 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 distillation recipe, using a Hybrid-KDA model as the student, 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 2-4x at 128K-token contexts.
-
When Perplexity Lies: Generation-Focused Distillation of Hybrid Sequence Models
-cross 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, on overlapping benchmarks, we show that a 7B distilled model that nearly matches its teacher to within 0.2 pp under log-likelihood scoring falls behind by 20.8 pp when it must generate answers autoregressively. We investigate this phenomenon with GenDistill, a multi-stage pipeline we designed for distilling a pretrained Transformer into an efficient Hybrid Kimi Delta Attention (Hybrid-KDA) student. Using it as a controlled testbed on Qwen3-0.6B, we systematically ablate six design axes (training objective, loss masking, training duration, dataset selection, parameter freezing, and architecture choice) and evaluate every choice under both log-likelihood and generation-based protocols. 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, so 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 distillation recipe, using a Hybrid-KDA model as the student, 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 2-4x at 128K-token contexts.
Versions are compared on the headline and summary the publisher puts in their feed. An edit to the body of an article that leaves both untouched will not appear here.
Why am I seeing this Ranked on source trust — arXiv
It ranks mainly on source trust: arXiv is the most reliable outlet we track on this subject, and is the only one on the story so far.
Link-outLink-out, because it did not clear the bar for a write-up. Link-out means we point at the publisher and say nothing of our own.
| Factor | Weight | Score | Contribution | Where it came from |
|---|---|---|---|---|
| Corroboration | 0.35 | 0.39 | +0.135 26% | 1 independent org on the story. Tier-3 aggregators never corroborate — they can show something is circulating, never that it is true. |
| Source trustleads | 0.25 | 0.85 | +0.212 41% | arXiv is the highest-trust source on this story and is first-party — the organisation announcing its own news. Trust is taken from the best source, not averaged. |
| Pickup rate | 0.20 | 0.00 | +0.000 0% | One counted organisation, so there is no spread to measure — nothing has picked this up to set a rate. |
| Freshness | 0.20 | 0.85 | +0.171 33% | Halves every 10 hours from the newest item on the story. This is the only factor that rewards a story for nothing more than being recent. |
Corroboration counts distinct organisations, once each, and only from tiers 1 and 2. Freshness halves every 10 hours, so this ranking is a snapshot and will differ at the next build.
What happened
-cross 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.
How this story arrived
Ordered by when each source was first observed, which is what the velocity figure is computed from. Publishers backdate; observed order does not.
- 01 Arxivfirst-party first seen When Perplexity Lies: Generation-Focused Distillation of Hybrid Sequence Models
Overclock clusters coverage from independent sources and grades it automatically. The figures above are computed, not editorial. This page summarises and links to reporting by the outlets named — follow the links for the original work.