# Different Teachers, Different Capabilities: Sub-1B On-Device Distillation for Structured Text Enrichment

arXiv: https://arxiv.org/abs/2607.08268 · HTML: https://mlsanity.com/paper/2607.08268/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 40 of its 44 references that are in the map.
- First posted on arXiv 9 Jul 2026.

## Builds on (most cited first, 30 of 40)

- [Distilling the Knowledge in a Neural Network](https://mlsanity.com/paper/1503.02531/) (2015)
- [Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena](https://mlsanity.com/paper/2306.05685/) (2023)
- [Self-Instruct: Aligning Language Models with Self-Generated Instructions](https://mlsanity.com/paper/2212.10560/) (2022)
- [G-Eval: NLG Evaluation using GPT-4 with Better Human Alignment](https://mlsanity.com/paper/2303.16634/) (2023)
- [FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation](https://mlsanity.com/paper/2305.14251/) (2023)
- [Large Language Models are not Fair Evaluators](https://mlsanity.com/paper/2305.17926/) (2023)
- [SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models](https://mlsanity.com/paper/2303.08896/) (2023)
- [Sequence-Level Knowledge Distillation](https://mlsanity.com/paper/1606.07947/) (2016)
- [Distilling Step-by-Step! Outperforming Larger Language Models with Less Training Data and Smaller Model Sizes](https://mlsanity.com/paper/2305.02301/) (2023)
- [LLM Evaluators Recognize and Favor Their Own Generations](https://mlsanity.com/paper/2404.13076/) (2024)
- [Improved Knowledge Distillation via Teacher Assistant](https://mlsanity.com/paper/1902.03393/) (2019)
- [Do NOT Think That Much for 2+3=? On the Overthinking of o1-Like LLMs](https://mlsanity.com/paper/2412.21187/) (2024)
- [Benchmarking Large Language Models for News Summarization](https://mlsanity.com/paper/2301.13848/) (2023)
- [Justice or Prejudice? Quantifying Biases in LLM-as-a-Judge](https://mlsanity.com/paper/2410.02736/) (2024)
- [Replacing Judges with Juries: Evaluating LLM Generations with a Panel of Diverse Models](https://mlsanity.com/paper/2404.18796/) (2024)
- [A Survey on Knowledge Distillation of Large Language Models](https://mlsanity.com/paper/2402.13116/) (2024)
- [A Survey on Model Compression for Large Language Models](https://mlsanity.com/paper/2308.07633/) (2023)
- [Self-Preference Bias in LLM-as-a-Judge](https://mlsanity.com/paper/2410.21819/) (2024)
- [MiniCheck: Efficient Fact-Checking of LLMs on Grounding Documents](https://mlsanity.com/paper/2404.10774/) (2024)
- [MobileLLM: Optimizing Sub-billion Parameter Language Models for On-Device Use Cases](https://mlsanity.com/paper/2402.14905/) (2024)
- [InFoBench: Evaluating Instruction Following Ability in Large Language Models](https://mlsanity.com/paper/2401.03601/) (2024)
- [CheckEval: A reliable LLM-as-a-Judge framework for evaluating text generation using checklists](https://mlsanity.com/paper/2403.18771/) (2024)
- [UniversalNER: Targeted Distillation from Large Language Models for Open Named Entity Recognition](https://mlsanity.com/paper/2308.03279/) (2023)
- [JSONSchemaBench: A Rigorous Benchmark of Structured Outputs for Language Models](https://mlsanity.com/paper/2501.10868/) (2025)
- [TIGERScore: Towards Building Explainable Metric for All Text Generation Tasks](https://mlsanity.com/paper/2310.00752/) (2023)
- [TICKing All the Boxes: Generated Checklists Improve LLM Evaluation and Generation](https://mlsanity.com/paper/2410.03608/) (2024)
- [Distillation Scaling Laws](https://mlsanity.com/paper/2502.08606/) (2025)
- [Are Reasoning Models More Prone to Hallucination?](https://mlsanity.com/paper/2505.23646/) (2025)
- [Unveiling the Key Factors for Distilling Chain-of-Thought Reasoning](https://mlsanity.com/paper/2502.18001/) (2025)
- [OptimalThinkingBench: Evaluating Over and Underthinking in LLMs](https://mlsanity.com/paper/2508.13141/) (2025)

## Abstract (arXiv)

High-volume structured extraction pays a large model's latency on every item, so distilling the task into a small on-device model is attractive: comparable output at a fraction of the time and cost. We measure what that distillation actually delivers, per sub-task. Each news article is mapped to one JSON object with a short summary and five categorical labels. We distill an 8B reasoning teacher (deepseek-r1:8b) into a 0.6B student (Qwen3-0.6B; QLoRA, three seeds), and add two teacher controls: a same-size non-reasoning teacher and a larger managed pipeline. A blinded, reference-free, three-judge panel scores every arm against the full article, alongside two non-distillation baselines, few-shot prompting and constrained decoding. The student runs at about 0.8 s per article against the teacher's 39 s, and recovers 58% of the base-to-teacher gap on summary quality, beating its primary baseline (constrained decoding) by +16.8 points and few-shot prompting by a secondary +4.9. A same-size non-reasoning teacher trains a student no better than the untuned base, so the summary gain follows from the teacher's reasoning nature rather than its scale. Capabilities then split by teacher: the reasoning teacher transfers writing quality and the managed pipeline transfers label diversity, while a same-size instruction teacher's students stay more grounded on the 22 short, thin-source articles in the 93-item test set (74 versus 55 faithful), where the reasoning-lineage student fabricates. That grounding difference is a consistent ordering rather than a significant aggregate effect, and the subgroup is small, so we report it as a direction. Because no single engine wins every field, the deliverable is a per-field routing map for on-device enrichment.