Papers with student model
Fabricator: An Open Source Toolkit for Generating Labeled Training Data with Teacher LLMs (2023.emnlp-demo)
Copied to clipboard
| Challenge: | Recent research addresses the bottleneck of producing labeled training data for NLP tasks. |
| Approach: | They propose a method that generates labeled data that can be used to train a downstream NLP model. |
| Outcome: | The proposed model enables an LLM to generate labeled data that can be used to train a downstream NLP model. |
Weight-Inherited Distillation for Task-Agnostic BERT Compression (2024.findings-naacl)
Copied to clipboard
| Challenge: | Knowledge Distillation (KD) is a predominant approach for BERT compression. |
| Approach: | They propose a weight-inherited distillation method which directly transfers knowledge from the teacher to a compact student model by inheriting the weights. |
| Outcome: | The proposed method outperforms state-of-the-art KD-based methods on GLUE and SQUAD benchmarks. |
Query Distillation: BERT-based Distillation for Ensemble Ranking (2020.coling-industry)
Copied to clipboard
| Challenge: | Recent years have witnessed substantial progress in the development of neural ranking networks, but an increasingly heavy computational burden due to growing numbers of parameters and the adoption of model ensembles. |
| Approach: | They propose a two-stage distillation method that allows a smaller student model to be trained while benefiting from the better performance of the teacher model. |
| Outcome: | The proposed method shows higher-quality rankings compared to the teacher model. |
Text2Sql: Pure Fine-Tuning and Pure Knowledge Distillation (2025.naacl-industry)
Copied to clipboard
| Challenge: | Text2Sql is a task that translates natural language questions and database schemas into SQL queries. |
| Approach: | They employ pure fine-tuning strategy to reduce redundancy by using only 53% of the baseline prompt length to fine- tune the model. |
| Outcome: | The model outperforms the baseline model by 8.2% and 8.6% in Test-suite accuracy (TS) and exact-set-match accuracy (EM) under the most refined Spider dev set of prompts, the model achieves 73.5% and 75.4%, respectively, approaching state-of-the-art (SOTA) levels. |
Revisiting Intermediate Layer Distillation for Compressing Language Models: An Overfitting Perspective (2023.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods for knowledge distillation (KD) are prone to overfitting to training datasets . recent advances in NLP have shown that using PLMs such as BERT and RoBERTa on downstream tasks is effective. |
| Approach: | They propose a consistency-regularized knowledge distillation method which mitigates overfitting of existing methods. |
| Outcome: | The proposed method outperforms existing methods on the GLUE benchmark and synthetic datasets. |
Do we need Label Regularization to Fine-tune Pre-trained Language Models? (2023.eacl-main)
Copied to clipboard
Ivan Kobyzev, Aref Jafari, Mehdi Rezagholizadeh, Tianda Li, Alan Do-Omri, Peng Lu, Pascal Poupart, Ali Ghodsi
| Challenge: | Knowledge Distillation (KD) is a label regularization technique that can be replaced with lighter teacher-free variants such as the label-smoothing technique. |
| Approach: | They propose to use knowledge distillation to train student models by deploying the teacher network during training. |
| Outcome: | The proposed method can be replaced with lighter teacher-free variants on PLMs with more than 600 distinct trials and ran each configuration five times. |
An Unsupervised Multiple-Task and Multiple-Teacher Model for Cross-lingual Named Entity Recognition (2022.acl-long)
Copied to clipboard
| Challenge: | Existing models for named entity recognition only consider the potential transferability between two identical tasks across both domains. |
| Approach: | They propose to use a similarity metric model to improve cross-lingual named entity recognition task on target domain. |
| Outcome: | Empirical studies on 7 different languages confirm the effectiveness of the proposed model. |
Evaluating Explanations: How Much Do Explanations from the Teacher Aid Students? (2022.tacl-1)
Copied to clipboard
Danish Pruthi, Rachit Bansal, Bhuwan Dhingra, Livio Baldini Soares, Michael Collins, Zachary C. Lipton, Graham Neubig, William W. Cohen
| Challenge: | Existing methods to explain predictions by highlighting salient features are often unstated. |
| Approach: | They propose a framework to quantify the value of explanations via the accuracy gains that they confer on a student model trained to simulate a teacher model. |
| Outcome: | The proposed framework allows principled, automatic, model-agnostic evaluation of attributions. |
Diversifying Neural Dialogue Generation via Negative Distillation (2022.naacl-main)
Copied to clipboard
| Challenge: | Existing approaches to generate generic responses are ignoring low-frequency but generic responses and bringing low- frequency but meaningless responses. |
| Approach: | They propose a negative training paradigm that reminds dialogue models not to generate high-frequency responses during training. |
| Outcome: | The proposed method outperforms previous methods in the generic response problem while minimizing low-frequency but meaningless responses. |
Dynamic Knowledge Distillation for Pre-trained Language Models (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods conduct knowledge distillation statically, e.g., student model aligns output distribution to teacher model on pre-defined training dataset. |
| Approach: | They propose a dynamic knowledge distillation that empowers the student to adjust the learning procedure according to its competency . they find it is promising and provide discussions on potential future directions towards more efficient methods . |
| Outcome: | The proposed method can boost student model performance while accelerating training . the proposed method reduces memory usage and accelerates model inference . |
Distilling Multilingual Transformers into CNNs for Scalable Intent Classification (2022.emnlp-industry)
Copied to clipboard
| Challenge: | Existing multilingual models for voice assistants are limited by their prohibitive inference time and limited performance. |
| Approach: | They propose to distill and deploy multilingual Transformer models for voice assistants using a teacher-student framework that uses teacher-trained models to supervise student model training. |
| Outcome: | The proposed model outperforms a teacher model trained on unlabelled data and achieves equivalent performance. |
Conversational Question Answering with Language Models Generated Reformulations over Knowledge Graph (2024.findings-acl)
Copied to clipboard
| Challenge: | State-of-the-art ConvQA methods struggle with inexplicit question-answer pairs, which can degrade ConvQ performance. |
| Approach: | They propose a reinforcement learning based model, CoRnNet, which utilizes question reformulations generated by large language models to improve ConvQA performance. |
| Outcome: | The proposed model outperforms state-of-the-art ConvQA models by using question reformulations generated by large language models (LLMs). |
Speed Without Sacrifice: Fine-Tuning Language Models with Medusa and Knowledge Distillation in Travel Applications (2025.acl-industry)
Copied to clipboard
Daniel Zagyva, Emmanouil Stergiadis, Laurens Van Der Maas, Aleksandra Dokic, Eran Fainman, Ilya Gusev, Moran Beladev
| Challenge: | Rapid growth of digital applications has intensified the demand for real-time natural language processing (NLP) capabilities. |
| Approach: | They propose a framework that combines Medusa and knowledge distillation to achieve compounded benefits in both model size and inference speed. |
| Outcome: | The proposed framework reduces inference latency by 10-20x while maintaining the student model’s performance quality. |
Skill-Aware Data Selection and Fine-Tuning for Data-Efficient Reasoning Distillation (2026.acl-short)
Copied to clipboard
| Challenge: | Large reasoning models such as DeepSeek-R1 and their distilled variants achieve impressive performance on complex reasoning tasks, yet their costs remain substantial. |
| Approach: | They propose a skill-centric distillation framework that efficiently transfers reasoning ability to weaker models with two components: (1) Skill-based data selection, which prioritizes examples targeting the student model’s weaker skills, and (2) Skillaware fine-tuning, which encourages explicit skill decomposition during problem solving. |
| Outcome: | The proposed framework surpasses baselines on Qwen3-4B and Qwend3-8B and focuses on skills emphasized during training. |
Memorization Inheritance in Sequence-Level Knowledge Distillation for Neural Machine Translation (2025.acl-short)
Copied to clipboard
| Challenge: | Memorization of noisy training data creates unexpected failure modes in neural machine translation models, thus presenting a reliability risk when deploying them in the real world. |
| Approach: | They propose a modification to sequence-level knowledge distillation (SeqKD) that intervenes in SeqKd to reduce memorization and hallucinations. |
| Outcome: | The proposed modification reduces memorization and hallucinations in the student model . |
MixTEA: Semi-supervised Entity Alignment with Mixture Teaching (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods to learn informative entity embeddings are insufficient for semi-supervised entity alignment. |
| Approach: | They propose a semi-supervised method which guides the model learning with an end-to-end mixture teaching of manually labeled mappings and probabilistic pseudo mappings. |
| Outcome: | The proposed method is superior to existing methods on benchmark datasets and further analyses. |
Align-to-Distill: Trainable Attention Alignment for Knowledge Distillation in Neural Machine Translation (2024.lrec-main)
Copied to clipboard
| Challenge: | Existing knowledge distillation approaches to NMT often rely on heuristics when deciding which teacher layers to distill from. |
| Approach: | They propose an approach to align student attention heads with their teacher counterparts by heuristics to solve a feature mapping problem. |
| Outcome: | The proposed strategy shows gains of +3.61 and +0.63 BLEU points for WMT-2022 DeDsb and WMT-2014 EnDe compared to baselines. |
Enhancing Long-Chain Reasoning Distillation through Error-Aware Self-Reflection (2026.findings-acl)
Copied to clipboard
Zhuoyang Wu, Xinze Li, Zhenghao Liu, Yukun Yan, Zhiyuan Liu, Minghe Yu, Cheng Yang, Yu Gu, Ge Yu, Maosong Sun
| Challenge: | Existing studies treat SLMs as student models and use long-form Chains-of-Thought (CoTs) as supervision signals for Supervised Fine-Tuning (SFT). Existing research focuses on distilling reasoning ability from LLMs to enhance the mathematical reasoning performance of small-scale models. |
| Approach: | They propose a framework that refines teacher CoTs through an error-aware reflection process to enable the student model to construct more tailored teacher Cots. |
| Outcome: | Experiments on multiple mathematical reasoning benchmarks show that ORION improves performance by more than 2% over all baselines. |
Accurate Knowledge Distillation via n-best Reranking (2024.naacl-long)
Copied to clipboard
| Challenge: | Existing studies using sequencelevel knowledge distillation (KD) have adopted this approach. |
| Approach: | They propose to utilize n-best reranking to enhance Sequence-Level Knowledge Distillation by utilizing a diverse set of models with different inductive biases, objective functions or architectures to pick the highest-quality hypotheses as labels. |
| Outcome: | The proposed approach is validated on the WMT’21 German English and Chinese english translation tasks. |
BiLD: Bi-directional Logits Difference Loss for Large Language Model Distillation (2025.coling-main)
Copied to clipboard
| Challenge: | Knowledge distillation (KD) is a method for reducing model size while preserving performance. |
| Approach: | They propose a method to distill large language models at the logit level by transferring knowledge from a large teacher model to a smaller student model. |
| Outcome: | The proposed method outperforms supervised fine-tuning, vanilla KL loss and five other distillation methods on 13 datasets. |
Teacher-Student Training for Debiasing: General Permutation Debiasing for Large Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have demonstrated impressive zero-shot capabilities and versatility in NLP tasks, however they often fail to maintain crucial invariances for specific tasks. |
| Approach: | They propose to distill the capabilities of a computationally intensive, debiased, teacher model into a more compact student model. |
| Outcome: | The proposed model outperforms larger, biased teacher models by reducing the invariances of the student model by incorporating encoders. |
RAIL-KD: RAndom Intermediate Layer Mapping for Knowledge Distillation (2022.findings-naacl)
Copied to clipboard
Md Akmal Haidar, Nithin Anchuri, Mehdi Rezagholizadeh, Abbas Ghaddar, Philippe Langlais, Pascal Poupart
| Challenge: | Existing methods for intermediate layer knowledge distillation suffer from computational burdens and engineering efforts for setting up a proper layer mapping. |
| Approach: | They propose a method where intermediate layers from teacher and student models are randomly selected to be distilled into intermediate layers of student models. |
| Outcome: | The proposed method outperforms state-of-the-art intermediate layer knowledge distillation methods on GLUE tasks and out-of domain test sets. |
MGCL: Multi-Granularity Clue Learning for Emotion-Cause Pair Extraction via Cross-Grained Knowledge Distillation (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Traditional methods often rely on coarse-grained clause-level annotations, which overlook valuable fine-grain clues. |
| Approach: | They propose a method that captures fine-grained clues from a weakly-supervised perspective efficiently by using a teacher model to give sub-clause clues without needing fine-grain annotations. |
| Outcome: | The proposed method achieves state-of-the-art performance while offering improved interpretability. |
Learning to Solve NLP Tasks in an Incremental Number of Languages (2021.acl-short)
Copied to clipboard
| Challenge: | Existing models that support multiple languages are expensive and time-consuming to retrain . retraining only with the new language data will result in Catastrophic Forgetting of previously acquired knowledge. |
| Approach: | They propose a Continual Learning strategy that updates a model to support new languages over time while maintaining consistent results on previously learned languages. |
| Outcome: | The proposed model is based on an existing model that "teaches" to a student while the student is also trained on a new language. |
Combining Curriculum Learning and Knowledge Distillation for Dialogue Generation (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Existing studies have shown that curriculum learning facilitates dialogue generation tasks while knowledge distillation can yield significant performance boosts for student models. |
| Approach: | They propose a combination of curriculum learning and knowledge distillation for dialogue generation models . they cluster training cases according to their complexity and employ an adversarial training strategy . |
| Outcome: | The proposed model improves compared with baselines. |
Learning Shortcut Models for Efficient Recursive Reasoning (2026.acl-srw)
Copied to clipboard
| Challenge: | Recent research shows that Transformer-style models can be made more efficient by sharing parameters over blocks. |
| Approach: | They propose a framework for distilling latent reasoning into a multiscale jump model that enables flexible test-time compute. |
| Outcome: | Experiments on ARC-AGI show that the proposed model achieves competitive accuracy compared to recursive baselines while requiring fewer sequential updates. |
GOVERN: Gradient Orientation Vote Ensemble for Multi-Teacher Reinforced Distillation (2024.emnlp-industry)
Copied to clipboard
| Challenge: | Pre-trained language models have achieved remarkable performance in OpenQA, but for practical deployment, knowledge distillation is crucial to maintain high performance while operating under computational constraints. |
| Approach: | They propose an algorithm to perform unsupervised knowledge distillation without the guidance of labels to achieve 99.5% of performance. |
| Outcome: | The proposed algorithm achieves 99.5% of performance in a commercial question-answering system. |
TrueTeacher: Learning Factual Consistency Evaluation with Large Language Models (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for evaluating factual consistency are limited in their effectiveness. |
| Approach: | They propose a method for generating synthetic data by annotating diverse model-generated summaries using a LLM. |
| Outcome: | The proposed method outperforms state-of-the-art models and the LLM teacher on TRUE benchmarks. |
Distilling Robustness into Natural Language Inference Models with Domain-Targeted Augmentation (2024.findings-acl)
Copied to clipboard
| Challenge: | Knowledge distillation optimises a smaller student model to behave similarly to a larger teacher model, retaining some performance benefits. |
| Approach: | They propose to augment the distillation with generated unlabelled examples that match the target distribution and upsamples data points among the training set that are similar to the target. |
| Outcome: | The proposed method outperforms previous robustness solutions on the task of natural language inference (NLI) it also improves performance on OOD domains even beyond the target domain. |
Overcoming Catastrophic Forgetting beyond Continual Learning: Balanced Training for Neural Machine Translation (2022.acl-long)
Copied to clipboard
| Challenge: | Neural networks tend to gradually forget the previously learned knowledge when learning multiple tasks sequentially from dynamic data distributions. |
| Approach: | They propose a method that iteratively provides complementary knowledge to student models by dynamically updating teacher models trained on specific data orders. |
| Outcome: | The proposed method improves on multiple machine translation tasks and improves performance over baseline systems. |
Symbolic Chain-of-Thought Distillation: Small Models Can Also “Think” Step-by-Step (2023.acl-long)
Copied to clipboard
| Challenge: | Symbolic Chain-of-thought Distillation (SCoTD) is a method to train a smaller student model on rationalizations sampled from a significantly larger teacher model. |
| Approach: | They propose a method to train a smaller student model on rationalizations from a larger teacher model. |
| Outcome: | The proposed method improves the performance of a student model in supervised and few-shot settings and especially for challenge sets. |
Teaching Small Language Models to Reason (2023.acl-short)
Copied to clipboard
| Challenge: | Chain of thought prompting improves reasoning capabilities of large language models but only emerges in models with tens of billions of parameters. |
| Approach: | They propose to fine tune a student model on chain of thought outputs generated by a larger teacher model. |
| Outcome: | The proposed method improves task performance across arithmetic, commonsense and symbolic reasoning datasets. |
Towards Developing a Multilingual and Code-Mixed Visual Question Answering System by Knowledge Distillation (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for training language-vision models only consider monolingual learning, especially English. |
| Approach: | They propose to extend an English language-vision model into a multilingual and code-mixed model by using knowledge distillation techniques. |
| Outcome: | The proposed model outperforms existing models on multilingual and code-mixed VQA datasets on eleven languages. |
Multi-stage Distillation Framework for Cross-Lingual Semantic Similarity Matching (2022.findings-naacl)
Copied to clipboard
Kunbo Ding, Weijie Liu, Yuejian Fang, Zhe Zhao, Qi Ju, Xuefeng Yang, Rong Tian, Zhu Tao, Haoyan Liu, Han Guo, Xingyu Bai, Weiquan Mao, Yudong Li, Weigang Guo, Taiqiang Wu, Ningyuan Sun
| Challenge: | Existing studies have shown that cross-lingual knowledge distillation can improve the performance of pre-trained models for cross-linguistic similarity matching tasks. |
| Approach: | They propose a multi-stage distillation framework for constructing a small-size but high-performance cross-lingual model using contrastive learning, bottleneck, and parameter recurrent strategies. |
| Outcome: | The proposed model can compress the size of XLM-R and MiniLM by more than 50% while the performance is only reduced by about 1%. |
Enhancing Knowledge Distillation of Large Language Models through Efficient Multi-Modal Distribution Alignment (2025.coling-main)
Copied to clipboard
| Challenge: | Existing knowledge distillation techniques for large language models are causing difficulties for student models to learn multi-modal probability distributions. |
| Approach: | They propose a ranking loss-based knowledge distillation method that encourages consistency of the ranking of peak predictions between teacher and student models. |
| Outcome: | The proposed method improves student models' ability to learn multi-modal distributions. |
Pre-training Distillation for Large Language Models: A Design Space Exploration (2025.acl-long)
Copied to clipboard
| Challenge: | Knowledge distillation (KD) aims to transfer knowledge from a large teacher model to a smaller student model for model compression. |
| Approach: | They extend knowledge distillation to the pre-training phase of large language models . they first conduct an experiment using a teacher LLM to distill a 1.9B student LLM . |
| Outcome: | The proposed model can be used to distill a 1.9B student model using a teacher LLM. |
MiniLMv2: Multi-Head Self-Attention Relation Distillation for Compressing Pretrained Transformers (2021.findings-acl)
Copied to clipboard
| Challenge: | Existing work on deep self-attention distillation for natural language processing tasks is limited by computational resources and latency. |
| Approach: | They generalize deep self-attention distillation in MINILM by using only self- attention relation distillation for taskagnostic compression of pretrained Transformers. |
| Outcome: | The proposed model outperforms the state-of-the-art in a multilingual and multilingual teacher model. |
Lion: Adversarial Distillation of Proprietary Large Language Models (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for transferring knowledge from a proprietary large language model to an open-source LLM have been criticized for not incorporating any feedback to boost student model's proficiency iteratively. |
| Approach: | They propose a novel adversarial distillation framework that integrates feedback to identify challenging instructions where the student model’s performance falls short. |
| Outcome: | The proposed framework surpasses existing state-of-the-art LLMs by 55.4% in challenging zero-shot reasoning benchmarks and 16.7% on AGIEval. |
DRP: Distilled Reasoning Pruning with Mathematical Skill-aware Step Decomposition for Efficient Large Reasoning Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing solutions to this problem are inference-time pruning and tuning-based distillation. |
| Approach: | They propose a framework that combines inference-time pruning with tuning-based distillation to enable efficient and accurate reasoning. |
| Outcome: | The proposed framework reduces token usage while improving accuracy on GSM8K and AIME tokens while avoiding performance drop. |
GKT: A Novel Guidance-Based Knowledge Transfer Framework For Efficient Cloud-edge Collaboration LLM Deployment (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing methods of acceleration require fine-tuning of considerably large models, such as Llama-7B, posing a challenge for average users. |
| Approach: | They propose a Guidance-based Knowledge Transfer framework that leverages a larger LLM as a 'teacher' and a smaller 'student' model to finalize responses. |
| Outcome: | The proposed framework achieves a maximum accuracy improvement of 14.18%, along with a 10.72 times speed-up on GSM8K and an accuracy improvement 14.00% along with 7.73 times speed up in CSQA. |
Annealing Knowledge Distillation (2021.eacl-main)
Copied to clipboard
| Challenge: | Knowledge distillation (KD) is a powerful model compression technique for deep neural networks. |
| Approach: | They propose a method to feed the rich information provided by teacher’s soft-targets incrementally and more efficiently by annealing the teacher output incrementally. |
| Outcome: | The proposed method can be used on image classification and NLP language inference tasks with BERT-based models on the GLUE benchmark. |
Performance-Guided LLM Knowledge Distillation for Efficient Text Classification at Scale (2024.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) face high computational demands at inference time due to high computational costs. |
| Approach: | They propose a cost-effective and high-throughput solution for large language models . PGKD distills the knowledge of LLMs into smaller, task-specific models based on teacher-student knowledge distillation . |
| Outcome: | PGKD outperforms BERT-based models and other knowledge distillation methods on multi-class classification datasets. |
Attention-Guided Answer Distillation for Machine Reading Comprehension (D18-1)
Copied to clipboard
| Challenge: | Existing approaches to reading comprehension systems are vulnerable to adversarial attacks. |
| Approach: | They propose to use knowledge distillation to transfer knowledge from an ensemble to a single model. |
| Outcome: | The proposed methods outperform the teacher on adversarial datasets and NarrativeQA benchmarks. |
PseudoReasoner: Leveraging Pseudo Labels for Commonsense Knowledge Base Population (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Commonsense Knowledge Base (CSKB) Population aims at reasoning over unseen entities and assertions on CSKBs, but it requires out-of-domain generalization ability as the source CSMB for training is of a relatively smaller scale (1M) . |
| Approach: | They propose a semi-supervised learning framework that uses a teacher model to provide pseudo labels on the unlabeled candidate dataset for a student model to learn from. |
| Outcome: | The proposed framework can improve the backbone model KG-BERT (RoBERTa-large) by 3.3 points on the overall performance and especially, 5.3 points on out-of-domain performance. |
Quantification of Large Language Model Distillation (2025.acl-long)
Copied to clipboard
Sunbowen Lee, Junting Zhou, Chang Ao, Kaige Li, Xeron Du, Sirui He, Haihong Wu, Tianci Liu, Jiaheng Liu, Hamid Alinejad-Rokny, Min Yang, Yitao Liang, Zhoufutu Wen, Shiwen Ni
| Challenge: | Existing studies have revealed the robustness degra-dation caused by data distillation. |
| Approach: | They propose a framework to evaluate and quantify model distillation . they aim to identify identity cognition contradictions and analyse multi-granularity response similarities across models to measure the extent of homogenization. |
| Outcome: | The proposed framework addresses two key aspects: (1) Identifying identity cognition contradictions to assess discrepancies in how models perceive and represent identity-related information; (2) Analyzing multi-granularity response similarities across models to measure the extent of homogenization. |
Improving Reasoning Capabilities in Small Models through Mixture-of-layers Distillation with Stepwise Attention on Key Information (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods focus on transferring teacher-generated rationales to student models, but do not explore teachers’ dynamic attention towards critical information during reasoning. |
| Approach: | They propose a method that transfers the teacher’s stepwise attention on key information to the student model and a Mixture of Layers module that allows dynamic alignment between the teacher and student. |
| Outcome: | The proposed framework achieves consistent performance improvements across multiple mathematical and commonsense reasoning datasets. |
MergeDistill: Merging Language Models using Pre-trained Distillation (2021.findings-acl)
Copied to clipboard
| Challenge: | Existing pre-trained multilingual language models often lack capacity and skewed data . this leads to inequitable representation of languages due to limited capacity and sub-optimal vocabularies. |
| Approach: | They propose a framework to merge pre-trained multilingual language models to maximize their assets with minimal dependencies. |
| Outcome: | The proposed framework outperforms teacher-trained models on multiple datasets and with a fixed model capacity. |
Unlearning Backdoor Attacks for LLMs with Weak-to-Strong Knowledge Distillation (2025.findings-acl)
Copied to clipboard
| Challenge: | Parameter-efficient fine-tuning (PEFT) can bridge the gap between large language models and downstream tasks, but is vulnerable to malicious attacks. |
| Approach: | They propose a weak-to-strong unlearning algorithm based on feature alignment knowledge distillation to defend against backdoor attacks . they first train a small-scale language model through full-parameter fine-tuning to serve as the clean teacher model and then guide the large-scale poisoned student model in unlearning the backdoor. |
| Outcome: | The proposed method can unlearn backdoor features without compromising model performance. |
ELAD: Explanation-Guided Large Language Models Active Distillation (2024.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are hindered by their memory inefficiency, computational demands, and the high costs of API inferences. |
| Approach: | They propose an Explanation-Guided LLMs Active Distillation framework that employs an active learning strategy to optimize the balance between annotation costs and model performance. |
| Outcome: | The proposed framework significantly improves the efficiency of LLMs knowledge distillation. |
RW-KD: Sample-wise Loss Terms Re-Weighting for Knowledge Distillation (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Knowledge Distillation (KD) is used to compress the pre-training and task-specific fine-tuning phases of large neural language models. |
| Approach: | They propose a sample-wise loss weighting method that re-weights the two losses for each sample. |
| Outcome: | The proposed method outperforms existing methods on 7 datasets of the GLUE benchmark. |
NewsBERT: Distilling Pre-trained Language Model for Intelligent News Application (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Existing language models are pre-trained and distilled on general corpus like Wikipedia, which has gaps with the news domain and may be suboptimal for news intelligence. |
| Approach: | They propose a method to distill existing language models on Wikipedia to enable efficient news intelligence. |
| Outcome: | The proposed model can be used to build and test a news intelligence application on Wikipedia and Wikipedia. |
Class-Incremental Few-Shot Event Detection (2024.lrec-main)
Copied to clipboard
| Challenge: | Existing methods to deal with new class of events with only a few labeled instances are challenging . old knowledge forgetting and new class overfitting are two problems in this task. |
| Approach: | They propose a task called class-incremental few-shot event detection to solve old knowledge forgetting and new class overfitting problems. |
| Outcome: | The proposed method reduces old knowledge forgetting and new class overfitting problems on two benchmark datasets. |
Collective Wisdom: Improving Low-resource Neural Machine Translation using Adaptive Knowledge Distillation (2020.coling-main)
Copied to clipboard
| Challenge: | Existing approaches to train high-quality NMT models in bilingually low-resource scenarios are limited by the scarcity of parallel sentence-pairs. |
| Approach: | They propose to distill the knowledge of teacher models to a single student model by using knowledge distillation. |
| Outcome: | The proposed approach achieves up to +0.9 BLEU score improvements compared to strong baselines. |
TinyThinker: Distilling Reasoning through Coarse-to-Fine Knowledge Internalization with Self-Reflection (2025.naacl-long)
Copied to clipboard
| Challenge: | Large Language Models exhibit impressive reasoning capabilities across diverse tasks . direct training on synthesized reasoning data may lead to superficial imitation of reasoning process, authors argue . |
| Approach: | They propose a framework that introduces a three-stage process that incrementally guides the student model through the reasoning process, progressively refining knowledge from coarse to fine granularity. |
| Outcome: | The proposed framework achieves superior performance on commonsense reasoning benchmarks and can be extended to other knowledge-intensive reasoning tasks. |
Dual-teacher Knowledge Distillation for Low-frequency Word Translation (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Neural machine translation models are trained on parallel corpora with unbalanced word frequency distribution, resulting in high-frequency words being ignored. |
| Approach: | They propose to employ a low-frequency teacher model that excels in translating low- frequency words to guide the learning of the student model. |
| Outcome: | The proposed method achieves +0.64 BLEU improvements over the state-of-the-art method on the low-frequency translation task while maintaining the translation quality of high-frequency words. |
Causal Distillation for Language Models (2022.naacl-main)
Copied to clipboard
Zhengxuan Wu, Atticus Geiger, Joshua Rozner, Elisa Kreiss, Hanson Lu, Thomas Icard, Christopher Potts, Noah Goodman
| Challenge: | Distillation efforts have led to language models that are more compact and efficient without serious drops in performance. |
| Approach: | They propose to augment distillation with a third objective that encourages the student model to imitate the causal dynamics of the teacher through a distillation interchange intervention training objective (DIITO). |
| Outcome: | The proposed method lowers perplexity on the WikiText-103M corpus and improves on the GLUE benchmark, SQuAD, and CoNLL-2003. |
Wider & Closer: Mixture of Short-channel Distillers for Zero-shot Cross-lingual Named Entity Recognition (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing mainstream methods for zero-shot cross-lingual named entity recognition ignore the rich and complementary information lying in the intermediate layers of pre-trained language models and domain-invariant information is easily lost during transfer. |
| Approach: | They propose a mixture of short-channel distillers to fully interact the rich hierarchical information in the teacher model and to transfer knowledge to the student model sufficiently and efficiently. |
| Outcome: | The proposed method shows great generalization and compatibility across languages and fields. |
MiniELM: A Lightweight and Adaptive Query Rewriting Framework for E-Commerce Search Optimization (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for rewriting query terms struggle with natural language understanding . generative methods face high inference latency and cost in offline settings . |
| Approach: | They propose a hybrid pipeline for rewriting query queries using offline knowledge distillation and online reinforcement learning. |
| Outcome: | The proposed pipeline improves query relevance, diversity, adaptability and cost-effective evaluation without manual annotations on Amazon ESCI dataset. |
Learn What Is Possible, Then Choose What Is Best: Disentangling One-To-Many Relations in Language Through Text-based Games (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Language models pre-trained on large self-supervised corpora, followed by task-specific fine-tuning has become the dominant paradigm in NLP. |
| Approach: | They propose to train language models pre-trained on large self-supervised corpora, followed by task-specific fine-tuning on the target domain. |
| Outcome: | The proposed model improves on the previous state-of-the-art model on the Jericho Walkthroughs dataset by 49%. |
Tiny-NewsRec: Effective and Efficient PLM-based News Recommendation (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing work fine tunes the PLM with the news recommendation task, which can cause a domain shift problem. |
| Approach: | They propose a self-supervised method to adapt general PLM to news domain with a contrastive matching task between news titles and news bodies. |
| Outcome: | The proposed method can improve both the effectiveness and efficiency of the large PLM-based news recommendation model while maintaining its performance. |
Beyond the Mode: Sequence-Level Distillation of Multilingual Translation Models for Low-Resource Language Pairs (2025.findings-naacl)
Copied to clipboard
Aarón Galiano-Jiménez, Juan Antonio Pérez-Ortiz, Felipe Sánchez-Martínez, Víctor M. Sánchez-Cartagena
| Challenge: | Existing multilingual pre-trained models for low-resource languages have outperformed those trained from scratch for low resources due to high hardware requirements. |
| Approach: | They propose to use beam search to decode the whole output distribution of the teacher to improve student learning. |
| Outcome: | The proposed methods improve student model performance and reduce gender bias amplification common to beam search based methods. |
Teach Small Models to Reason by Curriculum Distillation (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large Reasoning Models (LRMs) show strong System-2-style reasoning, but at the cost of significant computational overhead. |
| Approach: | They propose a two-stage curriculum distillation framework which builds a robust internal problem-solving student model and then teaches the student model to externalize this knowledge as explicit reasoning. |
| Outcome: | The proposed model outperforms single-stage baselines on mathematical benchmarks and significantly outperformed LRMs on complex tasks. |
Enhancing Chinese Word Segmentation via Pseudo Labels for Practicability (2021.findings-acl)
Copied to clipboard
| Challenge: | Pre-trained language models alleviate segmentation ambiguity and out-of-vocabulary (OOV) words. |
| Approach: | They propose a semisupervised neural method which distills knowledge from unlabeled data to a student model to improve both in-domain and out-of-domain CWS. |
| Outcome: | The proposed method can keep practicability of the lightweight student model and improve segmentation effectively on downstream Chinese NLP tasks. |
One Teacher is Enough? Pre-trained Language Model Distillation from Multiple Teachers (2021.findings-acl)
Copied to clipboard
| Challenge: | Pre-trained language models (PLMs) have huge model sizes and computational complexity, making it difficult to deploy them to low-latency and high-concurrence online systems. |
| Approach: | They propose a multi-teacher knowledge distillation framework for pre-trained language model compression that can train high-quality student model from multiple teacher PLMs. |
| Outcome: | The proposed framework can train high-quality student model from multiple teacher PLMs with shared pooling and prediction layers to align output space for better collaborative teaching. |
Multi-Level Knowledge Distillation for Out-of-Distribution Detection in Text (2023.acl-long)
Copied to clipboard
| Challenge: | Self-supervised representation learning has proved to be a valuable component for out-of-distribution (OoD) detection with only the texts of in-difference (ID) examples. |
| Approach: | They propose a method that integrates strengths and weaknesses of both methods . they use a fine-tuned model as the teacher to teach a randomly initialized student model . |
| Outcome: | The proposed method outperforms human evaluators in the pair-expert task on the Human ChatGPT Comparison Corpus. |
Learning from Imperfect Data: Towards Efficient Knowledge Distillation of Autoregressive Language Models for Text-to-SQL (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing text-to-SQL LLMs are computationally expensive and difficult to deploy in real-world applications. |
| Approach: | They propose to distill a larger teacher model into a smaller student model by using imperfect data to improve the KD. |
| Outcome: | The proposed method achieves the best tradeoff between performance and efficiency on 5 text-to-SQL benchmarks. |
ToDi: Token-wise Distillation via Fine-Grained Divergence Control (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) offer impressive performance but are impractical for resource-constrained deployment due to high latency and energy consumption. |
| Approach: | They propose a method that adaptively combines FKL and RKL per token using a sigmoid-based weighting function derived from the teacher-student probability log-ratio. |
| Outcome: | The proposed method outperforms baselines using uniform or less granular strategies across instruction-following benchmarks. |
Towards Efficient CoT Distillation: Self-Guided Rationale Selector for Better Performance with Fewer Rationales (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing work on rationale quality underestimates the importance of CoT distillation, focusing primarily on data quantity, which may result in transferring noisy or incorrect information to the student model. |
| Approach: | They propose a method which can discern and select high quality rationales for distillation and a Rationale Difficulty metric to measure the ability of the student model to generate the correct answer under a given rationale. |
| Outcome: | The proposed method achieves 4.6% accuracy improvement over baseline data on seven datasets over three tasks, controlling accuracy, diversity, and difficulty. |
Personalized Distillation: Empowering Open-Sourced LLMs with Adaptive Learning for Code Generation (2023.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies have shown that close-sourced LLMs lack the ability to integrate into real-world applications due to their high associated costs and ethical concerns. |
| Approach: | They propose to use student model to refine its own solution by querying ChatGPT to generate task instruction and solution pairs and querying data to refine model. |
| Outcome: | The proposed model outperforms standard distillation with only one third of the data. |
KNOT: Knowledge Distillation Using Optimal Transport for Solving NLP Tasks (2022.coling-1)
Copied to clipboard
| Challenge: | Knowledge Distillation using Optimal Transport (KNOT) aims to distill the natural language semantic knowledge from multiple teacher networks to a student network. |
| Approach: | They propose to distill natural language semantic knowledge from multiple teacher networks to a student network by learning to minimize the optimal transport cost of its assigned probability distribution over the labels to the weighted sum of probabilities predicted by the (local) teacher models. |
| Outcome: | The proposed method shows improvements in the global model’s SD performance over the baseline across three NLP tasks while performing on par with Entropy-based distillation on standard accuracy and F1 metrics. |
Reverse Thinking Makes LLMs Stronger Reasoners (2025.naacl-long)
Copied to clipboard
Justin Chen, Zifeng Wang, Hamid Palangi, Rujun Han, Sayna Ebrahimi, Long Le, Vincent Perot, Swaroop Mishra, Mohit Bansal, Chen-Yu Lee, Tomas Pfister
| Challenge: | Reverse-Enhanced Thinking (RevThink) is a framework for large language models to perform reverse thinking. |
| Approach: | They propose a framework for enhancing forward-backward reasoning by collecting data from a teacher model and employing three objectives to train a student model in a multi-task learning fashion. |
| Outcome: | The proposed framework outperforms a fine-tuning method trained on 10x more forward reasoning on 12 datasets covering commonsense, math, and logical reasoning. |
Patient Knowledge Distillation for BERT Model Compression (D19-1)
Copied to clipboard
| Challenge: | Pre-trained language models such as BERT have proven to be highly effective for natural language processing tasks, but the high demand for computing resources hinders their application in practice. |
| Approach: | They propose to compress an original large model (teacher) into an equally-effective lightweight shallow network (student) Empirically, this translates into improved results on multiple NLP tasks with a significant gain in training efficiency, without sacrificing model accuracy. |
| Outcome: | The proposed model reduces the computational cost of training models using the teacher model into a lightweight shallow network. |
MIDAS: Multi-level Intent, Domain, And Slot Knowledge Distillation for Multi-turn NLU (2025.findings-naacl)
Copied to clipboard
| Challenge: | Existing Large Language Models (LLMs) can generate coherent text, but they struggle to recognise user intent behind queries. |
| Approach: | They propose a novel approach leveraging multi-level intent, domain, and slot knowledge distillation for multi-turn NLU. |
| Outcome: | The proposed model improves multi-turn conversation understanding by integrating teacher teachers into a student model. |
Understanding and Improving Knowledge Distillation for Quantization Aware Training of Large Transformer Encoders (2022.emnlp-main)
Copied to clipboard
| Challenge: | Knowledge distillation (KD) has been used for quantization-aware training to improve the ability of a lightweight model with the transferred knowledge from the teacher. |
| Approach: | They propose two methods to improve attention recovery of quantized large Transformers by combining attention-map and attention-output losses. |
| Outcome: | The proposed methods achieve state-of-the-art accuracy for quantized large Transformer encoder models with sub-2-bit weight quantization. |
Staged Knowledge Distillation Through Least-to-Most Prompting: Optimizing Teacher Guidance via Difficulty-Aware Training (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Knowledge distillation (KD) enables the compression of large language models (LLMs) conventional methods suffer from training-inference mismatches and suboptimal performance due to expensive student-generated outputs. |
| Approach: | They propose a method that combines a CL strategy and adaptive loss design to reduce training mismatches and suboptimal performance. |
| Outcome: | L2M-KD outperforms existing white-box KD methods on instruction-following tasks . it outperformed existing methods, achieving superior student model performance with reduced overhead . |
Multi-Grained Knowledge Distillation for Named Entity Recognition (2021.naacl-main)
Copied to clipboard
| Challenge: | Pre-trained big models have delivered top performance in Seq2seq modeling, but their deployments in real-world applications are often hindered by excessive computations and memory demands. |
| Approach: | They propose a distillation scheme to efficiently transfer knowledge from big models to their cheaper counterparts. |
| Outcome: | The proposed scheme maximizes the assimilation of knowledge from the teacher model to the student model. |
Improved Knowledge Distillation for Pre-trained Language Models via Knowledge Selection (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Existing studies on knowledge distillation have shown that not all knowledge is necessary for learning a good student model. |
| Approach: | They propose an actor-critic approach to selecting appropriate knowledge to transfer during the process of knowledge distillation. |
| Outcome: | The proposed method outperforms several strong knowledge distillation baselines significantly on the GLUE datasets. |
Retrieved In-Context Principles from Previous Mistakes (2024.emnlp-main)
Copied to clipboard
| Challenge: | Recent advances in in-context learning (ICL) have limited customization and inadequate error coverage. |
| Approach: | They propose a method to retrieve in-context principles from mistakes to improve model performance. |
| Outcome: | The proposed framework enhances model performance when applied to various prompting strategies. |
Efficient End-to-End Visual Document Understanding with Rationale Distillation (2024.naacl-long)
Copied to clipboard
| Challenge: | Pre-processing tools such as optical character recognition (OCR) can map document image inputs to textual tokens, then large language models (LLMs) can reason over text. |
| Approach: | They propose a method that integrates outputs of OCR tools and larger multimodal models as intermediate "rationales" a student model is trained to predict rationales and answers based on visual documents . |
| Outcome: | The proposed model outperforms the base model on three visual document understanding benchmarks with only 1% higher computational cost. |
Enhancing Abstractiveness of Summarization Models through Calibrated Distillation (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods to generate abstractive summarizations are slow and abstractive, but we propose a novel approach to enhance the level of abstractiveness without sacrificing the informativeness of generated summaries. |
| Approach: | They propose a novel approach to enhance the level of abstractiveness without sacrificing the informativeness of generated summaries by exposing diverse pseudo summary with two supervision to the student model. |
| Outcome: | The proposed method outperforms previous methods in abstractive summarization distillation, producing highly abstractive and informative summaries. |
AD-KD: Attribution-Driven Knowledge Distillation for Language Model Compression (2023.acl-long)
Copied to clipboard
| Challenge: | Existing knowledge distillation methods focus on the transfer of model-specific knowledge but overlook data-specific information. |
| Approach: | They propose an attribution-driven knowledge distillation approach which explores the token-level rationale behind the teacher model and transfers attribution knowledge to the student model. |
| Outcome: | The proposed method outperforms state-of-the-art methods on the GLUE benchmark and shows that it is more efficient than existing methods. |
Improving Stance Detection with Multi-Dataset Learning and Knowledge Distillation (2021.emnlp-main)
Copied to clipboard
| Challenge: | stance detection is a method to determine whether a text author is in favor of, against or neutral toward a specific target. |
| Approach: | They propose a method that applies instance-specific temperature scaling to the teacher and student predictions. |
| Outcome: | The proposed method outperforms the state-of-the-art on all datasets and on multiple datasets. |
Enhancing Task-Specific Distillation in Small Data Regimes through Language Generation (2022.coling-1)
Copied to clipboard
| Challenge: | Large-scale pretrained language models have led to significant improvements in Natural Language Processing, but they come at the cost of high computational and storage requirements. |
| Approach: | They propose to distill knowledge from larger models to smaller ones through pseudo-labels on task-specific datasets. |
| Outcome: | The proposed approach improves on the SST-2, MRPC, YELP-2, and TREC-6 datasets. |
Towards Better Entity Linking with Multi-View Enhanced Distillation (2023.acl-long)
Copied to clipboard
Yi Liu, Yuan Tian, Jianxun Lian, Xinlong Wang, Yanan Cao, Fang Fang, Wen Zhang, Haizhen Huang, Weiwei Deng, Qi Zhang
| Challenge: | Entity linking is a fundamental task in Natural Language Processing (NLP), connecting mentions within unstructured contexts to their corresponding entities in a Knowledge Base (KB). |
| Approach: | They propose a dual-encoder framework that can efficiently match mentions to two-encoding frameworks by a global-view. |
| Outcome: | The proposed framework achieves state-of-the-art on several entity linking benchmarks. |
Single-/Multi-Source Cross-Lingual NER via Teacher-Student Learning on Unlabeled Data in Target Language (2020.acl-main)
Copied to clipboard
| Challenge: | Existing approaches to named entity recognition (NER) are limited by label projection with pairwise texts or direct model transfer. |
| Approach: | They propose a method where NER models in the source languages are used as teachers to train a student model on unlabeled data in the target language. |
| Outcome: | The proposed method outperforms existing state-of-the-art methods for single-source and multi-source cross-lingual NER on target languages. |
Revisiting Knowledge Distillation for Autoregressive Language Models (2024.acl-long)
Copied to clipboard
| Challenge: | Autoregressive language models (LMs) are expensive and memory intensive, preventing the development of industrial applications. |
| Approach: | They propose an adaptive teaching approach to improve the KD of autoregressive language models by distilling knowledge into a small student model. |
| Outcome: | The proposed method can achieve consistent and significant performance gains across all model types and sizes. |
Distilling the Essence: Efficient Reasoning Distillation via Sequence Truncation (2026.findings-acl)
Copied to clipboard
Wei-Rui Chen, Vignesh Kothapalli, Ata Fatahibaarzi, Hejian Sang, Shao Tang, Qingquan Song, Zhipeng Wang, Muhammad Abdul-Mageed
| Challenge: | a large reasoning model (LRM) training on large amounts of reasoning data is computationally expensive. |
| Approach: | They propose a method to quantify computation-quality tradeoffs as a function of sequence length. |
| Outcome: | The proposed method reduces training time, memory and FLOPs by 50% on long training sequences while retaining the full-sequence performance. |
Evolving Knowledge Distillation with Large Language Models and Active Learning (2024.lrec-main)
Copied to clipboard
| Challenge: | Existing studies have focused on the direct use of large language models for text generation and labeling, without fully exploring their potential to comprehend the target task and acquire valuable knowledge. |
| Approach: | They propose to distill the knowledge of large language models into smaller models by generating annotated data. |
| Outcome: | The proposed method improves the performance of small domain models while enhancing the ability of large language models. |
Accelerating Multiple Intent Detection and Slot Filling via Targeted Knowledge Distillation (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing non-autoregressive Spoken Language Understanding models suffer from multi-modality problem . current methods have little prior knowledge about the reference during inference . |
| Approach: | They propose a Targeted Knowledge Distillation Framework (TKDF) for multi-intent SLU that utilizes the knowledge distillation method to improve the performance. |
| Outcome: | The proposed model outperforms existing models on two public multi-intent datasets while speeding up by over 4.5 times. |
Distilled Dual-Encoder Model for Vision-Language Understanding (2022.emnlp-main)
Copied to clipboard
| Challenge: | Experimental results show that the proposed cross-modal attention distillation is crucial to the success of our framework. |
| Approach: | They propose a framework that distills knowledge of fusion-encoder teacher into dual-encoding student model. |
| Outcome: | The proposed model is competitive with the fusion-encoder teacher model in performance, but suffers from a lack of deep cross-modal interactions. |
One-Teacher and Multiple-Student Knowledge Distillation on Sentiment Classification (2022.coling-1)
Copied to clipboard
| Challenge: | Existing knowledge distillation models require large computing resources and long inference time to perform. |
| Approach: | They propose a one-teacher and multiple-student knowledge distillation approach to distill a deep pre-trained teacher model into multiple shallow student models with ensemble learning. |
| Outcome: | The proposed method achieves better results with fewer parameters and extremely high speedup ratios on three sentiment classification tasks. |
Calibrating Student Models for Emotion-related Tasks (2022.emnlp-main)
Copied to clipboard
| Challenge: | Knowledge distillation is an effective method to transfer knowledge from one network (a.k.a. teacher) to another (as student). |
| Approach: | They propose to use a mixup data augmentation technique to increase the accuracy of the model by providing better training signals to the student models. |
| Outcome: | The proposed method improves the calibration of student models while providing better training signals to the student models using training dynamics. |
Hard Gate Knowledge Distillation - Leverage Calibration for Robust and Reliable Language Model (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing knowledge distillation schemes focus on a teacher as a source of knowledge and a gauge to detect miscalibration of a student. |
| Approach: | They propose a method that uses a teacher model as a source of knowledge and a model as an error detector to detect miscalibration of a student. |
| Outcome: | The proposed scheme improves model generalization and significantly lowers calibration error. |
Find Your Optimal Teacher: Personalized Data Synthesis via Router-Guided Multi-Teacher Distillation (2026.acl-long)
Copied to clipboard
Hengyuan Zhang, Shiping Yang, Xiao Liang, Chenming Shang, Yuxuan Jiang, Chaofan Tao, Jing Xiong, Hayden Kwok-Hay So, Ruobing Xie, Angel X. Chang, Ngai Wong
| Challenge: | Existing studies show that stronger models are not always optimal teachers, suggesting a mismatch between the teacher’s output and the student’s learning ability. |
| Approach: | They propose a method that routes each prompt to its optimal teacher via a query-level router that jointly considers the student models’ learnability and teacher models’ response quality. |
| Outcome: | The proposed method outperforms baselines on six benchmarks including instruct tuning and math reasoning settings. |
Spelling-Aware Construction of Macaronic Texts for Teaching Foreign-Language Vocabulary (D19-1)
Copied to clipboard
| Challenge: | a machine foreign-language teacher replaces word tokens with glosses in a foreign language to ease the human reader into understanding the L2 vocabulary. |
| Approach: | They propose a machine foreign-language teacher that modifies text by replacing word tokens with glosses in a foreign language to ease the human reader into understanding the L2 . |
| Outcome: | The proposed model can learn representations for novel words and is a proxy for word guessing and learning ability of real human students. |
Multi-Sense Embeddings for Language Models and Knowledge Distillation (2025.findings-acl)
Copied to clipboard
| Challenge: | Transformer-based large language models generate different representations for the same token depending on context . however, words and tokens typically have only a limited number of senses . a knowledge distillation method can be used to learn a smaller student model . |
| Approach: | They propose a multi-sense embedding method that uses a clustering algorithm to generate a sense embeddable dictionary. |
| Outcome: | The proposed method offers significant space and inference time savings while maintaining competitive performance. |
A Study on Knowledge Distillation from Weak Teacher for Scaling Up Pre-trained Language Models (2023.findings-acl)
Copied to clipboard
| Challenge: | a study shows that DWT can be effective in the vision domain and natural language processing pre-training stages. |
| Approach: | They examine three key factors to optimize Distillation from Weak Teacher (DWT) DWT is a method of transferring knowledge from a weaker teacher model to a larger student model to improve its performance. |
| Outcome: | a new study examines three key factors to optimize DWT in NLP pre-training scenarios . the impact of teacher model quality and guidelines for adjusting the weighting value for DW T loss are examined . |
Rethinking Task-Specific Knowledge Distillation: Contextualized Corpus as Better Textbook (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for knowledge distillation use a two-stage paradigm: general distillation with a task-agnostic general corpus and task-specific distillation using augmented task- specific corpus. |
| Approach: | They propose a contextualized corpus that contextualizes task corpus with large-scale general corpus through relevance-based text retrieval to improve student learning. |
| Outcome: | The proposed model improves on the GLUE benchmark and shows that it is better than generalized corpus and augmented task-specific corpus. |
MLAS-LoRA: Language-Aware Parameters Detection and LoRA-Based Knowledge Transfer for Multilingual Machine Translation (2025.acl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) have demonstrated strong performance even with limited parallel data. |
| Approach: | They propose a multiple language-aware LoRA knowledge transfer framework that selectively adapts LLMs to MT by transferring knowledge from a large teacher to a small student model. |
| Outcome: | The proposed framework outperforms baseline models on multilingual language pairs by +1.7 BLEU on average. |
Multistage Collaborative Knowledge Distillation from a Large Language Model for Semi-Supervised Sequence Generation (2024.acl-long)
Copied to clipboard
Jiachen Zhao, Wenlong Zhao, Andrew Drozdov, Benjamin Rozonoyer, Md Arafat Sultan, Jay-Yoon Lee, Mohit Iyyer, Andrew McCallum
| Challenge: | Low-resource tasks such as semi-supervised sequence generation require expert knowledge and cost. |
| Approach: | They propose a method for semi-supervised sequence generation where few examples are too scarce to fine tune a model. |
| Outcome: | The proposed method can generalize better than its teacher to unseen examples on semi-supervised sequence generation tasks. |
Unveiling the Key Factors for Distilling Chain-of-Thought Reasoning (2025.findings-acl)
Copied to clipboard
Xinghao Chen, Zhijing Sun, Guo Wenjin, Miaoran Zhang, Yanjun Chen, Yirong Sun, Hui Su, Yijie Pan, Dietrich Klakow, Wenjie Li, Xiaoyu Shen
| Challenge: | Large Language Models (LLMs) excel in reasoning tasks through Chain-of-Thought prompting. |
| Approach: | They examine the factors influencing CoT distillation including granularity, format and teacher model. |
| Outcome: | The proposed model is based on four teacher models and seven student models across seven mathematical and commonsense reasoning datasets. |
Multi-Task Knowledge Distillation with Embedding Constraints for Scholarly Keyphrase Boundary Classification (2023.emnlp-main)
Copied to clipboard
| Challenge: | Scholarly keyphrase boundary classification is the task of identifying highly summative phrases from scientific papers and classifying them into a set of predefined classes. |
| Approach: | They propose a constraint which enforces the teachers and student similarity in the embedding space. |
| Outcome: | The proposed constraint outperforms previous studies and strong baselines on three datasets of scientific documents. |
What Makes it Ok to Set a Fire? Iterative Self-distillation of Contexts and Rationales for Disambiguating Defeasible Social and Moral Situations (2023.findings-emnlp)
Copied to clipboard
Kavel Rao, Liwei Jiang, Valentina Pyatkin, Yuling Gu, Niket Tandon, Nouha Dziri, Faeze Brahman, Yejin Choi
| Challenge: | Moral or ethical judgments rely heavily on the contexts in which they occur . a student model that produces defeasible contexts with improved validity, diversity, and defasibility is superior to intermediate student models . |
| Approach: | a new study uses a student model to provide contextualizations that make an action morally acceptable . the model is based on a dataset of 115K defeasible moral actions rated highly by human annotators . |
| Outcome: | The proposed model outperforms all intermediate models in a high-quality dataset . the model is based on 1.2M entries of contextualizations and rationales for 115K moral actions . |
Long-Chain Reasoning Distillation via Adaptive Prefix Alignment (2026.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have demonstrated remarkable reasoning capabilities, especially in solving complex mathematical problems. |
| Approach: | They propose a framework that exploits teacher CoTs for distillation through adaptive prefix alignment. |
| Outcome: | The proposed framework outperforms baseline models on multiple mathematical reasoning benchmarks by over 3%. |
Text Representation Distillation via Information Bottleneck Principle (2023.emnlp-main)
Copied to clipboard
| Challenge: | Pre-trained language models (PLMs) have recently shown great success in text representation field, however, the high computational cost and high-dimensional representation of PLMs pose significant challenges for practical applications. |
| Approach: | They propose a Knowledge Distillation method that distills large models into smaller representation models to reduce performance degradation after distillation. |
| Outcome: | Empirical results on two main downstream applications of the proposed method show that it reduces the risk of over-fitting and maximizes the mutual information between the model and the input data. |
Let GPT be a Math Tutor: Teaching Math Word Problem Solvers with Customized Exercise Generation (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing approaches for distilling large language models into smaller, more efficient student models are based on educational science principles such as knowledge tracing and personalized learning. |
| Approach: | They propose a method for distilling large language models into smaller, more efficient student models that are aligned with educational science principles such as knowledge tracing and personalized learning. |
| Outcome: | The proposed approach outperforms LLMs on three benchmarks while employing significantly fewer parameters. |
CorrSynth - A Correlated Sampling Method for Diverse Dataset Generation from LLMs (2024.emnlp-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) have demonstrated remarkable performance in diverse tasks using zero-shot and few-shot prompting. |
| Approach: | They propose a method which generates data that is more diverse and faithful to the input prompt using a correlated sampling strategy. |
| Outcome: | The proposed method overcomes the complexity drawbacks of other guidance-based techniques and improves student metrics and intrinsic metrics upon competitive baselines across four datasets. |
Selective Reflection-Tuning: Student-Selected Data Recycling for LLM Instruction-Tuning (2024.findings-acl)
Copied to clipboard
| Challenge: | Instruction tuning is critical to large language models but its success heavily relies on the training data quality. |
| Approach: | They propose a paradigm that synergizes a teacher LLM’s reflection and introspection with the data selection capability of the student LLM to automatically refine existing instruction-tuning data. |
| Outcome: | The proposed method achieves much stronger and top-tier 7B and 13B LLMs without collecting brand-new data. |
Mentor-KD: Making Small Language Models Better Multi-step Reasoners (2024.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have shown impressive emergent capabilities by leveraging Chain-of-Thought (CoT) prompting. |
| Approach: | They propose a Knowledge Distillation approach which transfers multi-step reasoning ability of Large Language Models (LLMs) to smaller LMs by fine-tuning language models of multi- step rationales generated by LLM teachers. |
| Outcome: | The proposed method is able to transfer multi-step reasoning ability of LLMs to smaller LMs while addressing data quality and soft label provision. |
Mitigating Shortcuts in Language Models with Soft Label Encoding (2024.lrec-main)
Copied to clipboard
| Challenge: | Recent studies have shown that large language models rely on spurious correlations in the data for natural language understanding (NLU) tasks. |
| Approach: | They propose a framework for debiasing shortcuts and a dummy class to encode shortcuts into a model and use it to generate soft labels. |
| Outcome: | The proposed framework significantly improves out-of-distribution generalization while maintaining satisfactory in-district accuracy. |
Error-driven Data-efficient Large Multimodal Model Tuning (2025.acl-long)
Copied to clipboard
| Challenge: | Large Multimodal Models (LMMs) have demonstrated impressive performance across numerous academic benchmarks, but task-specific tuning samples are often not readily available or expensive and time-consuming to obtain. |
| Approach: | They propose an error-driven data-efficient tuning framework that aims to efficiently adapt generic LMMs to newly emerging tasks without extensive task-specific training samples. |
| Outcome: | The proposed model achieves an average performance boost of 7.01% on seven tasks across three training data scales and three different training datascales. |
MoDE-CoTD: Chain-of-Thought Distillation for Complex Reasoning Tasks with Mixture of Decoupled LoRA-Experts (2024.lrec-main)
Copied to clipboard
| Challenge: | Current Chain-of-thought Distillation methods hinder CoT reasoning performance . student models are separately distilled from specific reasoning tasks . parameter update of student models severely harms CoT ability on unseen reasoning tasks. |
| Approach: | They propose a method which distills Chain-of-thought reasoning ability of large language models to much smaller student models. |
| Outcome: | The proposed method improves the reasoning ability of large language models on 14 datasets. |
Dual-Space Knowledge Distillation for Large Language Models (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing large language models (LLMs) have strong generalization abilities due to their huge model capacities. |
| Approach: | They propose a dual-space knowledge distillation framework that unifies the output spaces of the two models for KD. |
| Outcome: | The proposed framework outperforms existing white-box KD frameworks on task-agnostic instruction-following benchmarks and can automatically align representations of two models with different vocabularies. |
Maximizing the Effectiveness of Larger BERT Models for Compression (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods for capturing large BERT models as teachers do not fully exploit the potential advantages of larger teachers. |
| Approach: | They propose a method that leverages a pretrained teacher model to guide the training of a lightweight student model to enhance knowledge transfer. |
| Outcome: | The proposed method enhances knowledge transfer by leveraging a pretrained teacher model to guide the training of a lightweight student model. |
CoTD-PO: Chain-of-Thought Distillation with Preference Optimization (2025.findings-emnlp)
Copied to clipboard
Lujie Niu, Haochen Sun, Fangkun Zhao, Sheng Chen, Zimeng Bai, Jiawei Zhang, Caixia Yuan, Xiaojie Wang
| Challenge: | Existing methods for chain-of-thought distillation suffer from a distribution mismatch between teacher-generated training trajectories and the student model's own generative distribution. |
| Approach: | They propose a framework that shifts the training paradigm from passive imitation to active trajectory exploration by allowing students to sample their own answer paths. |
| Outcome: | The proposed method outperforms standard CoT distillation baselines while mitigating mode collapse and preserving semantic diversity. |
Prophecy Distillation for Boosting Abstractive Summarization (2024.lrec-main)
Copied to clipboard
| Challenge: | Abstractive summarization models with maximum likelihood estimation generate unfaithful facts alongside ambiguous focus. |
| Approach: | They propose a framework which learns a regular summarization model to mimic the behavior of being guided by prophecy for boosting abstractive summaries. |
| Outcome: | The proposed model achieves new or matched state-of-the-art on four well-known datasets. |
Fast and Effective On-Policy Distillation from Reasoning Prefixes (2026.findings-acl)
Copied to clipboard
Dongxu Zhang, Zhichao Yang, Sepehr Janghorbani, Jun Han, Andrew Ressler II, Qian Qian, Gregory D Lyng, Sanjit Singh Batra, Robert E. Tillman
| Challenge: | On-policy distillation (OPD) requires expensive on-the-fly sampling of the student policy during training, which substantially increases training cost. |
| Approach: | They propose to use on-policy distillation to sample trajectories from student model . they propose to terminate the sampling early during distillation . |
| Outcome: | The proposed method matches the performance of full OPD in long reasoning outputs while reducing training FLOP by 2x–40x. |
DPED: Multi-Layer Noise Distillation for Privacy-Preserving Text Embeddings (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods to train text embedding models under differential privacy constraints are difficult due to high dimensionality of language data and the presence of rare, identifying linguistic features. |
| Approach: | They propose a framework that leverages teacher-student distillation with noise injection to learn high-quality embeddings while providing differential privacy guarantees. |
| Outcome: | The proposed framework outperforms standard differentially private training methods on benchmark datasets and provides higher privacy-utility trade-offs. |
TAeKD: Teacher Assistant Enhanced Knowledge Distillation for Closed-Source Multilingual Neural Machine Translation (2024.lrec-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) have produced impressive results in the field of Multilingual Neural Machine Translation (MNMT). |
| Approach: | They propose a Teacher Assistant enhanced Knowledge Distillation method to augment knowledge transfer capacity from closed-source MNMT models. |
| Outcome: | The proposed method outperforms the state-of-the-art KD methods on both WMT22 and FLORES-101 test sets. |
Temporal reasoning for timeline summarisation in social media (2025.acl-long)
Copied to clipboard
| Challenge: | Existing temporal reasoning datasets focus on pair-wise event relationships. |
| Approach: | They propose a temporal reasoning dataset focused on temporal relationships among sequential events within narratives that combines temporal thinking with timeline summarisation through a knowledge distillation framework. |
| Outcome: | The proposed model achieves superior performance on mental health-related timeline summarisation tasks, highlighting the importance and generalisability of leveraging temporal reasoning to improve timeline summaries. |
ThinkTuning: Instilling Cognitive Reflections without Distillation (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent advances in test-time scaling have led to the emergence of thinking LLMs that exhibit self-reflective behaviors and multi-step reasoning. |
| Approach: | They propose a GRPO-based interactive training approach that augments the rollouts of a student model with the guidance of . a teacher poses a problem, lets the student try an answer, then gives corrective feedback–enough to point the mind in the right direction and then show the correct solution. |
| Outcome: | The proposed method shows 3.69% improvement over zero-shot baselines and 2.08% and 3.99% improvement over the vanilla-GRPO baselines. |
MedCoach: Enhancing Medical Reasoning in LLMs via Knowledge Graph-Augmented Chain-of-Thought Distillation (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for training specialized reasoning models for the medical domain are limited due to the scarcity of high-quality, large-scale Chain-of-Thought (CoT) data. |
| Approach: | They propose a framework that introduces a dedicated coach role to guide the student model through question decomposition. |
| Outcome: | The proposed framework smooths the learning curve in medical reasoning by facilitating domain adaptation before advancing to complex long-chain reasoning. |
Which Reasoning Trajectories Teach Students to Reason Better? A Simple Metric of Informative Alignment (2026.acl-long)
Copied to clipboard
Yuming Yang, Mingyoung Lai, Wanxu Zhao, Xiaoran Fan, Zhiheng Xi, Mingqi Wu, Chiyue Huang, Jun Zhao, Haijun Lv, Jian Tong, Yunhua Zhou, Yicheng Zou, Qipeng Guo, Tao Gui, Qi Zhang, Xuanjing Huang
| Challenge: | Existing methods assess suitability primarily through student likelihood, favoring trajectories that align closely with the student model’s current behavior but overlooking more informative ones. |
| Approach: | They propose a Rank–Surprisal Ratio metric that captures both alignment and informativeness to assess the suitability of a reasoning trajectory. |
| Outcome: | The proposed metric captures both alignment and informativeness to assess the suitability of a reasoning trajectory. |
EnsemW2S: Enhancing Weak-to-Strong Generalization with Large Language Model Ensembles (2026.findings-acl)
Copied to clipboard
Aakriti Agrawal, Mucong Ding, Chenghao Deng, Zora Che, Arjun Rajaram, Anirudh Satheesh, Bang An, C. Bayan Bruss, John Langford, Furong Huang
| Challenge: | Large Language Models (LLMs) are rapidly approaching and potentially exceeding human-level performance . a novel method aims to improve weak experts' generalization abilities by training them on limited human- level data . |
| Approach: | They propose a method that iteratively combines multiple weak experts to improve their generalization performance by training on limited human-level data. |
| Outcome: | The proposed method improves weak experts' generalization abilities by iterating on weak models and stronger student models. |