Papers with distilled models
DistilQwen2.5: Industrial Practices of Training Distilled Open Lightweight Language Models (2025.acl-industry)
Copied to clipboard
| Challenge: | Existing studies on distilled lightweight LLMs have focused on transferring knowledge from a larger model (the teacher) to a smaller model (sector). |
| Approach: | They propose a family of distilled, lightweight LLMs derived from Qwen2.5 models. |
| Outcome: | Experimental results show that the distilled models have significantly stronger instruction-following capabilities than the original models. |
Distilling the Knowledge of Romanian BERTs Using Multiple Teachers (2022.lrec-1)
Copied to clipboard
Andrei-Marius Avram, Darius Catrina, Dumitru-Clementin Cercel, Mihai Dascalu, Traian Rebedea, Vasile Pais, Dan Tufis
| Challenge: | Existing approaches to train pre-trained language models focus on the English language, thus widening the gap when considering low-resource languages. |
| Approach: | They propose three versions of distilled BERT models for the Romanian language . they argue that the models offer performance comparable to their teachers . |
| Outcome: | The proposed models perform comparable to their teachers, while being twice as fast on a GPU and 35% smaller. |
LEAP: Layer-wise Exit-Aware Pretraining for Efficient Transformer Inference (2026.acl-industry)
Copied to clipboard
Shashank Kapadia, Deep Narayan Mishra, Sujal Reddy Alugubelli, Haoan Wang, Saipraveen Vabbilisetty, Rishi Bhatia, Anupriya Sharma
| Challenge: | Layer-aligned distillation and convergence-based early exit are dominant computational efficiency paradigms for transformer inference. |
| Approach: | They propose a training objective that aligns intermediate student layers to teacher representations and reconciles this incompatibility with standard distillation. |
| Outcome: | The proposed model achieves 1.61 measured wall-clock speedup with 91.9% of samples exiting by layer 7 and 1.80 theoretical layer reduction, where standard distilled models achieve zero effective speedup. |
Generation, Distillation and Evaluation of Motivational Interviewing-Style Reflections with a Foundational Language Model (2024.eacl-long)
Copied to clipboard
| Challenge: | Motivational Interviewing (MI) is a counselling technique used to guide people towards behaviour change. |
| Approach: | They propose a method for distilling reflections from a foundational language model into smaller models that can be owned and controlled. |
| Outcome: | The proposed method achieves 100% success rate on hold-out test set and 90% on the GPT-2 XL. |
SafeRoute: Adaptive Model Selection for Efficient and Accurate Safety Guardrails in Large Language Models (2025.findings-acl)
Copied to clipboard
Seanie Lee, Dong Bok Lee, Dominik Wagner, Minki Kang, Haebin Seong, Tobias Bocklet, Juho Lee, Sung Ju Hwang
| Challenge: | Deploying large language models (LLMs) requires robust safety guard models to detect and block harmful user prompts. |
| Approach: | They propose a binary router that selectively applies a larger safety guard model to the data that the router considers hard. |
| Outcome: | The proposed method outperforms baselines on multiple benchmark datasets on hard and hard examples. |
Multi-Stage Balanced Distillation: Addressing Long-Tail Challenges in Sequence-Level Knowledge Distillation (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Knowledge distillation (KD) is a promising solution for large language models, but their deployment remains computationally expensive. |
| Approach: | They propose a framework which iteratively balances training data within a fixed computational budget and enables the transfer of knowledge from expensive teacher LLMs to smaller student models. |
| Outcome: | The proposed framework achieves state-of-the-art performance across diverse long-tailed datasets, enhancing both the efficiency and efficacy of the distilled models. |
GOLD: Generalized Knowledge Distillation via Out-of-Distribution-Guided Language Data Generation (2024.findings-naacl)
Copied to clipboard
| Challenge: | Existing methods to generate data using LLMs are limited by sampling from the center of original content distribution. |
| Approach: | They propose a task-agnostic data generation and knowledge distillation framework for LLMs that employs an iterative out-of-distribution-guided feedback mechanism to generate data. |
| Outcome: | The proposed framework outperforms prior arts and the LLM on 10 different classification tasks and noisey generated data. |
Distilling Reasoning Capabilities into Smaller Language Models (2023.findings-acl)
Copied to clipboard
| Challenge: | a step-by-step reasoning approach like chain of thought has proved to be effective in eliciting reasoning abilities in large language models. |
| Approach: | They propose a knowledge distillation approach that leverages CoT reasoning capabilities of larger models and distills them into smaller models. |
| Outcome: | The proposed scheme boosts the performance of smaller models over 70% on multiple reasoning datasets. |
Distilling Efficient Language-Specific Models for Cross-Lingual Transfer (2023.findings-acl)
Copied to clipboard
| Challenge: | Massively multilingual Transformers (MMTs) are widely used for cross-lingual transfer learning. |
| Approach: | They propose to extract compressed, language-specific models from MMTs which retain the capacity of the original MMT for cross-lingual transfer. |
| Outcome: | The proposed model outperforms models trained from scratch in zero-shot cross-lingual transfer across benchmarks. |
Latent Agents: A Post-Training Procedure for Internalized Multi-Agent Debate (2026.acl-long)
Copied to clipboard
| Challenge: | Multi-agent debate is compute-intensive and requires long transcripts before answering questions. |
| Approach: | They propose a framework that distills multi-agent debate into a single LLM by combining debate structure learning with internalization via dynamic reward scheduling and length clipping. |
| Outcome: | The proposed model matches or exceeds explicit multi-agent debate performance using 93% fewer tokens across multiple models and benchmarks. |
Block Pruning For Faster Transformers (2021.emnlp-main)
Copied to clipboard
| Challenge: | Pruning methods have proven to be effective at reducing model size, while distillation methods are proven for speeding up inference. |
| Approach: | They propose a block pruning approach that integrates structured pruning methods with the movement pruning paradigm for fine-tuning. |
| Outcome: | The proposed model is 2.4x faster, 74% smaller and faster than distilled models on classification and generation tasks. |
Marco-o1 v2: Towards Widening The Distillation Bottleneck for Reasoning Models (2025.acl-long)
Copied to clipboard
Huifeng Yin, Yu Zhao, Minghao Wu, Xuanfan Ni, Bo Zeng, Huaiyu.wh Huaiyu.wh, Tianqi Shi, Liangying Shao, Chenyang Lyu, Longyue Wang, Weihua Luo, Kaifu Zhang
| Challenge: | Recent efforts to distill large reasoning models into smaller lightweight models have shown competitive performances. |
| Approach: | They propose to distill long Chain-of-Thought data to improve SFT and RL methods by constructing data from scratch using Monte Carlo Tree Search. |
| Outcome: | The proposed method significantly improves reasoning performance on various benchmarks such as math (GSM8K, MATH, AIME). |
Temporal Consistency for LLM Reasoning Process Error Identification (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Empirical evaluations show consistent performance improvements over baseline methods . 7B/8B distilled models outperform all 70B/72B models and GPT-4o on ProcessBench . |
| Approach: | They propose a temporal consistency method that leverages consistency in a sequence of self-reflection actions to improve verification accuracy. |
| Outcome: | The proposed method outperforms existing methods on three benchmarks . it leverages consistency in a sequence of self-reflection actions to improve accuracy . |
Evaluating Cultural Knowledge and Reasoning in LLMs Through Persian Allusions (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Allusion recognition is a critical test of LLMs' ability to deploy stored information in open-ended, figurative settings. |
| Approach: | They propose a framework for evaluating Persian literary allusions through annotations and LLM-generated texts incorporating allusion in novel contexts. |
| Outcome: | The proposed framework evaluates Persian literary allusions through annotations and LLM-generated texts incorporating allusion in novel contexts. |
ERRV: Eliciting Efficient Reasoning through Reasoning Vectors for Policy Optimization in Large Language Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing efforts to improve reasoning efficiency of large language models focus on modifying the reinforcement learning reward, such as adding length penalties. |
| Approach: | They propose a training framework that elicits efficient reasoning through reasoning vectors and a framework that allows the model to generate high-quality responses during reinforcement learning. |
| Outcome: | The proposed framework reduces reasoning length by 30% while maintaining stability, while retaining high accuracy. |
Grammar as Control: Modular Language Generation for the Long Tail (2026.acl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) can bootstrap language technologies for long-tail languages . however, without structured guidance, they produce narrow, unrepresentative samples . |
| Approach: | They propose a framework that transforms descriptive grammars into explicit control mechanisms that guide LLMs to generate typologically balanced synthetic data for downstream training. |
| Outcome: | The proposed framework improves typological entropy and yields a "student-beats-teacher" effect across three low-resource languages. |
LEAF: Knowledge Distillation of Text Embedding Models with Teacher-Aligned Representations (2026.acl-long)
Copied to clipboard
| Challenge: | a new framework for text embedding models is available for free . asymmetrical architectures allow for flexible asymmetry, while asynchronous architectures require small batches . |
| Approach: | They propose a knowledge distillation framework for text embedding models that is compatible with their teacher . they publish leaf-ir, a 23M parameters information retrieval oriented model that ranks no.1 on BEIR . |
| Outcome: | The proposed model is compatible with teacher, enabling flexible asymmetric architectures . it sets a new state-of-the-art (SOTA) on BEIR, and achieves no.1 on the leaderboard . |