SENTINEL generates targeted tasks from model failures in a Controller-Proposer-Solver loop, raising Pass^1 from 66.4 to 74.9 on Tau2-Bench Retail and outperforming standard RL.
hub Canonical reference
A Survey on Knowledge Distillation of Large Language Models
Canonical reference. 83% of citing Pith papers cite this work as background.
46
Pith papers citing it
51
external citations · Pith
Background
83% of classified citations
abstract
In the era of Large Language Models (LLMs), Knowledge Distillation (KD) emerges as a pivotal methodology for transferring advanced capabilities from leading proprietary LLMs, such as GPT-4, to their open-source counterparts like LLaMA and Mistral. Additionally, as open-source LLMs flourish, KD plays a crucial role in both compressing these models, and facilitating their self-improvement by employing themselves as teachers. This paper presents a comprehensive survey of KD's role within the realm of LLM, highlighting its critical function in imparting advanced knowledge to smaller models and its utility in model compression and self-improvement. Our survey is meticulously structured around three foundational pillars: \textit{algorithm}, \textit{skill}, and \textit{verticalization} -- providing a comprehensive examination of KD mechanisms, the enhancement of specific cognitive abilities, and their practical implications across diverse fields. Crucially, the survey navigates the intricate interplay between data augmentation (DA) and KD, illustrating how DA emerges as a powerful paradigm within the KD framework to bolster LLMs' performance. By leveraging DA to generate context-rich, skill-specific training data, KD transcends traditional boundaries, enabling open-source models to approximate the contextual adeptness, ethical alignment, and deep semantic insights characteristic of their proprietary counterparts. This work aims to provide an insightful guide for researchers and practitioners, offering a detailed overview of current methodologies in KD and proposing future research directions. Importantly, we firmly advocate for compliance with the legal terms that regulate the use of LLMs, ensuring ethical and lawful application of KD of LLMs. An associated Github repository is available at https://github.com/Tebmer/Awesome-Knowledge-Distillation-of-LLMs.
hub tools
citation-role summary
background 6
citation-polarity summary
roles
background 6representative citing papers
KAT detects persistent low-KL agreement traps in on-policy distillation via a dynamic threshold to filter weak supervision, improving avg@k by 2.66% and pass@k by 3.43% on four math benchmarks while shortening rollouts by 59.73%.
Model collapse occurs in structured interactive learning if and only if the directed interaction graph satisfies a specific topological condition, with finite-sample guarantees for linear regression and asymptotic results for M-estimators.
CoDistill-GRPO lets small and large models mutually improve via co-distillation in GRPO, raising small-model math accuracy by over 11 points while cutting large-model training time by about 18%.
Chain-based Distillation constructs a sequence of anchor models to enable efficient initialization of variable-sized SLMs through interpolation, with bridge distillation for cross-architecture transfer, yielding better performance than scratch training.
AOPD modifies on-policy distillation by using localized divergence minimization for non-positive advantages instead of negative reinforcement, yielding average gains of 4.09/8.34 over standard OPD on math reasoning benchmarks under strong/weak initialization.
LOVER creates an unsupervised logic-regularized verifier that reaches 95% of supervised verifier performance on reasoning tasks across 10 datasets.
ReflectMT internalizes reflection via two-stage RL to enable direct high-quality machine translation that outperforms explicit reasoning models like DeepSeek-R1 on WMT24 while using 94% fewer tokens.
A semi-supervised framework distills vision foundation models into compact instance segmentation experts that outperform their teachers by up to 11.9 AP on Cityscapes and 8.6 AP on ADE20K while being 11 times smaller.
CodeRL+ integrates variable-level execution trajectory inference into RLVR training to align textual code representations with execution semantics, delivering 4.6% relative pass@1 gains and generalization to code-reasoning and test-output tasks.
CODI compresses explicit CoT into continuous space via self-distillation and is the first implicit method to match explicit CoT performance on GSM8k at GPT-2 scale with 3.1x compression and 28.2% higher accuracy than prior implicit approaches.
Distilling an 8B reasoning teacher into a 0.6B student recovers most summary quality at ~50× speed, but teacher type—not scale alone—determines which capabilities transfer.
DuoMem distills from a 72B teacher to 4B student via context and parameter space, achieving 77.9% success on ALFWorld vs 4.3% baseline.
ARKD uses an RL policy network to adaptively balance FKL and RKL in LLM distillation, claiming gains of 0.4-0.6 points on Rouge-L and BertScore over baselines.
LLM-labeled training sets for entity matching produce student models with F1 scores within 2 points of benchmark-trained models on five datasets at a cost of $28-41 versus 470 hours of manual work.
Activation steering on early layers improves diversity of synthetic data for low-resource languages and often boosts downstream classifier performance compared to non-steered prompting.
A new conversion method (KEPT) transforms Orion-14B into Llama-format models while preserving benchmark performance using ~123M tokens of distillation.
DASD improves math reasoning in LLMs by adaptively directing self-distillation based on per-token entropy to balance exploration and step accuracy, outperforming prior self-distillation and RLVR baselines on six benchmarks.
FireFly inverts task synthesis by exploring real MCP servers first via pairwise tool graphs and sub-DAG sampling, then generates 5,144 verified tasks backward from outcomes to train a 4B model that matches Claude Sonnet 4.6 on tool-calling benchmarks.
OpenJarvis decomposes personal AI into Intelligence, Engine, Agents, Tools & Memory, and Learning primitives and applies LLM-guided spec search to produce on-device configurations that reach within 3.2 pp of cloud baselines on average across eight tasks.
MOPD improves on-policy distillation by using peer successes and failures from multiple rollouts to construct more informative teacher signals, yielding consistent gains over baselines on reasoning benchmarks.
PARSE trains a prompt-aware linear router on dense-model outputs to select dynamic SVD ranks, improving accuracy up to 10% at 0.6 compression ratio on LLaMA-7B while delivering 2.5x prefill and 2.4x decode speedups.
SimCT enlarges the supervision space in cross-tokenizer on-policy distillation using short jointly tokenizable multi-token continuations, producing consistent gains over shared-token baselines on math and code benchmarks.
UniSD unifies self-distillation components for autoregressive LLMs and its full integrated version improves base models by 5.4 points and baselines by 2.8 points across six benchmarks.
citing papers explorer
-
SENTINEL: Failure-Driven Reinforcement Learning for Training Tool-Using Language Model Agents
SENTINEL generates targeted tasks from model failures in a Controller-Proposer-Solver loop, raising Pass^1 from 66.4 to 74.9 on Tau2-Bench Retail and outperforming standard RL.
-
Escaping the KL Agreement Trap in On-Policy Distillation
KAT detects persistent low-KL agreement traps in on-policy distillation via a dynamic threshold to filter weak supervision, improving avg@k by 2.66% and pass@k by 3.43% on four math benchmarks while shortening rollouts by 59.73%.
-
When Does Model Collapse Occur in Structured Interactive Learning?
Model collapse occurs in structured interactive learning if and only if the directed interaction graph satisfies a specific topological condition, with finite-sample guarantees for linear regression and asymptotic results for M-estimators.
-
CoDistill-GRPO: A Co-Distillation Recipe for Efficient Group Relative Policy Optimization
CoDistill-GRPO lets small and large models mutually improve via co-distillation in GRPO, raising small-model math accuracy by over 11 points while cutting large-model training time by about 18%.
-
Chain-based Distillation for Effective Initialization of Variable-Sized Small Language Models
Chain-based Distillation constructs a sequence of anchor models to enable efficient initialization of variable-sized SLMs through interpolation, with bridge distillation for cross-architecture transfer, yielding better performance than scratch training.
-
Asymmetric On-Policy Distillation: Bridging Exploitation and Imitation at the Token Level
AOPD modifies on-policy distillation by using localized divergence minimization for non-positive advantages instead of negative reinforcement, yielding average gains of 4.09/8.34 over standard OPD on math reasoning benchmarks under strong/weak initialization.
-
Logic-Regularized Verifier Elicits Reasoning from LLMs
LOVER creates an unsupervised logic-regularized verifier that reaches 95% of supervised verifier performance on reasoning tasks across 10 datasets.
-
ReflectMT: Internalizing Reflection for Efficient and High-Quality Machine Translation
ReflectMT internalizes reflection via two-stage RL to enable direct high-quality machine translation that outperforms explicit reasoning models like DeepSeek-R1 on WMT24 while using 94% fewer tokens.
-
Training a Student Expert via Semi-Supervised Foundation Model Distillation
A semi-supervised framework distills vision foundation models into compact instance segmentation experts that outperform their teachers by up to 11.9 AP on Cityscapes and 8.6 AP on ADE20K while being 11 times smaller.
-
CodeRL+: Improving Code Generation via Reinforcement with Execution Semantics Alignment
CodeRL+ integrates variable-level execution trajectory inference into RLVR training to align textual code representations with execution semantics, delivering 4.6% relative pass@1 gains and generalization to code-reasoning and test-output tasks.
-
CODI: Compressing Chain-of-Thought into Continuous Space via Self-Distillation
CODI compresses explicit CoT into continuous space via self-distillation and is the first implicit method to match explicit CoT performance on GSM8k at GPT-2 scale with 3.1x compression and 28.2% higher accuracy than prior implicit approaches.
-
Different Teachers, Different Capabilities: Sub-1B On-Device Distillation for Structured Text Enrichment
Distilling an 8B reasoning teacher into a 0.6B student recovers most summary quality at ~50× speed, but teacher type—not scale alone—determines which capabilities transfer.
-
DuoMem: Towards Capable On-Device Memory Agents via Dual-Space Distillation
DuoMem distills from a 72B teacher to 4B student via context and parameter space, achieving 77.9% success on ALFWorld vs 4.3% baseline.
-
ARKD: Adaptive Reinforcement Learning-Guided Bidirectional KL Divergence Distillation for Text Generation
ARKD uses an RL policy network to adaptively balance FKL and RKL in LLM distillation, claiming gains of 0.4-0.6 points on Rouge-L and BertScore over baselines.
-
Labeling Training Data for Entity Matching Using Large Language Models
LLM-labeled training sets for entity matching produce student models with F1 scores within 2 points of benchmark-trained models on five datasets at a cost of $28-41 versus 470 hours of manual work.
-
Want Better Synthetic Data? Steer It: Activation Steering for Low-Resource Language Generation
Activation steering on early layers improves diversity of synthetic data for low-resource languages and often boosts downstream classifier performance compared to non-steered prompting.
-
Llamion Technical Report
A new conversion method (KEPT) transforms Orion-14B into Llama-format models while preserving benchmark performance using ~123M tokens of distillation.
-
Tailoring Teaching to Aptitude: Direction-Adaptive Self-Distillation for LLM Reasoning
DASD improves math reasoning in LLMs by adaptively directing self-distillation based on per-token entropy to balance exploration and step accuracy, outperforming prior self-distillation and RLVR baselines on six benchmarks.
-
Firefly: Illuminating Large-Scale Verified Tool-Call Data Generation from Real APIs
FireFly inverts task synthesis by exploring real MCP servers first via pairwise tool graphs and sub-DAG sampling, then generates 5,144 verified tasks backward from outcomes to train a 4B model that matches Claude Sonnet 4.6 on tool-calling benchmarks.
-
OpenJarvis: Personal AI, On Personal Devices
OpenJarvis decomposes personal AI into Intelligence, Engine, Agents, Tools & Memory, and Learning primitives and applies LLM-guided spec search to produce on-device configurations that reach within 3.2 pp of cloud baselines on average across eight tasks.
-
Multi-Rollout On-Policy Distillation via Peer Successes and Failures
MOPD improves on-policy distillation by using peer successes and failures from multiple rollouts to construct more informative teacher signals, yielding consistent gains over baselines on reasoning benchmarks.
-
Different Prompts, Different Ranks: Prompt-aware Dynamic Rank Selection for SVD-based LLM Compression
PARSE trains a prompt-aware linear router on dense-model outputs to select dynamic SVD ranks, improving accuracy up to 10% at 0.6 compression ratio on LLaMA-7B while delivering 2.5x prefill and 2.4x decode speedups.
-
SimCT: Recovering Lost Supervision for Cross-Tokenizer On-Policy Distillation
SimCT enlarges the supervision space in cross-tokenizer on-policy distillation using short jointly tokenizable multi-token continuations, producing consistent gains over shared-token baselines on math and code benchmarks.
-
UniSD: Towards a Unified Self-Distillation Framework for Large Language Models
UniSD unifies self-distillation components for autoregressive LLMs and its full integrated version improves base models by 5.4 points and baselines by 2.8 points across six benchmarks.
-
DeCo-DETR: Decoupled Cognition DETR for efficient Open-Vocabulary Object Detection
DeCo-DETR builds hierarchical semantic prototypes offline and uses decoupled training streams to deliver competitive zero-shot open-vocabulary detection with improved inference speed.
-
LLM-AutoDP: Automatic Data Processing via LLM Agents for Model Fine-tuning
LLM agents iteratively generate and optimize data processing strategies for fine-tuning, delivering over 80% win rates versus unprocessed data and 65% versus LLM-based AutoML baselines while cutting search time by up to 10x.
-
VERDI: VLM-Embedded Reasoning for Autonomous Driving
VERDI aligns perception, prediction, and planning outputs of end-to-end AD models with VLM-generated text features at training time to embed structured reasoning, yielding up to 11% better l2 distance and 10% higher non-collision rate in closed-loop tests.
-
RPAM: A Principled Metric for Evaluating Associations in Language Models with High Predictive Validity in Downstream Outputs
Relative Probability Association Metric (RPAM) measures LM associations via softmax-normalized continuation probabilities and correlates strongly with human associations and downstream LM behavior across three models.
-
Geometric Foundation Model Distillation for Efficient Lunar 3D Reconstruction
Distillation of a 688M-parameter MASt3R teacher yields up to 7x smaller students that retain most lunar reconstruction accuracy and outperform sparse-supervised baselines.
-
Cross-Resolution Semantic Transfer for Robust Text-to-Image Retrieval in Low-Resolution Surveillance
CRST improves ultra-low-resolution text-to-image person retrieval by 5.7% Rank-1 and 5.3% mAP on average across three datasets while stabilizing mixed-resolution galleries.
-
Decision-Making with Lightweight Confidence-Aware Language Model for Autonomous Driving
Lightweight confidence-aware LM distilled from multi-agent CoT demonstrations achieves SOTA success rates on nuPlan benchmark for AD decision-making with low inference latency.
-
ReAD: Reinforcement-Guided Capability Distillation for Large Language Models
ReAD applies a contextual bandit to allocate fixed-token distillation budget across interdependent LLM capabilities, yielding higher task utility and fewer negative spillovers than standard methods.
-
Large Language Model Post-Training: A Unified View of Off-Policy and On-Policy Learning
LLM post-training is unified as off-policy or on-policy interventions that expand support for useful behaviors, reshape policies within reachable states, or consolidate behavior across training stages.
-
A drone-based framework for coral habitat mapping via weakly supervised segmentation
A multi-scale weakly supervised framework converts underwater point classifications into coarse masks to train UAV coral segmentation models, then refines them via self-training to reach 86.07% pixel accuracy and 52.23% mIoU without pixel-level labels.
-
Model Compression vs. Adversarial Robustness: An Empirical Study on Language Models for Code
Empirical tests show compressed code language models retain task performance but suffer markedly lower robustness under four standard adversarial attacks.
-
RAP: Runtime Adaptive Pruning for LLM Inference
RAP is a reinforcement learning framework for runtime-adaptive pruning of LLMs that jointly optimizes model weights and KV-cache usage under varying memory budgets.
-
Quantum Software Engineering in Practice: FPGA and AI Integration for Quantum Certification
QAccCert integrates FPGAs and LLMs within a Quantum Software Engineering framework to optimize CHSH entanglement certification parameters in ideal quantum simulations.
-
Enhancing deep learning models for time series classification via knowledge distillation
Knowledge distillation most benefits intermediate-complexity students for time series classification, cutting parameters sharply while matching teacher accuracy across FCN, Inception, and ConvTran on UCR.
-
AIGP: An LLM-Based Framework for Long-Term Value Alignment in E-Commerce Pricing
AIGP combines LLMs with offline RL and DPO to produce interpretable pricing policies that improved GMV by 13.21%, ROI by 7.59%, and milestone achievement by 8.20% in 14-day online tests versus baseline.
-
PRIDE: Privileged Information-enhanced Distillation for Empathetic Dialogue Generation
PRIDE distills empathetic reasoning from large teacher LLMs to smaller students via an empathy prompt, multi-source attention, and dual-alignment loss using privileged information available only at training time.
-
LUNA-AD: Lightweight Uncertainty-Aware Language Model with Lifelong Learning for Autonomous Driving
LUNA-AD introduces a tri-system model with multi-agent hypothesis exploration, distilled lightweight inference, and reflection-driven lifelong learning that claims state-of-the-art success rates on nuPlan benchmarks with reduced latency.
-
Knowledge Distillation Must Account for What It Loses
Knowledge distillation evaluations must report lost teacher capabilities via a Distillation Loss Statement rather than relying solely on task scores.
-
ADMEDTAGGER: an annotation framework for distillation of expert knowledge for the Polish medical language
Llama 3.1 annotates Polish medical texts to train DistilBERT classifiers achieving F1 scores above 0.80 that are 500 times smaller than the teacher model.
-
Large Language Models for Multi-Robot Systems: A Survey
A survey that categorizes LLM uses in multi-robot systems across task allocation, motion planning, action generation, and human interaction, while noting challenges and future research opportunities.
- When Context Returns: Toward Robust Internalization in On-Policy Distillation
- SOD: Step-wise On-policy Distillation for Small Language Model Agents