Papers with teacher models
Discourse Structures Guided Fine-grained Propaganda Identification (2023.emnlp-main)
Copied to clipboard
| Challenge: | Using teacher-predicted probabilities and knowledge distillation frameworks to identify propaganda content is important. |
| Approach: | They propose to integrate local and global discourse structures for propaganda discovery and construct two teacher models for identifying PDTB-style discourse relations between nearby sentences and common discourse roles of sentences in a news article respectively. |
| Outcome: | The proposed models improve accuracy and recall of propaganda content identification at sentence-level and token-level. |
Trace-of-Thought Prompting: Investigating Prompt-Based Knowledge Distillation Through Question Decomposition (2024.acl-srw)
Copied to clipboard
| Challenge: | Trace-of-Thought Prompting allows small neural networks to emulate larger, teacher models with reduced computational demands. |
| Approach: | They propose a framework to distill critical reasoning capabilities from teacher models to student models . they use problem decomposition to enhance interpretability and facilitate human-in-the-loop interventions . |
| Outcome: | a new framework enables small neural networks to emulate the performance of larger, teacher models . it leverages problem decomposition to enhance interpretability and facilitate human-in-the-loop interventions . the proposed framework is available on github.com/trace-of-thought/trac-of_thought-prompting/main . |
Think Like You Execute: Verifiable Chain of Thought from Program Traces (2026.acl-industry)
Copied to clipboard
Shailja Thakur, Vaibhav Saxena, Rohan Kulkarni, Shivdeep Singh, Parameswaran Selvam, Hiroshi Kanayama, Hima Patel
| Challenge: | Current synthetic Chain-of-Thought (CoT) training data often consists of plausible-sounding explanations generated by teacher models, not verifiable accounts of actual program behavior. |
| Approach: | They propose to ground CoT generation directly in program execution traces to improve reasoning capabilities. |
| Outcome: | The proposed pipeline improves performance on live code benchmarks and on cruxEval-output and cruxeval-input. |
Distill-C: Enhanced NL2SQL via Distilled Customization with LLMs (2025.naacl-industry)
Copied to clipboard
Cong Duy Vu Hoang, Gioacchino Tangari, Clemence Lanfranchi, Dalu Guo, Paul Cayet, Steve Siu, Don Dharmasiri, Yuan-Fang Li, Long Duong, Damien Hilloulin, Rhicheek Patra, Sungpack Hong, Hassan Chafi
| Challenge: | Domain- and customer-specific requirements complicate the problem of NL2SQL customization. |
| Approach: | They propose a distilled customization framework tailored for NL2SQL tasks. |
| Outcome: | The proposed framework outperforms teacher models on three benchmarks and achieves an average improvement of 36% in execution accuracy. |
AMR-Evol: Adaptive Modular Response Evolution Elicits Better Knowledge Distillation for Large Language Models in Code Generation (2024.emnlp-main)
Copied to clipboard
| Challenge: | proprietary large language models (LLMs) have demonstrated impressive code generation performance. |
| Approach: | They propose an adaptive module-based model that refines the direct response distillation process by modular decomposition and adaptive response evolution. |
| Outcome: | The proposed framework outperforms baseline model and code generation methods on three popular benchmarks. |
POINTS-Reader: Distillation-Free Adaptation of Vision-Language Models for Document Conversion (2025.emnlp-main)
Copied to clipboard
Yuan Liu, Zhongyin Zhao, Le Tian, Haicheng Wang, Xubing Ye, Yangxiu You, Zilin Yu, Chuhan Wu, Zhou Xiao, Yang Yu, Jie Zhou
| Challenge: | Existing approaches to training document conversion models with manual annotation are costly and time-consuming, and training student models by distilling outputs from teacher models can significantly limit their performance in real-world applications. |
| Approach: | They propose a fully automated framework for constructing high-quality document extraction datasets and models capable of handling diverse document formats and layouts. |
| Outcome: | The proposed model outperforms existing models and improves on annotated documents. |
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. |
Learning from Diverse Reasoning Paths with Routing and Collaboration (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies suggest that the reasoning abilities of large language models (LLMs) grows with model size and pre-training data. |
| Approach: | They propose to combine quality filtering, conditional routing, and cooperative peer teaching to transfer knowledge from powerful teacher models to compact and transparent students. |
| Outcome: | Experiments show that QR-Distill is superior to traditional methods. |
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. |
Sparse Distillation: Speeding Up Text Classification by Using Bigger Student Models (2022.naacl-main)
Copied to clipboard
| Challenge: | Existing methods to reduce inference cost by distilling transformer models into lightweight student models are limited for high-volume use cases. |
| Approach: | They propose to distill state-of-the-art transformer models into lightweight student models to reduce computation cost at inference time. |
| Outcome: | The proposed pipeline achieves up to 600x speed-up on GPUs and CPUs on six single-sentence text classification tasks and in domain generalization settings. |
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. |
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. |
XtremeDistil: Multi-stage Distillation for Massive Multilingual Models (2020.acl-main)
Copied to clipboard
| Challenge: | Existing work on pre-trained language models focuses on reducing the size of these models into shallow ones. |
| Approach: | They propose a knowledge distillation technique that leverages teacher internal representations to reduce the size of pre-trained language models. |
| Outcome: | The proposed method outperforms previous methods in multilingual Named Entity Recognition (NER) it reduces the size of teacher models by 35x while retaining 95% of its F1 score. |
Multi-Granularity Semantic Revision for Large Language Model Distillation (2026.acl-long)
Copied to clipboard
Xiaoyu Liu, Yun Zhang, Wei Li, Simiao Li, Xudong Huang, Hanting Chen, Yehui Tang, Jie Hu, Zhiwei Xiong, Yunhe Wang
| Challenge: | Existing methods for generating large language models rely on student-generated outputs, which introduce generation errors and misguide the distillation process. |
| Approach: | They propose a multi-granularity semantic revision method for LLM distillation that corrects errors using teacher-generated tokens and re-generates the sequence to minimize errors. |
| Outcome: | The proposed method reduces errors and misguides distillation on student models and improves consistency between teacher and student outputs. |
Addressing Overthinking in Large Vision-Language Models via Gated Perception-Reasoning Optimization (2026.findings-acl)
Copied to clipboard
Xingjian Diao, Zheyuan Liu, Chunhui Zhang, Weiyi Wu, Keyi Kong, Lin Shi, Kaize Ding, Soroush Vosoughi, Jiang Gui
| Challenge: | Prior work has attempted to mitigate this issue by using adaptive reasoning strategies, but these methods overlook a fundamental bottleneck: visual perception failures. |
| Approach: | They propose a meta-reasoning controller that dynamically routes computation among three decision paths at each generation step. |
| Outcome: | The proposed method outperforms slow-thinking methods while producing shorter responses. |
Meta-KD: A Meta Knowledge Distillation Framework for Language Model Compression across Domains (2021.acl-long)
Copied to clipboard
| Challenge: | Pre-trained language models have been successful in NLP tasks, but their large size and long inference time limit their deployment in real-time applications. |
| Approach: | They propose a meta-teacher model that captures transferable knowledge across domains and passes it to students. |
| Outcome: | The proposed model can distill large teacher models into small student models with guidance from the meta-teacher. |
Few Shot Rationale Generation using Self-Training with Dual Teachers (2023.findings-acl)
Copied to clipboard
| Challenge: | Existing models that generate free-text explanations for annotated labels are expensive and require a large annotation dataset. |
| Approach: | They propose a self-training approach leveraging both labeled and unlabeled data to further improve few-shot models by combining teacher models and a multi-tasking student model. |
| Outcome: | The proposed model improves on three public datasets and can generate a free-text explanation for predicted labels. |
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. |
Structure-Level Knowledge Distillation For Multilingual Sequence Labeling (2020.acl-main)
Copied to clipboard
| Challenge: | Existing multilingual models still underperform individual monolingual models due to model capacity limitations. |
| Approach: | They propose to distill the structural knowledge of several monolingual models (teachers) to the unified multilingual model (student). |
| Outcome: | The proposed model outperforms strong baseline models and teacher models on 4 multilingual tasks with 25 datasets and has stronger zero-shot generalizability. |
MTA4DPR: Multi-Teaching-Assistants Based Iterative Knowledge Distillation for Dense Passage Retrieval (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies have shown the effectiveness of knowledge distillation in DPR, but there is a performance gap between the teacher and the distilled student. |
| Approach: | They propose an iterative knowledge distillation method which transfers knowledge from teacher to student with help of multiple assistants in an iterated manner. |
| Outcome: | The proposed method achieves state-of-the-art performance among models with same parameters on multiple datasets and is competitive when compared with larger models. |
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. |
Reasoning with OmniThought: A Large CoT Dataset with Verbosity and Cognitive Difficulty Annotations (2026.acl-long)
Copied to clipboard
| Challenge: | Existing resources often fail to provide extensive reasoning problems with coherent CoT processes distilled from multiple teacher models. |
| Approach: | They propose a large-scale dataset featuring 2 million CoT processes generated by multiple powerful LRMs. |
| Outcome: | The proposed dataset features 2 million CoT processes and is validated by multiple powerful LRMs. |
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. |
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. |
Debate, Reflect, and Distill: Multi-Agent Feedback with Tree-Structured Preference Optimization for Efficient Language Model Enhancement (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for enhancing small models struggle to yield substantial and lasting performance gains. |
| Approach: | They propose a Debate and Reflect framework that orchestrates multi-turn debates between smaller models and stronger teacher models. |
| Outcome: | The proposed framework outperforms existing methods by a large margin in smaller models. |
From Mimicking to Integrating: Knowledge Integration for Pre-Trained Language Models (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Existing models for natural language processing (NLP) are fine-tuned and released for research and deployments. |
| Approach: | They propose a PLM reuse paradigm that merges teacher-PLM knowledge into a student model. |
| Outcome: | The proposed paradigm can reduce the computational cost and environmental side-effects of retraining the PLM from scratch. |
Assistant-Guided Mitigation of Teacher Preference Bias in LLM-as-a-Judge (2025.findings-emnlp)
Copied to clipboard
| Challenge: | LLM-as-a-Judge uses large language models to evaluate the quality of LLM generated responses, but training proxy judge models using evaluation data generated by powerful teacher models introduces a critical yet previously overlooked issue: teacher preference bias. |
| Approach: | They propose a new setting that incorporates an additional assistant model, which is not biased toward the teacher model’s responses, to complement the training data. |
| Outcome: | The proposed model reduces teacher preference bias while maintaining strong performance across six evaluation benchmarks. |
Guiding Through Complexity: What Makes Good Supervision for Hard Reasoning Tasks? (2025.naacl-long)
Copied to clipboard
| Challenge: | Using weak teacher models to effectively supervise LLMs can improve performance on hard reasoning tasks. |
| Approach: | They propose two data-driven supervision strategies that offer supervision data at different quality levels upon tasks of varying complexity. |
| Outcome: | The proposed methods outperform "perfectly correct" supervision on harder subtasks even when the outcome error rate is high. |
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. |
Efficiently Selecting Response Generation Strategies for Synthetic Data Construction by Self-Aligned Perplexity (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Using a small sample of data, we find that perplexity is suboptimal in characterizing “familiarity” . |
| Approach: | They propose a method that assesses a small subset of generated data to estimate suitability for a specific target LLM. |
| Outcome: | The proposed method assesses a small subset of generated data to estimate suitability for a specific target LLM. |
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. |
Can LLM Watermarks Robustly Prevent Unauthorized Knowledge Distillation? (2025.acl-long)
Copied to clipboard
| Challenge: | Large Language Model (LLM) watermarking is radioactive and enables the detection of watermarks inherited by student models when trained on the outputs of watermarked teacher models. |
| Approach: | They propose two types of watermark removal attacks that allow student models to perform untraceable knowledge distillation while avoiding watermark inheritance. |
| Outcome: | The proposed attacks eliminate inherited watermarks while maintaining knowledge transfer efficiency and low computational overhead. |
Referee: Reference-Free Sentence Summarization with Sharper Controllability through Symbolic Knowledge Distillation (2022.emnlp-main)
Copied to clipboard
| Challenge: | a new framework for sentence summarization is available that can be trained reference-free . a high-quality dataset of sentence-summary pairs with varying degrees of compression ratios is obtained . |
| Approach: | They propose a framework for sentence summarization that can be trained reference-free . they propose 'referee' that iteratively distills latent knowledge into better models . |
| Outcome: | The proposed framework outperforms existing models in the use of explicit examples from teacher models without compromising the quality of the summarization. |
Distillation Enhanced Generative Retrieval (2024.findings-acl)
Copied to clipboard
| Challenge: | Generative retrieval is a promising new paradigm in text retrieval that generates identifier strings of relevant passages as the retrieval target. |
| Approach: | They propose a framework that leverages generative language models to enhance generative retrieval by distillation. |
| Outcome: | The proposed framework achieves state-of-the-art performance among the generative retrieval methods. |
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. |
Self-Evolution Knowledge Distillation for LLM-based Machine Translation (2025.coling-main)
Copied to clipboard
| Challenge: | Existing knowledge distillation strategies for large language models minimize output distributions between student and teacher models indiscriminately for each token. |
| Approach: | They propose a distillation strategy that integrates teacher and one-hot distribution of ground truth into the student distribution as prior knowledge, which promotes the distillation process. |
| Outcome: | The proposed method brings an average improvement of approximately 1.4 SacreBLEU points across four translation directions in the WMT22 test sets. |
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. |
Large Language Models Are Reasoning Teachers (2023.acl-long)
Copied to clipboard
| Challenge: | Recent studies have shown that chain-of-thought (CoT) prompting can elicit language models to solve complex reasoning tasks step-by-step. |
| Approach: | They propose a method that uses large model samples as reasoning teachers to fine-tune smaller models. |
| Outcome: | The proposed method outperforms prompt-based methods and the teacher model in many tasks and extends it by leveraging the teacher's ability to generate multiple rationales for each original sample. |
TwT: Thinking without Tokens by Habitual Reasoning Distillation with Multi-Teachers’ Guidance (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have made significant strides in problem-solving by incorporating reasoning processes, but this enhanced reasoning capability results in an increased number of output tokens during inference, leading to higher computational costs. |
| Approach: | They propose a method that internalizes explicit reasoning into the model’s habitual behavior through a Teacher-Guided compression strategy inspired by human cognition. |
| Outcome: | The proposed method reduces inference-time costs while maintaining high performance while preserving high quality and diversity of the distillation dataset. |
Experience is the Teacher: Reusing Atomic Thoughts from LLMs to Improve Medical Dialogue (2026.findings-acl)
Copied to clipboard
Guangya Yu, Hui Luo, Qi Ye, Ruihui Hou, Weiyan Zhang, Mingxi Shang, Xuanwu Li, ChunMing Wang, Tong Ruan
| Challenge: | Recent large reasoning models (LLMs) lack dynamic and diverse thinking capabilities . reusing atomic thoughts provides a practical pathway toward dynamic reasoning . |
| Approach: | They propose a framework that extracts atomic thoughts from teacher models and reuses them to guide reasoning and generate responses. |
| Outcome: | The proposed framework extracts atomic thoughts from teacher models and reuses them to guide reasoning and generate responses. |
EvoCoT: Overcoming the Exploration Bottleneck in Reinforcement Learning for LLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to reinforcement learning with verifiable reward (RLVR) are limited by difficulty or lack of exploration. |
| Approach: | They propose a self-evolving curriculum learning framework based on chain-of-thought reasoning optimization that constrains exploration space by self-generating and verifying CoT trajectories. |
| Outcome: | The proposed framework enables LLMs to solve previously unsolved problems without external supervision and is compatible with various RL fine-tuning methods. |
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. |
Model Calibration for Emotion Detection (2025.findings-emnlp)
Copied to clipboard
| Challenge: | a MixUp method is used to calibrate emotion detection models based on knowledge distillation and the MixUp data augmentation technique. |
| Approach: | They propose a method that uses knowledge distillation and the MixUp data augmentation technique to enhance the trustworthiness of emotion detection models. |
| Outcome: | The proposed method improves the accuracy of the teacher models and the student models. |
Comprehensive and Efficient Distillation for Lightweight Sentiment Analysis Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent efforts to develop lightweight and practical sentiment analysis models are limited by manual instruction and large-scale user texts. |
| Approach: | They propose a framework for sentiment analysis that uses attribute-based instruction construction and difficulty-based data filtering to distill knowledge. |
| Outcome: | The proposed framework outperforms baseline methods in data efficiency and performance. |
Beyond Logits: Aligning Feature Dynamics for Effective Knowledge Distillation (2025.acl-long)
Copied to clipboard
Guoqiang Gong, Jiaxing Wang, Jin Xu, Deping Xiang, Zicheng Zhang, Leqi Shen, Yifeng Zhang, JunhuaShu JunhuaShu, ZhaolongXing ZhaolongXing, Zhen Chen, Pengzhang Liu, Ke Zhang
| Challenge: | Knowledge distillation (KD) compresses large language models into lightweight versions called student models. |
| Approach: | They propose to align the entire feature dynamics between teacher and student models by using two additional loss terms to achieve this. |
| Outcome: | The proposed method matches the entire feature dynamics between teacher and student models rather than just the final states. |
PARIF: Pushing the Pareto Frontier of Instruction Following and Reasoning with Curriculum Reinforcement Learning (2026.acl-long)
Copied to clipboard
| Challenge: | Existing alignment methods struggle to balance general reasoning with instruction-following (IF) this is hindered by dependency on teacher models, reward hacking, and reasoning-answer inconsistencies. |
| Approach: | They propose a two-stage curriculum learning framework based on Reinforcement Learning from Verifiable Rewards to enhance both IF and general reasoning capabilities. |
| Outcome: | The proposed framework outperforms leading models on six representative IF tasks while achieving a 21.25% relative average improvement over the original model. |
AutoRubric: Rubric-Based Generative Rewards for Faithful Multimodal Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Multimodal large language models (MLLMs) have advanced from perception tasks to complex multi-step reasoning. |
| Approach: | They propose a framework that integrates reinforcement learning with verifiable rewards with process-level supervision through automatically collected rubric-based generative rewards. |
| Outcome: | The proposed framework achieves state-of-the-art performance on six multimodal reasoning benchmarks and significantly improves reasoning faithfulness in dedicated evaluations. |
Learning to Refine: Self-Refinement of Parallel Reasoning in LLMs (2026.findings-acl)
Copied to clipboard
Qibin Wang, Pu Zhao, Shaohan Huang, Fangkai Yang, Lu Wang, Furu Wei, Qingwei Lin, Saravan Rajmohan, Dongmei Zhang
| Challenge: | Existing approaches to test-time scaling are limited due to the quality of candidate responses. |
| Approach: | They propose a new metric to quantify the relative improvement of self-refinement beyond majority voting. |
| Outcome: | The proposed method achieves state-of-the-art performance across five benchmarks over other methods. |
Mosaic-IT: Cost-Free Compositional Data Synthesis for Instruction Tuning (2025.findings-acl)
Copied to clipboard
| Challenge: | Current instruction tuning relies on teacher models or human intervention to generate and refine the instructions and responses for training, which are costly, non-sustainable, and may lack diversity. |
| Approach: | They propose a human/model-free compositional data synthesis method that can create rich and diverse augmentations from existing instruction tuning data to enhance large language models. |
| Outcome: | The proposed method improves performance over benchmarks and reduces training costs by 80% compared with original instruction tuning. |
Beyond Output Matching: Bidirectional Alignment for Enhanced In-Context Learning (2025.acl-long)
Copied to clipboard
Chengwei Qin, Wenhan Xia, Fangkai Jiao, Chen Chen, Yuchen Hu, Bosheng Ding, Ruirui Chen, Shafiq Joty
| Challenge: | Existing methods to train student models on the generated outputs of teacher models are not efficient for ICL. |
| Approach: | They propose to align the output of smaller (student) models with that of larger (teacher) models by incorporating a ranking loss and aligning the token-level output distribution. |
| Outcome: | The proposed model outperforms baseline models on a variety of tasks involving language understanding, reasoning, and coding. |
Making VLMs More Robot-Friendly: Self-Critical Distillation of Low-Level Procedural Reasoning (2025.emnlp-main)
Copied to clipboard
Chan Young Park, Jillian Fisher, Marius Memmel, Dipika Khullar, Seoho Yun, Abhishek Gupta, Yejin Choi
| Challenge: | Large language models (LLMs) are human-centric, but omit low-level, spatially grounded details needed for robotic execution. |
| Approach: | They propose a lightweight framework for vision-language procedural planning that enables iteratively critique, revise and verify their own plans without external supervision or teacher models. |
| Outcome: | a new framework outperforms weaker models 100X the size in vision-language procedural planning . the framework generates higher-quality, execution-ready plans that can be used at inference and fine-tuning . |
TRAC: Teacher-Guided Token Reward with Adaptive Calibration for Robust Policy Optimization (2026.acl-long)
Copied to clipboard
| Challenge: | Current reward models for reinforcement learning (RL) rely on outcome rewards that propagate a single scalar value across all tokens based on final correctness. |
| Approach: | They propose a framework that derives dense token-level supervision from LLMs . they use a multi-granularity calibration mechanism to modulate teacher influence . |
| Outcome: | The proposed framework evaluates teacher reliability across problem-level expertise, trajectory-level discrimination, and token-level confidence. |