Papers by Mehdi Ali
Not Far Away, Not So Close: Sample Efficient Nearest Neighbour Data Augmentation via MiniMax (2021.findings-acl)
Copied to clipboard
| Challenge: | Existing kNN-based augmentation techniques blindly incorporate all samples, but MiniMax-kNN uses a subset of augmented samples to maximize KL-divergence between teacher and student models. |
| Approach: | They propose a semi-supervised approach to augmented data augmentation using kNN. |
| Outcome: | The proposed method outperforms existing kNN-based augmentation techniques on several classification tasks and requires fewer augmented examples and less computation to achieve superior performance. |
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. |
When Chosen Wisely, More Data Is What You Need: A Universal Sample-Efficient Strategy For Data Augmentation (2022.findings-acl)
Copied to clipboard
| Challenge: | Existing DA methods naively add a certain number of augmented samples without considering the quality and the added computational cost of these samples. |
| Approach: | They propose a data-augmented DA technique that generates or reweights augmented samples . they say it is faster to train and can be plugged into any DA method . |
| Outcome: | The proposed technique is faster to train and more efficient than existing methods. |
Improving Generalization of Pre-trained Language Models via Stochastic Weight Averaging (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Recent studies show that the flatness of the local minimum correlates well with better generalization. |
| Approach: | They propose to use a method encouraging convergence to a flatter minimum to fine-tune PLMs. |
| Outcome: | The proposed method outperforms state-of-the-art methods on NLP tasks without extra computation cost. |
Investigating Multilingual Instruction-Tuning: Do Polyglot Models Demand for Multilingual Instructions? (2024.emnlp-main)
Copied to clipboard
Alexander Weber, Klaudia Thellmann, Jan Ebert, Nicolas Flores-Herr, Jens Lehmann, Michael Fromm, Mehdi Ali
| Challenge: | a study of multilingual pre-trained LLMs on parallel instruction-tuning benchmarks shows that instruction-following models can be used across languages by up to 9.9%. |
| Approach: | They conduct an extensive study of the performance of multilingual pre-trained LLMs instruction-tuned on parallel instruction-uning datasets. |
| Outcome: | The proposed model improves cross-lingual instruction following capabilities by 9.9% on a large and mid-sized LLM on parallel instruction-tuning datasets. |
Judging Quality Across Languages: A Multilingual Approach to Pretraining Data Filtering with Language Models (2025.emnlp-main)
Copied to clipboard
Mehdi Ali, Manuel Brack, Max Lübbering, Elias Wendt, Abbas Goher Khan, Richard Rutmann, Alex Jude, Maurice Kraus, Alexander Arno Weber, Felix Stollenwerk, David Kaczér, Florian Mai, Lucie Flek, Rafet Sifa, Nicolas Flores-Herr, Joachim Koehler, Patrick Schramowski, Michael Fromm, Kristian Kersting
| Challenge: | Existing open-source multilingual datasets rely on heuristic filtering methods restricting both their cross-lingual transferability and scalability. |
| Approach: | They propose a systematic approach that curates diverse and high-quality multilingual data at scale while significantly reducing computational demands. |
| Outcome: | Evaluated empirically across 35 languages, the proposed approach outperforms current heuristic filtering methods like Fineweb2 and improves model training quality and retention rates. |
How to Select One Among All ? An Empirical Study Towards the Robustness of Knowledge Distillation in Natural Language Understanding (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Knowledge Distillation (KD) is a model compression algorithm that helps transfer knowledge in a large neural network into a smaller one. |
| Approach: | They propose a framework to assess adversarial robustness of multiple KD algorithms. |
| Outcome: | The proposed algorithm achieves state-of-the-art on the GLUE benchmark and out-of domain generalization and adversarial robustness compared to competitive methods. |
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. |
Universal-KD: Attention-based Output-Grounded Intermediate Layer Knowledge Distillation (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for intermediate layer matching are limited due to huge over-parameterization . |
| Approach: | They propose to match intermediate layers of teacher and student in output space via attention-based layer projection. |
| Outcome: | The proposed method outperforms existing methods on GLUE tasks. |
QDyLoRA: Quantized Dynamic Low-Rank Adaptation for Efficient Large Language Model Tuning (2024.emnlp-industry)
Copied to clipboard
Hossein Rajabzadeh, Mojtaba Valipour, Tianshu Zhu, Marzieh Tahaei, Hyock Kwon, Ali Ghodsi, Boxing Chen, Mehdi Rezagholizadeh
| Challenge: | Existing methods to fine tune large language models require huge memory, limiting the choice to acquire Larger models. |
| Approach: | They propose an efficient quantization approach for dynamic low-rank adaptation that can efficiently fine tune large language models on a set of pre-defined LoRA ranks. |
| Outcome: | The proposed method outperforms QLoRA and is competitive to QLouRA and outperformed when employing its optimal rank. |
Continuation KD: Improved Knowledge Distillation through the Lens of Continuation Optimization (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for knowledge distillation (KD) do not mitigate the noise in the teacher’s output: modeling the noisy behaviour of the teacher can distract the student from learning more useful features. |
| Approach: | They propose a method that optimizes the highly non-convex KD objective by starting with the smoothed version of this objective and making it more complex as the training proceeds. |
| Outcome: | The proposed method achieves state-of-the-art performance on NLU and computer vision tasks. |
Pro-KD: Progressive Distillation by Following the Footsteps of the Teacher (2022.coling-1)
Copied to clipboard
| Challenge: | Knowledge distillation (KD) is a powerful tool for deep learning applications. |
| Approach: | They propose a method which defines a smoother training path for the student by following the training footprints of the teacher rather than solely relying on distilling from a single mature fully-trained teacher. |
| Outcome: | The proposed technique is quite effective in mitigating the capacity-gap problem and the checkpoint search problem. |
Tokenizer Choice For LLM Training: Negligible or Crucial? (2024.findings-naacl)
Copied to clipboard
Mehdi Ali, Michael Fromm, Klaudia Thellmann, Richard Rutmann, Max Lübbering, Johannes Leveling, Katrin Klug, Jan Ebert, Niclas Doll, Jasper Buschhoff, Charvi Jain, Alexander Weber, Lena Jurkschat, Hammam Abdelwahab, Chelsea John, Pedro Ortiz Suarez, Malte Ostendorff, Samuel Weinbach, Rafet Sifa, Stefan Kesselheim, Nicolas Flores-Herr
| Challenge: | Recent success of large language models has been driven by curating the training dataset composition, scaling of model architectures and advancements in pretraining objectives, leaving tokenizer influence as a blind spot. |
| Approach: | They conduct a comprehensive study on the influence of tokenizer choice on LLM downstream performance by training 24 mono- and multilingual LLMs at a 2.6B parameter scale. |
| Outcome: | The proposed model can significantly impact the model's downstream performance and training costs. |
Efficient Citer: Tuning Large Language Models for Enhanced Answer Quality and Verification (2024.findings-naacl)
Copied to clipboard
Marzieh Tahaei, Aref Jafari, Ahmad Rashid, David Alfonso-Hermelo, Khalil Bibi, Yimeng Wu, Ali Ghodsi, Boxing Chen, Mehdi Rezagholizadeh
| Challenge: | Existing models with explicit citations lack the ability to verify information generated by these models. |
| Approach: | They construct a citation training dataset and fine-tune two models to address the challenge of explicit citations efficiently. |
| Outcome: | The proposed models surpass ChatGPT and exhibit exceptional out-of-domain generalization in both human and automatic evaluation. |
Sorted LLaMA: Unlocking the Potential of Intermediate Layers of Large Language Models for Dynamic Inference (2024.findings-eacl)
Copied to clipboard
| Challenge: | Large language models excel at understanding and generating human-like text, but their widespread deployment can be prohibitively expensive. |
| Approach: | They propose a method that makes large language models dynamic without Pre-Training . they use modularity in networks and sort sub-models based on computation/accuracy in a nested manner. |
| Outcome: | The proposed method can make large language models dynamic without pre-training and replace standard fine-tuning with sorted fine- tuning. |
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. |
Kronecker Decomposition for GPT Compression (2022.acl-short)
Copied to clipboard
| Challenge: | GPT is an auto-regressive Transformer-based pre-trained language model . but its huge size can be prohibitive for deploying on low capacity devices . |
| Approach: | They use a Kronecker decomposition technique to compress GPT models . they use ILKD to refine the model on downstream tasks . |
| Outcome: | The proposed model outperforms the existing DistilGPT2 model on language modeling and general language understanding evaluation benchmark tasks. |
LLM Parameters for Math Across Languages: Shared or Separate? (2026.acl-srw)
Copied to clipboard
Behzad Shomali, Luisa Victor, Tim Selbach, Ali Hamza Bashir, David Berghaus, Joachim Koehler, Mehdi Ali, Markus Frey
| Challenge: | Existing research on large language models (LLMs) has focused on performance or representational properties, but it remains unclear whether these differences reflect language-specific parameters or a shared mechanism. |
| Approach: | They propose to localize and compare model parameters that support mathematical reasoning across languages. |
| Outcome: | The proposed analysis shows that the model parameters in English and lower-resource languages exhibit partial cross-lingual overlap with systematic language-dependent differences. |
DyLoRA: Parameter-Efficient Tuning of Pre-trained Models using Dynamic Search-Free Low-Rank Adaptation (2023.eacl-main)
Copied to clipboard
| Challenge: | Pre-training/fine-tuning of pre-training models has become more expensive and resource-hungry. |
| Approach: | They propose a low-rank adaptation technique that trains LoRA blocks for a range of ranks instead of a single rank. |
| Outcome: | The proposed method trains LoRA blocks for a range of ranks instead of a single rank . it can train dynamic search-free models with DyLoRA at least 4 to 7 times faster than LoRA . |
KroneckerBERT: Significant Compression of Pre-trained Language Models Through Kronecker Decomposition and Knowledge Distillation (2022.naacl-main)
Copied to clipboard
| Challenge: | a recent study shows that over-parameterized pre-trained language models are unsuitable for low-capacity devices. |
| Approach: | They propose a transformer-based pre-trained language model that is overparameterized . they use a two-stage knowledge distillation scheme to train the model . |
| Outcome: | The proposed model outperforms state-of-the-art models on well-known NLP benchmarks. |