Michal Lukasik

Yihan Wang, Si Si, Daliang Li et al.

Pretrained large language models (LLMs) are general purpose problem solvers applicable to a diverse set of tasks with prompts. They can be further improved towards a specific task by fine-tuning on a specialized dataset. However, fine-tuning usually makes the model narrowly specialized on this dataset with reduced general in-context learning performances, which is undesirable whenever the fine-tuned model needs to handle additional tasks where no fine-tuning data is available. In this work, we first demonstrate that fine-tuning on a single task indeed decreases LLMs' general in-context learning performance. We discover one important cause of such forgetting, format specialization, where the model overfits to the format of the fine-tuned task.We further show that format specialization happens at the very beginning of fine-tuning. To solve this problem, we propose Prompt Tuning with MOdel Tuning (ProMoT), a simple yet effective two-stage fine-tuning framework that reduces format specialization and improves generalization.ProMoT offloads task-specific format learning into additional and removable parameters by first doing prompt tuning and then fine-tuning the model itself with this soft prompt attached. With experiments on several fine-tuning tasks and 8 in-context evaluation tasks, we show that ProMoT achieves comparable performance on fine-tuned tasks to standard fine-tuning, but with much less loss of in-context learning performances across a board range of out-of-domain evaluation tasks. More importantly, ProMoT can even enhance generalization on in-context learning tasks that are semantically related to the fine-tuned task, e.g. ProMoT on En-Fr translation significantly improves performance on other language pairs, and ProMoT on NLI improves performance on summarization. Experiments also show that ProMoT can improve the generalization performance of multi-task training.

Daliang Li, Ankit Singh Rawat, Manzil Zaheer et al.

Large language models (LLMs) have led to a series of breakthroughs in natural language processing (NLP), owing to their excellent understanding and generation abilities. Remarkably, what further sets these models apart is the massive amounts of world knowledge they internalize during pretraining. While many downstream applications provide the model with an informational context to aid its performance on the underlying task, how the model's world knowledge interacts with the factual information presented in the context remains under explored. As a desirable behavior, an LLM should give precedence to the context whenever it contains task-relevant information that conflicts with the model's memorized knowledge. This enables model predictions to be grounded in the context, which can then be used to update or correct specific model predictions without frequent retraining. By contrast, when the context is irrelevant to the task, the model should ignore it and fall back on its internal knowledge. In this paper, we undertake a first joint study of the aforementioned two properties, namely controllability and robustness, in the context of LLMs. We demonstrate that state-of-the-art T5 and PaLM (both pretrained and finetuned) could exhibit poor controllability and robustness, which do not scale with increasing model size. As a solution, we propose a novel method - Knowledge Aware FineTuning (KAFT) - to strengthen both controllability and robustness by incorporating counterfactual and irrelevant contexts to standard supervised datasets. Our comprehensive evaluation showcases the utility of KAFT across model architectures and sizes.

Zitong Yang, Michal Lukasik, Vaishnavh Nagarajan et al.

The impressive generalization performance of modern neural networks is attributed in part to their ability to implicitly memorize complex training patterns. Inspired by this, we explore a novel mechanism to improve model generalization via explicit memorization. Specifically, we propose the residual-memorization (ResMem) algorithm, a new method that augments an existing prediction model (e.g. a neural network) by fitting the model's residuals with a $k$-nearest neighbor based regressor. The final prediction is then the sum of the original model and the fitted residual regressor. By construction, ResMem can explicitly memorize the training labels. Empirically, we show that ResMem consistently improves the test set generalization of the original prediction model across various standard vision and natural language processing benchmarks. Theoretically, we formulate a stylized linear regression problem and rigorously show that ResMem results in a more favorable test risk over the base predictor.

Serena Wang, Harikrishna Narasimhan, Yichen Zhou et al.

Knowledge distillation has proven to be an effective technique in improving the performance a student model using predictions from a teacher model. However, recent work has shown that gains in average efficiency are not uniform across subgroups in the data, and in particular can often come at the cost of accuracy on rare subgroups and classes. To preserve strong performance across classes that may follow a long-tailed distribution, we develop distillation techniques that are tailored to improve the student's worst-class performance. Specifically, we introduce robust optimization objectives in different combinations for the teacher and student, and further allow for training with any tradeoff between the overall accuracy and the robust worst-class objective. We show empirically that our robust distillation techniques not only achieve better worst-class performance, but also lead to Pareto improvement in the tradeoff between overall performance and worst-class performance compared to other baseline methods. Theoretically, we provide insights into what makes a good teacher when the goal is to train a robust student.

Michal Lukasik, Vaishnavh Nagarajan, Ankit Singh Rawat et al.

The success of modern neural networks has prompted study of the connection between memorisation and generalisation: overparameterised models generalise well, despite being able to perfectly fit (memorise) completely random labels. To carefully study this issue, Feldman proposed a metric to quantify the degree of memorisation of individual training examples, and empirically computed the corresponding memorisation profile of a ResNet on image classification bench-marks. While an exciting first glimpse into what real-world models memorise, this leaves open a fundamental question: do larger neural models memorise more? We present a comprehensive empirical analysis of this question on image classification benchmarks. We find that training examples exhibit an unexpectedly diverse set of memorisation trajectories across model sizes: most samples experience decreased memorisation under larger models, while the rest exhibit cap-shaped or increasing memorisation. We show that various proxies for the Feldman memorization score fail to capture these fundamental trends. Lastly, we find that knowledge distillation, an effective and popular model compression technique, tends to inhibit memorisation, while also improving generalisation. Specifically, memorisation is mostly inhibited on examples with increasing memorisation trajectories, thus pointing at how distillation improves generalisation.

Lin Chen, Michal Lukasik, Wittawat Jitkrittum et al.

Classical wisdom in machine learning holds that the generalization error can be decomposed into bias and variance, and these two terms exhibit a \emph{trade-off}. However, in this paper, we show that for an ensemble of deep learning based classification models, bias and variance are \emph{aligned} at a sample level, where squared bias is approximately \emph{equal} to variance for correctly classified sample points. We present empirical evidence confirming this phenomenon in a variety of deep learning models and datasets. Moreover, we study this phenomenon from two theoretical perspectives: calibration and neural collapse. We first show theoretically that under the assumption that the models are well calibrated, we can observe the bias-variance alignment. Second, starting from the picture provided by the neural collapse theory, we show an approximate correlation between bias and variance.

Yasi Zhang, Tianyu Chen, Mingyuan Zhou et al.

Large language models (LLMs) are increasingly deployed as automated evaluators that assign numeric scores to model outputs, a paradigm known as LLM-as-a-Judge. However, standard Reinforcement Learning (RL) methods typically rely on binary rewards (e.g., 0-1 accuracy), thereby ignoring the ordinal structure inherent in regression tasks; for instance, they fail to recognize that predicting 4 is significantly better than predicting 1 when the ground truth is 5. Conversely, existing regression-aware approaches are often confined to Supervised Fine-Tuning (SFT), limiting their ability to explore optimal reasoning paths. To bridge this gap, we propose \textbf{REAL} (\underline{RE}gression-\underline{A}ware Reinforcement \underline{L}earning), a principled RL framework designed to optimize regression rewards, and also proven to be optimal for correlation metrics. A key technical challenge is that the regression objective is explicitly policy-dependent, thus invalidating standard policy gradient methods. To address this, we employ the generalized policy gradient estimator, which naturally decomposes optimization into two complementary components: (1) exploration over Chain-of-Thought (CoT) trajectory, and (2) regression-aware prediction refinement of the final score. Extensive experiments across model scales (8B to 32B) demonstrate that REAL consistently outperforms both regression-aware SFT baselines and standard RL methods, exhibiting significantly better generalization on out-of-domain benchmarks. On Qwen3-32B specifically, we achieve gains of +8.40 Pearson and +7.20 Spearman correlation over the SFT baseline, and +18.30/+11.20 over the base model. These findings highlight the critical value of integrating regression objectives into RL exploration for accurate LLM evaluation.

Cheng-Han Chiang, Hung-yi Lee, Michal Lukasik

The LLM-as-a-judge paradigm uses large language models (LLMs) for automated text evaluation, where a numerical assessment is assigned by an LLM to the input text following scoring rubrics. Existing methods for LLM-as-a-judge use cross-entropy (CE) loss for fine-tuning, which neglects the numeric nature of score prediction. Recent work addresses numerical prediction limitations of LLM fine-tuning through regression-aware fine-tuning, which, however, does not consider chain-of-thought (CoT) reasoning for score prediction. In this paper, we introduce TRACT (Two-stage Regression-Aware fine-tuning with CoT), a method combining CoT reasoning with regression-aware training. TRACT consists of two stages: first, seed LLM is fine-tuned to generate CoTs, which serve as supervision for the second stage fine-tuning. The training objective of TRACT combines the CE loss for learning the CoT reasoning capabilities, and the regression-aware loss for the score prediction. Experiments across four LLM-as-a-judge datasets and two LLMs show that TRACT significantly outperforms existing methods. Extensive ablation studies validate the importance of each component in TRACT.

Michal Lukasik, Harikrishna Narasimhan, Aditya Krishna Menon et al.

Large language models (LLMs) have shown strong results on a range of applications, including regression and scoring tasks. Typically, one obtains outputs from an LLM via autoregressive sampling from the model's output distribution. We show that this inference strategy can be sub-optimal for common regression and scoring evaluation metrics. As a remedy, we build on prior work on Minimum Bayes Risk decoding, and propose alternate inference strategies that estimate the Bayes-optimal solution for regression and scoring metrics in closed-form from sampled responses. We show that our proposal significantly improves over baselines across datasets and models.

Michal Lukasik, Lin Chen, Harikrishna Narasimhan et al.

Bipartite ranking is a fundamental supervised learning problem, with the goal of learning a ranking over instances with maximal Area Under the ROC Curve (AUC) against a single binary target label. However, one may often observe multiple binary target labels, e.g., from distinct human annotators. How can one synthesize such labels into a single coherent ranking? In this work, we formally analyze two approaches to this problem -- loss aggregation and label aggregation -- by characterizing their Bayes-optimal solutions. We show that while both approaches can yield Pareto-optimal solutions, loss aggregation can exhibit label dictatorship: one can inadvertently (and undesirably) favor one label over others. This suggests that label aggregation can be preferable to loss aggregation, which we empirically verify.

Wittawat Jitkrittum, Michal Lukasik, Ananda Theertha Suresh et al.

Multi-party computation (MPC) is a branch of cryptography where multiple non-colluding parties execute a well designed protocol to securely compute a function. With the non-colluding party assumption, MPC has a cryptographic guarantee that the parties will not learn sensitive information from the computation process, making it an appealing framework for applications that involve privacy-sensitive user data. In this paper, we study training and inference of neural networks under the MPC setup. This is challenging because the elementary operations of neural networks such as the ReLU activation function and matrix-vector multiplications are very expensive to compute due to the added multi-party communication overhead. To address this, we propose the HD-cos network that uses 1) cosine as activation function, 2) the Hadamard-Diagonal transformation to replace the unstructured linear transformations. We show that both of the approaches enjoy strong theoretical motivations and efficient computation under the MPC setup. We demonstrate on multiple public datasets that HD-cos matches the quality of the more expensive baselines.

Srinadh Bhojanapalli, Ayan Chakrabarti, Andreas Veit et al.

Pairwise dot product-based attention allows Transformers to exchange information between tokens in an input-dependent way, and is key to their success across diverse applications in language and vision. However, a typical Transformer model computes such pairwise attention scores repeatedly for the same sequence, in multiple heads in multiple layers. We systematically analyze the empirical similarity of these scores across heads and layers and find them to be considerably redundant, especially adjacent layers showing high similarity. Motivated by these findings, we propose a novel architecture that reuses attention scores computed in one layer in multiple subsequent layers. Experiments on a number of standard benchmarks show that reusing attention delivers performance equivalent to or better than standard transformers, while reducing both compute and memory usage.

Michal Lukasik, Srinadh Bhojanapalli, Aditya Krishna Menon et al.

Knowledge distillation is widely used as a means of improving the performance of a relatively simple student model using the predictions from a complex teacher model. Several works have shown that distillation significantly boosts the student's overall performance; however, are these gains uniform across all data subgroups? In this paper, we show that distillation can harm performance on certain subgroups, e.g., classes with few associated samples. We trace this behaviour to errors made by the teacher distribution being transferred to and amplified by the student model. To mitigate this problem, we present techniques which soften the teacher influence for subgroups where it is less reliable. Experiments on several image classification benchmarks show that these modifications of distillation maintain boost in overall accuracy, while additionally ensuring improvement in subgroup performance.

Srinadh Bhojanapalli, Ayan Chakrabarti, Himanshu Jain et al.

State-of-the-art transformer models use pairwise dot-product based self-attention, which comes at a computational cost quadratic in the input sequence length. In this paper, we investigate the global structure of attention scores computed using this dot product mechanism on a typical distribution of inputs, and study the principal components of their variation. Through eigen analysis of full attention score matrices, as well as of their individual rows, we find that most of the variation among attention scores lie in a low-dimensional eigenspace. Moreover, we find significant overlap between these eigenspaces for different layers and even different transformer models. Based on this, we propose to compute scores only for a partial subset of token pairs, and use them to estimate scores for the remaining pairs. Beyond investigating the accuracy of reconstructing attention scores themselves, we investigate training transformer models that employ these approximations, and analyze the effect on overall accuracy. Our analysis and the proposed method provide insights into how to balance the benefits of exact pair-wise attention and its significant computational expense.

Michal Lukasik, Himanshu Jain, Aditya Krishna Menon et al.

Label smoothing has been shown to be an effective regularization strategy in classification, that prevents overfitting and helps in label de-noising. However, extending such methods directly to seq2seq settings, such as Machine Translation, is challenging: the large target output space of such problems makes it intractable to apply label smoothing over all possible outputs. Most existing approaches for seq2seq settings either do token level smoothing, or smooth over sequences generated by randomly substituting tokens in the target sequence. Unlike these works, in this paper, we propose a technique that smooths over \emph{well formed} relevant sequences that not only have sufficient n-gram overlap with the target sequence, but are also \emph{semantically similar}. Our method shows a consistent and significant improvement over the state-of-the-art techniques on different datasets.

Aleksandar Bojchevski, Johannes Gasteiger, Bryan Perozzi et al.

Graph neural networks (GNNs) have emerged as a powerful approach for solving many network mining tasks. However, learning on large graphs remains a challenge - many recently proposed scalable GNN approaches rely on an expensive message-passing procedure to propagate information through the graph. We present the PPRGo model which utilizes an efficient approximation of information diffusion in GNNs resulting in significant speed gains while maintaining state-of-the-art prediction performance. In addition to being faster, PPRGo is inherently scalable, and can be trivially parallelized for large datasets like those found in industry settings. We demonstrate that PPRGo outperforms baselines in both distributed and single-machine training environments on a number of commonly used academic graphs. To better analyze the scalability of large-scale graph learning methods, we introduce a novel benchmark graph with 12.4 million nodes, 173 million edges, and 2.8 million node features. We show that training PPRGo from scratch and predicting labels for all nodes in this graph takes under 2 minutes on a single machine, far outpacing other baselines on the same graph. We discuss the practical application of PPRGo to solve large-scale node classification problems at Google.

Michal Lukasik, Boris Dadachev, Gonçalo Simões et al.

Document and discourse segmentation are two fundamental NLP tasks pertaining to breaking up text into constituents, which are commonly used to help downstream tasks such as information retrieval or text summarization. In this work, we propose three transformer-based architectures and provide comprehensive comparisons with previously proposed approaches on three standard datasets. We establish a new state-of-the-art, reducing in particular the error rates by a large margin in all cases. We further analyze model sizes and find that we can build models with many fewer parameters while keeping good performance, thus facilitating real-world applications.

Michal Lukasik, Srinadh Bhojanapalli, Aditya Krishna Menon et al.

Label smoothing is commonly used in training deep learning models, wherein one-hot training labels are mixed with uniform label vectors. Empirically, smoothing has been shown to improve both predictive performance and model calibration. In this paper, we study whether label smoothing is also effective as a means of coping with label noise. While label smoothing apparently amplifies this problem --- being equivalent to injecting symmetric noise to the labels --- we show how it relates to a general family of loss-correction techniques from the label noise literature. Building on this connection, we show that label smoothing is competitive with loss-correction under label noise. Further, we show that when distilling models from noisy data, label smoothing of the teacher is beneficial; this is in contrast to recent findings for noise-free problems, and sheds further light on settings where label smoothing is beneficial.

Arkaitz Zubiaga, Elena Kochkina, Maria Liakata et al.

Rumour stance classification, defined as classifying the stance of specific social media posts into one of supporting, denying, querying or commenting on an earlier post, is becoming of increasing interest to researchers. While most previous work has focused on using individual tweets as classifier inputs, here we report on the performance of sequential classifiers that exploit the discourse features inherent in social media interactions or 'conversational threads'. Testing the effectiveness of four sequential classifiers -- Hawkes Processes, Linear-Chain Conditional Random Fields (Linear CRF), Tree-Structured Conditional Random Fields (Tree CRF) and Long Short Term Memory networks (LSTM) -- on eight datasets associated with breaking news stories, and looking at different types of local and contextual features, our work sheds new light on the development of accurate stance classifiers. We show that sequential classifiers that exploit the use of discourse properties in social media conversations while using only local features, outperform non-sequential classifiers. Furthermore, we show that LSTM using a reduced set of features can outperform the other sequential classifiers; this performance is consistent across datasets and across types of stances. To conclude, our work also analyses the different features under study, identifying those that best help characterise and distinguish between stances, such as supporting tweets being more likely to be accompanied by evidence than denying tweets. We also set forth a number of directions for future research.

Arkaitz Zubiaga, Elena Kochkina, Maria Liakata et al.

Rumour stance classification, the task that determines if each tweet in a collection discussing a rumour is supporting, denying, questioning or simply commenting on the rumour, has been attracting substantial interest. Here we introduce a novel approach that makes use of the sequence of transitions observed in tree-structured conversation threads in Twitter. The conversation threads are formed by harvesting users' replies to one another, which results in a nested tree-like structure. Previous work addressing the stance classification task has treated each tweet as a separate unit. Here we analyse tweets by virtue of their position in a sequence and test two sequential classifiers, Linear-Chain CRF and Tree CRF, each of which makes different assumptions about the conversational structure. We experiment with eight Twitter datasets, collected during breaking news, and show that exploiting the sequential structure of Twitter conversations achieves significant improvements over the non-sequential methods. Our work is the first to model Twitter conversations as a tree structure in this manner, introducing a novel way of tackling NLP tasks on Twitter conversations.

Michal Lukasik, Kalina Bontcheva, Trevor Cohn et al.

Social media tend to be rife with rumours while new reports are released piecemeal during breaking news. Interestingly, one can mine multiple reactions expressed by social media users in those situations, exploring their stance towards rumours, ultimately enabling the flagging of highly disputed rumours as being potentially false. In this work, we set out to develop an automated, supervised classifier that uses multi-task learning to classify the stance expressed in each individual tweet in a rumourous conversation as either supporting, denying or questioning the rumour. Using a classifier based on Gaussian Processes, and exploring its effectiveness on two datasets with very different characteristics and varying distributions of stances, we show that our approach consistently outperforms competitive baseline classifiers. Our classifier is especially effective in estimating the distribution of different types of stance associated with a given rumour, which we set forth as a desired characteristic for a rumour-tracking system that will warn both ordinary users of Twitter and professional news practitioners when a rumour is being rebutted.

Michal Lukasik, Trevor Cohn, Kalina Bontcheva

Social media is a rich source of rumours and corresponding community reactions. Rumours reflect different characteristics, some shared and some individual. We formulate the problem of classifying tweet level judgements of rumours as a supervised learning task. Both supervised and unsupervised domain adaptation are considered, in which tweets from a rumour are classified on the basis of other annotated rumours. We demonstrate how multi-task learning helps achieve good results on rumours from the 2011 England riots.