The Curse of Depth in Large Language Models

In this paper, we introduce the Curse of Depth, a concept that highlights, explains, and addresses the recent observation in modern Large Language Models(LLMs) where nearly half of the layers are less effective than expected. We first confirm the wide existence of this phenomenon across the most popular families of LLMs such as Llama, Mistral, DeepSeek, and Qwen. Our analysis, theoretically and empirically, identifies that the underlying reason for the ineffectiveness of deep layers in LLMs is the widespread usage of Pre-Layer Normalization (Pre-LN). While Pre-LN stabilizes the training of Transformer LLMs, its output variance exponentially grows with the model depth, which undesirably causes the derivative of the deep Transformer blocks to be an identity matrix, and therefore barely contributes to the training. To resolve this training pitfall, we propose LayerNorm Scaling, which scales the variance of output of the layer normalization inversely by the square root of its depth. This simple modification mitigates the output variance explosion of deeper Transformer layers, improving their contribution. Our experimental results, spanning model sizes from 130M to 1B, demonstrate that LayerNorm Scaling significantly enhances LLM pre-training performance compared to Pre-LN. Moreover, this improvement seamlessly carries over to supervised fine-tuning. All these gains can be attributed to the fact that LayerNorm Scaling enables deeper layers to contribute more effectively during training.

Sebra: Debiasing Through Self-Guided Bias Ranking

Ranking samples by fine-grained estimates of spuriosity (the degree to which spurious cues are present) has recently been shown to significantly benefit bias mitigation, over the traditional binary biased-\textit{vs}-unbiased partitioning of train sets. However, this spuriosity ranking comes with the requirement of human supervision. In this paper, we propose a debiasing framework based on our novel \ul{Se}lf-Guided \ul{B}ias \ul{Ra}nking (\emph{Sebra}), that mitigates biases (spurious correlations) via an automatic ranking of data points by spuriosity within their respective classes. Sebra leverages a key local symmetry in Empirical Risk Minimization (ERM) training -- the ease of learning a sample via ERM inversely correlates with its spuriousity; the fewer spurious correlations a sample exhibits, the harder it is to learn, and vice versa. However, globally across iterations, ERM tends to deviate from this symmetry. Sebra dynamically steers ERM to correct this deviation, facilitating the sequential learning of attributes in increasing order of difficulty, \ie, decreasing order of spuriosity. As a result, the sequence in which Sebra learns samples naturally provides spuriousity rankings. We use the resulting fine-grained bias characterization in a contrastive learning framework to mitigate biases from multiple sources. Extensive experiments show that Sebra consistently outperforms previous state-of-the-art unsupervised debiasing techniques across multiple standard benchmarks, including UrbanCars, BAR, CelebA, and ImageNet-1K. Code, pre-trained models, and training logs are available at https://kadarsh22.github.io/sebra_iclr25/.

Mix-LN: Unleashing the Power of Deeper Layers by Combining Pre-LN and Post-LN

Large Language Models (LLMs) have achieved remarkable success, yet recent findings reveal that their deeper layers often contribute minimally and can be pruned without affecting overall performance. While some view this as an opportunity for model compression, we identify it as a training shortfall rooted in the widespread use of Pre-Layer Normalization (Pre-LN). We demonstrate that Pre-LN, commonly employed in models like GPT and LLaMA, leads to diminished gradient norms in its deeper layers, reducing their effectiveness. In contrast, Post-Layer Normalization (Post-LN) preserves larger gradient norms in deeper layers but suffers from vanishing gradients in earlier layers. To address this, we introduce Mix-LN, a novel normalization technique that combines the strengths of Pre-LN and Post-LN within the same model. Mix-LN applies Post-LN to the earlier layers and Pre-LN to the deeper layers, ensuring more uniform gradients across layers. This allows all parts of the network--both shallow and deep layers--to contribute effectively to training. Extensive experiments with various model sizes from 70M to 7B demonstrate that Mix-LN consistently outperforms both Pre-LN and Post-LN, promoting more balanced, healthier gradient norms throughout the network, and enhancing the overall quality of LLM pre-training. Furthermore, we demonstrate that models pre-trained with Mix-LN learn better compared to those using Pre-LN or Post-LN during supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF), highlighting the critical importance of high-quality deep layers. By effectively addressing the inefficiencies of deep layers in current LLMs, Mix-LN unlocks their potential, enhancing model capacity without increasing model size. Our code is available at https://github.com/pixeli99/MixLN.

Aspect-Based Few-Shot Learning

We generalize the formulation of few-shot learning by introducing the concept of an aspect. In the traditional formulation of few-shot learning, there is an underlying assumption that a single "true" label defines the content of each data point. This label serves as a basis for the comparison between the query object and the objects in the support set. However, when a human expert is asked to execute the same task without a predefined set of labels, they typically consider the rest of the data points in the support set as context. This context specifies the level of abstraction and the aspect from which the comparison can be made. In this work, we introduce a novel architecture and training procedure that develops a context given the query and support set and implements aspect-based few-shot learning that is not limited to a predetermined set of classes. We demonstrate that our method is capable of forming and using an aspect for few-shot learning on the Geometric Shapes and Sprites dataset. The results validate the feasibility of our approach compared to traditional few-shot learning.

Condense, Don't Just Prune: Enhancing Efficiency and Performance in MoE Layer Pruning

Mixture-of-Experts (MOE) has garnered significant attention for their ability to scale up neural networks while utilizing the same or even fewer active parameters. However, MoE does not relieve the massive memory requirements of networks, which limits their practicality in real-world applications, especially in the era of large language models (LLMs). While recent work explores the possibility of removing entire layers of MoE to reduce memory, the performance degradation is still notable. In this paper, we propose Condense-MoE (CD-MoE} that, instead of dropping the entire MoE layer, condenses the big, sparse MoE layer into a small but dense layer with only a few experts that are activated for all tokens. Our approach is specifically designed for fine-grained MoE with shared experts, where Feed-Forward Networks are split into many small experts, with certain experts isolated to serve as shared experts that are always activated. We demonstrate the effectiveness of our method across multiple MoE models such as DeepSeekMoE and QwenMoE on various benchmarks. Specifically, for the DeepSeekMoE-16B model, our approach maintains nearly 90% of the average accuracy while reducing memory usage by 30% and enhancing inference speed by 30%. Moreover, we show that with lightweight expert fine-tuning, the pruned model can achieve further improvements on specific tasks. Our code are available at https://github.com/duterscmy/CD-MoE/tree/main.

Pushing the Limits of Sparsity: A Bag of Tricks for Extreme Pruning

Pruning of deep neural networks has been an effective technique for reducing model size while preserving most of the performance of dense networks, crucial for deploying models on memory and power-constrained devices. While recent sparse learning methods have shown promising performance up to moderate sparsity levels such as 95% and 98%, accuracy quickly deteriorates when pushing sparsities to extreme levels. Obtaining sparse networks at such extreme sparsity levels presents unique challenges, such as fragile gradient flow and heightened risk of layer collapse. In this work, we explore network performance beyond the commonly studied sparsities, and propose a collection of techniques that enable the continuous learning of networks without accuracy collapse even at extreme sparsities, including 99.90%, 99.95% and 99.99% on ResNet architectures. Our approach combines 1) Dynamic ReLU phasing, where DyReLU initially allows for richer parameter exploration before being gradually replaced by standard ReLU, 2) weight sharing which reuses parameters within a residual layer while maintaining the same number of learnable parameters, and 3) cyclic sparsity, where both sparsity levels and sparsity patterns evolve dynamically throughout training to better encourage parameter exploration. We evaluate our method, which we term Extreme Adaptive Sparse Training (EAST) at extreme sparsities using ResNet-34 and ResNet-50 on CIFAR-10, CIFAR-100, and ImageNet, achieving significant performance improvements over state-of-the-art methods we compared with.

OWLed: Outlier-weighed Layerwise Pruning for Efficient Autonomous Driving Framework

The integration of Large Language Models (LLMs) into autonomous driving systems offers promising enhancements in environmental understanding and decision-making. However, the substantial computational demands of deploying LLMs locally on vehicles render this approach unfeasible for real-world automotive applications. To address this challenge, we introduce OWLed, the Outlier-Weighed Layerwise Pruning for Efficient Autonomous Driving Framework that leverages outlier-weighted layerwise sparsity for model compression. Our method assigns non-uniform sparsity ratios to different layers based on the distribution of outlier features, significantly reducing the model size without the need for fine-tuning. To ensure the compressed model adapts well to autonomous driving tasks, we incorporate driving environment data into both the calibration and pruning processes. Our empirical studies reveal that the encoder component is more sensitive to pruning than the LLM, highlighting its critical role in the system. Experimental results demonstrate that OWLed outperforms existing methods in perception, action prediction, and language understanding while substantially lowering computational requirements. These findings underscore the potential of combining advanced pruning techniques with LLMs to develop efficient and robust autonomous driving systems capable of handling complex scenarios. Code will be made publicly available.

Multimodal Contrastive Learning of Urban Space Representations from POI Data

Existing methods for learning urban space representations from Point-of-Interest (POI) data face several limitations, including issues with geographical delineation, inadequate spatial information modelling, underutilisation of POI semantic attributes, and computational inefficiencies. To address these issues, we propose CaLLiPer (Contrastive Language-Location Pre-training), a novel representation learning model that directly embeds continuous urban spaces into vector representations that can capture the spatial and semantic distribution of urban environment. This model leverages a multimodal contrastive learning objective, aligning location embeddings with textual POI descriptions, thereby bypassing the need for complex training corpus construction and negative sampling. We validate CaLLiPer's effectiveness by applying it to learning urban space representations in London, UK, where it demonstrates 5-15% improvement in predictive performance for land use classification and socioeconomic mapping tasks compared to state-of-the-art methods. Visualisations of the learned representations further illustrate our model's advantages in capturing spatial variations in urban semantics with high accuracy and fine resolution. Additionally, CaLLiPer achieves reduced training time, showcasing its efficiency and scalability. This work provides a promising pathway for scalable, semantically rich urban space representation learning that can support the development of geospatial foundation models. The implementation code is available at https://github.com/xlwang233/CaLLiPer.

TODO: Enhancing LLM Alignment with Ternary Preferences

Aligning large language models (LLMs) with human intent is critical for enhancing their performance across a variety of tasks. Standard alignment techniques, such as Direct Preference Optimization (DPO), often rely on the binary Bradley-Terry (BT) model, which can struggle to capture the complexities of human preferences -- particularly in the presence of noisy or inconsistent labels and frequent ties. To address these limitations, we introduce the Tie-rank Oriented Bradley-Terry model (TOBT), an extension of the BT model that explicitly incorporates ties, enabling more nuanced preference representation. Building on this, we propose Tie-rank Oriented Direct Preference Optimization (TODO), a novel alignment algorithm that leverages TOBT's ternary ranking system to improve preference alignment. In evaluations on Mistral-7B and Llama 3-8B models, TODO consistently outperforms DPO in modeling preferences across both in-distribution and out-of-distribution datasets. Additional assessments using MT Bench and benchmarks such as Piqa, ARC-c, and MMLU further demonstrate TODO's superior alignment performance. Notably, TODO also shows strong results in binary preference alignment, highlighting its versatility and potential for broader integration into LLM alignment. The implementation details can be found in https://github.com/XXares/TODO.

Full-Rank No More: Low-Rank Weight Training for Modern Speech Recognition Models

This paper investigates the under-explored area of low-rank weight training for large-scale Conformer-based speech recognition models from scratch. Our study demonstrates the viability of this training paradigm for such models, yielding several notable findings. Firstly, we discover that applying a low-rank structure exclusively to the attention modules can unexpectedly enhance performance, even with a significant rank reduction of 12%. In contrast, feed-forward layers present greater challenges, as they begin to exhibit performance degradation with a moderate 50% rank reduction. Furthermore, we find that both initialization and layer-wise rank assignment play critical roles in successful low-rank training. Specifically, employing SVD initialization and linear layer-wise rank mapping significantly boosts the efficacy of low-rank weight training. Building on these insights, we introduce the Low-Rank Speech Model from Scratch (LR-SMS), an approach that achieves performance parity with full-rank training while delivering substantial reductions in parameters count (by at least 2x), and training time speedups (by 1.3x for ASR and 1.15x for AVSR).