Early Exit Is a Natural Capability in Transformer-based Models: An Empirical Study on Early Exit without Joint Optimization

Large language models (LLMs) exhibit exceptional performance across various downstream tasks. However, they encounter limitations due to slow inference speeds stemming from their extensive parameters. The early exit (EE) is an approach that aims to accelerate auto-regressive decoding. EE generates outputs from intermediate layers instead of using the whole model, which offers a promising solution to this challenge. However, additional output layers and joint optimization used in conventional EE hinder the application of EE in LLMs. In this paper, we explore the possibility of LLMs EE without additional output layers and joint optimization. Our findings indicate that EE is a natural capability within transformer-based models. While joint optimization does not give model EE capability, it must be employed to address challenges by improving the accuracy of locating the optimal EE layer through gating functions. Additionally, our study reveals patterns in EE behavior from a sub-word perspective based on the LLaMA model and the potential possibility for EE based on sub-layers.

Predictor-Corrector Enhanced Transformers with Exponential Moving Average Coefficient Learning

Residual networks, as discrete approximations of Ordinary Differential Equations (ODEs), have inspired significant advancements in neural network design, including multistep methods, high-order methods, and multi-particle dynamical systems. The precision of the solution to ODEs significantly affects parameter optimization, thereby impacting model performance. In this work, we present a series of advanced explorations of Transformer architecture design to minimize the error compared to the true ``solution.'' First, we introduce a predictor-corrector learning framework to minimize truncation errors, which consists of a high-order predictor and a multistep corrector. Second, we propose an exponential moving average-based coefficient learning method to strengthen our higher-order predictor. Extensive experiments on large-scale machine translation, abstractive summarization, language modeling, and natural language understanding benchmarks demonstrate the superiority of our approach. On the WMT'14 English-German and English-French tasks, our model achieved BLEU scores of 30.95 and 44.27, respectively. Furthermore, on the OPUS multilingual machine translation task, our model surpasses a robust 3.8B DeepNet by an average of 2.9 SacreBLEU, using only 1/3 parameters. Notably, it also beats LLama models by 5.7 accuracy points on the LM Harness Evaluation.

Forgetting Curve: A Reliable Method for Evaluating Memorization Capability for Long-context Models

Numerous recent works target to extend effective context length for language models and various methods, tasks and benchmarks exist to measure model's effective memorization length. However, through thorough investigations, we find limitations for currently existing evaluations on model's memorization capability. We provide an extensive survey for limitations in this work and propose a new method called forgetting curve to measure the memorization capability of long-context models. We show that forgetting curve has the advantage of being robust to the tested corpus and the experimental settings, of not relying on prompts and can be applied to any model size. We apply our forgetting curve to a large variety of models involving both transformer and RNN/SSM based architectures. Our measurement provides empirical evidence for the effectiveness of transformer extension techniques while raises questions for the effective length of RNN/SSM based models. We also examine the difference between our measurement and existing benchmarks as well as popular metrics for various models. Our code and results can be found at https://github.com/1azybug/ForgettingCurve.

LRHP: Learning Representations for Human Preferences via Preference Pairs

To improve human-preference alignment training, current research has developed numerous preference datasets consisting of preference pairs labeled as "preferred" or "dispreferred". These preference pairs are typically used to encode human preferences into a single numerical value through reward modeling, which acts as a reward signal during reinforcement learning from human feedback (RLHF). However, representing these human preferences as a numerical value complicates the analysis of these preferences and restricts their broader applications other than RLHF. In contrast, in this work, we introduce a preference representation learning task that aims to construct a richer and more structured representation of human preferences. We further develop a more generalizable framework, Learning Representations for Human Preferences via preference pairs (namely LRHP), which extends beyond traditional reward modeling to tackle this task. We verify the utility of preference representations in two downstream tasks: preference data selection and preference margin prediction. Building upon the human preferences in representations, we achieve strong performance in both tasks, significantly outperforming baselines.

A Modular-based Strategy for Mitigating Gradient Conflicts in Simultaneous Speech Translation

Simultaneous Speech Translation (SimulST) involves generating target language text while continuously processing streaming speech input, presenting significant real-time challenges. Multi-task learning is often employed to enhance SimulST performance but introduces optimization conflicts between primary and auxiliary tasks, potentially compromising overall efficiency. The existing model-level conflict resolution methods are not well-suited for this task which exacerbates inefficiencies and leads to high GPU memory consumption. To address these challenges, we propose a Modular Gradient Conflict Mitigation (MGCM) strategy that detects conflicts at a finer-grained modular level and resolves them utilizing gradient projection. Experimental results demonstrate that MGCM significantly improves SimulST performance, particularly under medium and high latency conditions, achieving a 0.68 BLEU score gain in offline tasks. Additionally, MGCM reduces GPU memory consumption by over 95\% compared to other conflict mitigation methods, establishing it as a robust solution for SimulST tasks.

More Effective LLM Compressed Tokens with Uniformly Spread Position Identifiers and Compression Loss

Compressing Transformer inputs into compressd tokens allows running LLMs with improved speed and cost efficiency. Based on the compression method ICAE, we carefully examine the position identifier choices for compressed tokens and also propose a new compression loss. We demonstrate empirically that our proposed methods achieve significantly higher compression ratios (15x compared to 4x for ICAE), while being able to attain comparable reconstruction performance.

NDP: Next Distribution Prediction as a More Broad Target

Large language models (LLMs) trained on next-token prediction (NTP) paradigm have demonstrated powerful capabilities. However, the existing NTP paradigm contains several limitations, particularly related to planned task complications and error propagation during inference. In our work, we extend the critique of NTP, highlighting its limitation also due to training with a narrow objective: the prediction of a sub-optimal one-hot distribution. To support this critique, we conducted a pre-experiment treating the output distribution from powerful LLMs as efficient world data compression. By evaluating the similarity between the $n$-gram distribution and the one-hot distribution with LLMs, we observed that the $n$-gram distributions align more closely with the output distribution of LLMs. Based on this insight, we introduce Next Distribution Prediction (NDP), which uses $n$-gram distributions to replace the one-hot targets, enhancing learning without extra online training time. We conducted experiments across translation, general task, language transfer, and medical domain adaptation. Compared to NTP, NDP can achieve up to +2.97 COMET improvement in translation tasks, +0.61 average improvement in general tasks, and incredible +10.75 average improvement in the medical domain. This demonstrates the concrete benefits of addressing the target narrowing problem, pointing to a new direction for future work on improving NTP.

RoVRM: A Robust Visual Reward Model Optimized via Auxiliary Textual Preference Data

Large vision-language models (LVLMs) often fail to align with human preferences, leading to issues like generating misleading content without proper visual context (also known as hallucination). A promising solution to this problem is using human-preference alignment techniques, such as best-of-n sampling and reinforcement learning. However, these techniques face the difficulty arising from the scarcity of visual preference data, which is required to train a visual reward model (VRM). In this work, we continue the line of research. We present a Robust Visual Reward Model (RoVRM) which improves human-preference alignment for LVLMs. RoVRM leverages auxiliary textual preference data through a three-phase progressive training and optimal transport-based preference data selection to effectively mitigate the scarcity of visual preference data. We experiment with RoVRM on the commonly used vision-language tasks based on the LLaVA-1.5-7B and -13B models. Experimental results demonstrate that RoVRM consistently outperforms traditional VRMs. Furthermore, our three-phase progressive training and preference data selection approaches can yield consistent performance gains over ranking-based alignment techniques, such as direct preference optimization.

Cross-layer Attention Sharing for Large Language Models

As large language models (LLMs) evolve, the increase in model depth and parameter number leads to substantial redundancy. To enhance the efficiency of the attention mechanism, previous works primarily compress the KV cache or group attention heads, while largely overlooking redundancy between layers. Our comprehensive analyses across various LLMs show that highly similar attention patterns persist within most layers. It's intuitive to save the computation by sharing attention weights across layers. However, further analysis reveals two challenges: (1) Directly sharing the weight matrix without carefully rearranging the attention heads proves to be ineffective; (2) Shallow layers are vulnerable to small deviations in attention weights. Driven by these insights, we introduce LiSA, a lightweight substitute for self-attention in well-trained LLMs. LiSA employs tiny feed-forward networks to align attention heads between adjacent layers and low-rank matrices to approximate differences in layer-wise attention weights. Evaluations encompassing 13 typical benchmarks demonstrate that LiSA maintains high response quality in terms of accuracy and perplexity while reducing redundant attention calculations within 53-84% of the total layers. Our implementations of LiSA achieve a 6X compression of Q and K, with maximum throughput improvements of 19.5% for LLaMA3-8B and 32.3% for LLaMA2-7B.

Translate-and-Revise: Boosting Large Language Models for Constrained Translation

Imposing constraints on machine translation systems presents a challenging issue because these systems are not trained to make use of constraints in generating adequate, fluent translations. In this paper, we leverage the capabilities of large language models (LLMs) for constrained translation, given that LLMs can easily adapt to this task by taking translation instructions and constraints as prompts. However, LLMs cannot always guarantee the adequacy of translation, and, in some cases, ignore the given constraints. This is in part because LLMs might be overly confident in their predictions, overriding the influence of the constraints. To overcome this overiding behaviour, we propose to add a revision process that encourages LLMs to correct the outputs by prompting them about the constraints that have not yet been met. We evaluate our approach on four constrained translation tasks, encompassing both lexical and structural constraints in multiple constraint domains. Experiments show 15\% improvement in constraint-based translation accuracy over standard LLMs and the approach also significantly outperforms neural machine translation (NMT) state-of-the-art methods.