The attention mechanism is a fundamental component of the Transformer model, contributing to interactions among distinct tokens. In general, the attention scores are determined simply by the key-query products. However, this work’s occasional trial (combining DAPE and NoPE) of including additional MLPs on attention scores without position encoding indicates that the classical key-query multiplication may limit the performance of Transformers. In this work, we conceptualize attention as a feature map and apply the convolution operator (for neighboring attention scores across different heads) to mimic the processing methods in computer vision. Specifically, **the main contribution of this paper is identifying and interpreting the Transformer length extrapolation problem as a result of the limited expressiveness of the naive query and key dot product, and we successfully translate the length extrapolation issue into a well-understood feature map processing problem**, which is called Convolutional Data-Adaptive Position Encoding (CDAPE).The novel insight, which can be adapted to various attention-related models, reveals that the current Transformer architecture has the potential for further evolution. Extensive experiments demonstrate that treating attention as a feature map and applying convolution as a processing method significantly enhances Transformer performance.

The capacity of Large Language Models (LLMs) to comprehend and reason over long contexts is pivotal for advancements in diverse fields. Yet, they still stuggle with capturing long-distance dependencies within sequences to deeply understand semantics. To address this issue, we introduce Query-aware Inference for LLMs (Q-LLM), a system designed to process extensive sequences akin to human cognition. By focusing on memory data relevant to a given query, Q-LLM can accurately capture pertinent information within a fixed window size and provide precise answers to queries. It doesn’t require extra training and can be seamlessly integrated with any LLMs. Q-LLM using LLaMA3 (QuickLLaMA) can read Harry Potter within 30s and accurately answer the questions. On widely recognized benchmarks, Q-LLM improved by 7.17% compared to the current state-of-the-art on LLaMA3, and by 3.26% on Mistral on the ∞-bench. In the Needle-in-a-Haystack and BABILong task, Q-LLM improved upon the current SOTA by 7.0% and 6.1%. Our code is in https://github.com/dvlab-research/Q-LLM.

Deploying large language models (LLMs) for long-context inference remains challenging due to their substantial memory and computational demands. While techniques such as Key-Value (KV) cache compression are designed to reduce memory usage, they often neglect the structured sparsity inherent in the relationship between hidden states and their corresponding KV cache. In this work, we explore the role of uncertainty as a potential indicator of sparsity within LLMs. We propose UNComp, an uncertainty-aware framework that leverages truncated matrix entropy to identify areas of low information content, thereby revealing sparsity patterns that can be used for adaptive compression. Unlike traditional methods that apply uniform compression, UNComp dynamically adjusts its approach to compression, guided by uncertainty measures that reflect the importance of various model components. Our analysis shows that sparsity patterns, when derived from uncertainty estimates, can be exploited to reveal special long-range dependencies, such as retrieval heads and retrieval layers. This perspective not only enhances our understanding of how compression can be optimized but also provides new insights into the inherent sparsity of LLMs during long-context inference. By focusing on uncertainty to analyze the sparsity pattern in detail, UNComp reduces the KV cache size to 4.74% of the original, achieves a 6% prefill speedup, and improves throughput by 6.4× — not only delivering strong lossless compression performance, but also validating the effectiveness of the underlying theoretical tool. Our codes are submitted with the paper.

The softmax function is crucial in Transformer attention, which normalizes each row of the attention scores with summation to one. **Usually, tokens with larger attention scores are important for the final prediction.However, the softmax function can face a gradient vanishing issue for such important tokens (e.g., probabilities close to one), leading to optimization difficulties for the important tokens so that the performance may not be better.**In this paper, we propose Self-Adjust Softmax (SA-Softmax) to address this issue by modifying softmax(z) to z ⋅ softmax(z) and its normalized variant ^{(z - min(z_\min,0))}⁄_{max(0,zmax)-min(zmin,0)} ⋅ softmax(z).We theoretically show that SA-Softmax provides enhanced gradient properties compared to the vanilla softmax function.Moreover, Attention can be seamlessly integrated into existing Transformer models to their attention mechanisms with minor adjustments.We conducted experiments to evaluate the empirical performance of Transformer models using compared to the vanilla softmax function. These experiments, involving models with up to 2.7 billion parameters, are conducted across diverse datasets, language tasks, and positional encoding methods.

In recent years, developing compact and efficient large language models (LLMs) has emerged as a thriving area of research. However, traditional Supervised Fine-Tuning (SFT), which relies on singular ground truth labels, often fails to capture token-level dependencies and linguistic diversity. To address these limitations, we propose a logits-based fine-tuning framework that integrates the strengths of supervised learning and knowledge distillation. Our approach constructs enriched training targets by combining teacher logits with ground truth labels, preserving both correctness and linguistic diversity. This ensures more reliable and effective training. To validate our approach, we constructed a large-scale 1.2M logits dataset and trained a series of science-focused models. Experimental results demonstrate that our method achieves significant improvements over current SOTA, with accuracy gains of 18% on Mawps and 22.7% on TabMWP. Across nine widely used mathematical benchmarks, our method consistently outperforms prior SFT models, achieving an average improvement of 7.28%. All code and datasets will be open-sourced.

Automated theorem proving (ATP) has become an appealing domain for exploring the reasoning ability of the recent successful generative language models. However, current ATP benchmarks are mainly focus on symbolic inference, but rarely involve the understanding of complex number combination reasoning. In this work, we propose TRIGO, an ATP benchmark that not only requires a model to reduce a trigonometric expression with step-by-step proof but also evaluates a generative LM’s reasoning ability on formulas and capability to manipulate, group, and factor number terms. We gather trigonometric expressions and their reduced forms from web, annotate the simplification process manually, and translate it into the “Lean” formal language system. We then automatically generate additional examples from the annotated samples to expand the dataset. Furthermore, we also create three automatically generated training and testing datasets of varying difficulty and distributions. Our extensive experiments show our proposed TRIGO poses a new challenge for advanced generative LM’s including GPT-4 which is pre-trained on a considerable amount of open-source formal theorem-proving language data, and provide a new tool to study the generative LM’s ability on both formal and mathematical reasoning.