arXiv subjects

Qingfu Zhu

Publications and source records attributed to Qingfu Zhu.

At least 37 records · Page 2Linked to original sources

arXiv2026

ProxyAttn: Guided Sparse Attention via Representative Heads

The quadratic complexity of attention mechanisms limits the efficiency of Large Language Models (LLMs) on long-text tasks. Recently, methods that dynamically estimate block importance have enabled efficient block sparse attention, leading to significant acceleration in long-text pre-filling of LLMs. However, their coarse-grained estimation inevitably leads to performance degradation at high sparsity rates. In this work, we propose ProxyAttn, a training-free sparse attention algorithm that achieves more precise block estimation by compressing the dimension of attention heads. Based on our observation of the similarity among multiple attention heads, we use the scores of pooled representative heads to approximate the scores for all heads. To account for the varying sparsity among heads, we also propose a block-aware dynamic budget estimation method. By combining the scores from representative proxy heads with multi-head dynamic budgets, we achieve a more fine-grained block importance evaluation at low computational cost. Experiments on a variety of mainstream models and extensive benchmarks confirm the underlying similarity among attention heads. Leveraging a fine-grained estimation, the proposed method achieves substantial gains in performance and efficiency compared to existing methods. More precisely, ProxyAttn can achieve up to 10.3x attention acceleration and 2.4x prefilling acceleration without significant performance loss. Our code is available at https://github.com/wyxstriker/ProxyAttn.

cs.CL↗

arXiv2026

Is Compression Really Linear with Code Intelligence?

Understanding the relationship between data compression and the capabilities of Large Language Models (LLMs) is crucial, especially in specialized domains like code intelligence. Prior work posited a linear relationship between compression and general intelligence. However, it overlooked the multifaceted nature of code that encompasses diverse programming languages and tasks, and struggled with fair evaluation of modern Code LLMs. We address this by evaluating a diverse array of open-source Code LLMs on comprehensive multi-language, multi-task code benchmarks. To address the challenge of efficient and fair evaluation of pre-trained LLMs' code intelligence, we introduce \textit{Format Annealing}, a lightweight, transparent training methodology designed to assess the intrinsic capabilities of these pre-trained models equitably. Compression efficacy, measured as bits-per-character (BPC), is determined using a novel, large-scale, and previously unseen code validation set derived from GitHub. Our empirical results reveal a fundamental logarithmic relationship between measured code intelligence and BPC. This finding refines prior hypotheses of linearity, which we suggest are likely observations of the logarithmic curve's tail under specific, limited conditions. Our work provides a more nuanced understanding of compression's role in developing code intelligence and contributes a robust evaluation framework in the code domain.

cs.CL↗

arXiv2026

How Many Code and Test Cases Are Enough? Evaluating Test Cases Generation from a Binary-Matrix Perspective

Evaluating test cases automatically generated by Large Language Models (LLMs) is a critical yet challenging task. Existing benchmarks often evaluate the exclusion ratio on large, unstructured collections of wrong codes, suffering from high computational costs and score inflation. Furthermore, they inadvertently reward generators that detect common, trivial bugs, while failing to penalize their inability to identify rare yet critical faults. In this work, we connect two fundamental questions: (1) What is the minimal set of wrong codes sufficient to represent the entire error space? and (2) What is the minimal set of test cases needed to distinguish them? We introduce a novel framework that formalizes benchmark construction as finding an optimal diagnostic basis in a binary code-test matrix, where rows represent wrong codes and columns represent test case results. The rank of this matrix specifies the minimal number of independent error patterns (wrong codes) and provides a tight upper bound on the number of test cases required for complete fault coverage. Our objective is to identify a basis of size equal to the matrix rank that maximizes internal diversity. To tackle this NP-hard problem, we propose WrongSelect, an efficient approximation algorithm to select maximally diverse wrong codes. Applying this framework to millions of competitive programming submissions, we construct TC-Bench, a compact, diverse, and inflation-resistant benchmark. Extensive experiments show that even the most advanced test case generation methods achieve only ~60% exclusion rates on TC-Bench, exposing a significant gap in their diagnostic power and highlighting substantial room for future improvement. Our dataset is available at: https://huggingface.co/datasets/Luoberta/TC-Bench and our code is at: https://github.com/Luowaterbi/TC-Bench.

cs.CL↗

arXiv2026

When Does Context Help? Error Dynamics of Contextual Information in Large Language Models

Contextual information at inference time, such as demonstrations, retrieved knowledge, or interaction history, can substantially improve large language models (LLMs) without parameter updates, yet its theoretical role remains poorly understood beyond specific settings such as in-context learning (ICL). We present a unified theoretical framework for analyzing the effect of arbitrary contextual information in Transformer-based LLMs. Our analysis characterizes contextual influence through output error dynamics. In a single-layer Transformer, we prove that the context-conditioned error vector decomposes additively into the baseline error vector and a contextual correction vector. This yields necessary geometric conditions for error reduction: the contextual correction must align with the negative baseline error and satisfy a norm constraint. We further show that the contextual correction norm admits an explicit upper bound determined by context-query relevance and complementarity. These results extend to multi-context and multi-layer Transformers. Experiments across ICL, retrieval-augmented generation, and memory evolution validate our theory and motivate a principled context selection strategy that improves performance by $0.6\%$.

cs.CL↗

arXiv2026

How Do Language Models Understand Tables? A Mechanistic Analysis of Cell Location

While Large Language Models (LLMs) are increasingly deployed for table-related tasks, the internal mechanisms enabling them to process linearized two-dimensional structured tables remain opaque. In this work, we investigate the process of table understanding by dissecting the atomic task of cell location. Through activation patching and complementary interpretability techniques, we delineate the table understanding mechanism into a sequential three-stage pipeline: Semantic Binding, Coordinate Localization, and Information Extraction. We demonstrate that models locate the target cell via an ordinal mechanism that counts discrete delimiters to resolve coordinates. Furthermore, column indices are encoded within a linear subspace that allows for precise steering of model focus through vector arithmetic. Finally, we reveal that models generalize to multi-cell location tasks by multiplexing the identical attention heads identified during atomic location. Our findings provide a comprehensive explanation of table understanding within Transformer architectures.

cs.CL↗

arXiv2026

Seer Self-Consistency: Advance Budget Estimation for Adaptive Test-Time Scaling

Test-time scaling improves the inference performance of Large Language Models (LLMs) but also incurs substantial computational costs. Although recent studies have reduced token consumption through dynamic self-consistency, they remain constrained by the high latency of sequential requests. In this paper, we propose SeerSC, a dynamic self-consistency framework that simultaneously improves token efficiency and latency by integrating System 1 and System 2 reasoning. Specifically, we utilize the rapid System 1 to compute the answer entropy for given queries. This score is then used to evaluate the potential of samples for scaling, enabling dynamic self-consistency under System 2. Benefiting from the advance and accurate estimation provided by System 1, the proposed method can reduce token usage while simultaneously achieving a significant decrease in latency through parallel generation. It outperforms existing methods, achieving up to a 47% reduction in token consumption and a 43% reduction in inference latency without significant performance loss.

cs.CL↗

arXiv2026

Judge Q: Trainable Queries for Optimized Information Retention in KV Cache Eviction

Large language models (LLMs) utilize key-value (KV) cache to store historical information during sequence processing. The size of KV cache grows linearly as the length of the sequence extends, which seriously affects memory usage and decoding efficiency. Current methods for KV cache eviction typically utilize the last window from the pre-filling phase as queries to compute the KV importance scores for eviction. Although this scheme is simple to implement, it tends to overly focus on local information, potentially leading to the neglect or omission of crucial global information. To mitigate this issue, we propose Judge Q, a novel training method which incorporates a soft token list. This method only tunes the model's embedding layer at a low training cost. By concatenating the soft token list at the end of the input sequence, we train these tokens' attention map to the original input sequence to align with that of the actual decoded tokens. In this way, the queries corresponding to the soft tokens can effectively capture global information and better evaluate the importance of the keys and values within the KV cache, thus maintaining decoding quality when KV cache is evicted. Under the same eviction budget, our method exhibits less performance degradation compared to existing eviction approaches. We validate our approach through experiments conducted on models such as Llama-3.1-8B-Instruct and Mistral-7B-Instruct-v0.3, using benchmarks including LongBench, RULER, and Needle-in-a-Haystack. Results indicate an improvement of approximately 1 point on the LongBench and over 3 points on RULER. This proposed methodology can be seamlessly integrated into existing open-source models with minimal training overhead, thereby enhancing performance in KV cache eviction scenarios.

cs.CL↗

arXiv2025

CAMERA: Multi-Matrix Joint Compression for MoE Models via Micro-Expert Redundancy Analysis

Large Language Models (LLMs) with Mixture-of-Experts (MoE) architectures are distinguished by their strong performance scaling with increasing parameters across a wide range of tasks, yet they also suffer from substantial computational and storage overheads. Notably, the performance gains of MoE models do not scale proportionally with the growth in expert parameters. While prior works attempt to reduce parameters via expert-level pruning, merging, or decomposition, they still suffer from challenges in both performance and computational efficiency. In this paper, we address these challenges by introducing micro-expert as a finer-grained compression unit that spans across matrices. We first establish a more fundamental perspective, viewing MoE layers as mixtures of micro-experts, and present CAMERA, a lightweight and training-free framework for identifying micro-expert redundancy. Our analysis uncovers significant variance in micro-expert contributions during decoding. Based on this insight, we further propose CAMERA-P, a structured micro-expert pruning framework, and CAMERA-Q, a mixed-precision quantization idea designed for micro-experts. Extensive experiments on nine downstream tasks show that CAMERA-P consistently outperforms strong baselines under pruning ratios ranging from 20% to 60%. Furthermore, CAMERA-Q achieves superior results under aggressive 2-bit quantization, surpassing existing matrix- and channel-level ideas. Notably, our method enables complete micro-expert analysis of Qwen2-57B-A14B in less than 5 minutes on a single NVIDIA A100-40GB GPU.

cs.CL↗

arXiv2025

Lookahead Q-Cache: Achieving More Consistent KV Cache Eviction via Pseudo Query

Large language models (LLMs) rely on key-value cache (KV cache) to accelerate decoding by reducing redundant computations. However, the KV cache memory usage grows substantially with longer text sequences, posing challenges for efficient deployment. Existing KV cache eviction methods prune tokens using prefilling-stage attention scores, causing inconsistency with actual inference queries, especially under tight memory budgets. In this paper, we propose Lookahead Q-Cache (LAQ), a novel eviction framework that generates low-cost pseudo lookahead queries to better approximate the true decoding-stage queries. By using these lookahead queries as the observation window for importance estimation, LAQ achieves more consistent and accurate KV cache eviction aligned with real inference scenarios. Experimental results on LongBench and Needle-in-a-Haystack benchmarks show that LAQ outperforms existing methods across various budget levels, achieving a 1 $\sim$ 4 point improvement on LongBench under limited cache budget. Moreover, LAQ is complementary to existing approaches and can be flexibly combined to yield further improvements.

cs.CL↗

arXiv2025

Automated Snippet-Alignment Data Augmentation for Code Translation

Code translation aims to translate the code from its source language to the target language and is used in various software development scenarios. Recent developments in Large Language Models (LLMs) have showcased their capabilities in code translation, and parallel corpora play a crucial role in training models for code translation. Parallel corpora can be categorized into program-alignment (PA) and snippet-alignment (SA) data. Although PA data has complete context and is suitable for semantic alignment learning, it may not provide adequate fine-grained training signals due to its extended length, while the brevity of SA data enables more fine-grained alignment learning. Due to limited parallel corpora, researchers explore several augmentation methods for code translation. Previous studies mainly focus on augmenting PA data. In this paper, we propose a data augmentation method that leverages LLMs to generate SA data automatically. To fully leverage both PA data and SA data, we explore a simple yet effective two-stage training strategy, which consistently enhances model performance compared to fine-tuning solely on PA data. Experiments on TransCoder-test demonstrate that our augmented SA data combined with the two-stage training approach yields consistent improvements over the baseline, achieving a maximum gain of 3.78% on pass@k.

cs.SE↗

arXiv2025

Bounds of Chain-of-Thought Robustness: Reasoning Steps, Embed Norms, and Beyond

Existing research indicates that the output of Chain-of-Thought (CoT) is significantly affected by input perturbations. Although many methods aim to mitigate such impact by optimizing prompts, a theoretical explanation of how these perturbations influence CoT outputs remains an open area of research. This gap limits our in-depth understanding of how input perturbations propagate during the reasoning process and hinders further improvements in prompt optimization methods. Therefore, in this paper, we theoretically analyze the effect of input perturbations on the fluctuation of CoT outputs. We first derive an upper bound for input perturbations under the condition that the output fluctuation is within an acceptable range, based on which we prove that: (i) This upper bound is positively correlated with the number of reasoning steps in the CoT; (ii) Even an infinitely long reasoning process cannot eliminate the impact of input perturbations. We then apply these conclusions to the Linear Self-Attention (LSA) model, which can be viewed as a simplified version of the Transformer. For the LSA model, we prove that the upper bound for input perturbation is negatively correlated with the norms of the input embedding and hidden state vectors. To validate this theoretical analysis, we conduct experiments on three mainstream datasets and four mainstream models. The experimental results align with our theoretical analysis, empirically demonstrating the correctness of our findings.

cs.CL↗

arXiv2025

CommonKV: Compressing KV Cache with Cross-layer Parameter Sharing

Large Language Models (LLMs) confront significant memory challenges due to the escalating KV cache with increasing sequence length. As a crucial technique, existing cross-layer KV cache sharing methods either necessitate modified model architectures with subsequent pre-training or incur significant performance degradation at high compression rates. To mitigate these challenges, we propose CommonKV, a training-free method for cross-layer KV cache compression through adjacent parameters sharing. Inspired by the high similarity observed in cross-layer hidden states, we utilize Singular Value Decomposition (SVD) to achieve weight sharing across adjacent parameters, resulting in a more easily mergeable latent KV cache. Furthermore, we also introduce an adaptive budget allocation strategy. It dynamically assigns compression budgets based on cosine similarity, ensuring that dissimilar caches are not over-compressed. Experiments across multiple backbone models and benchmarks including LongBench and Ruler demonstrate that the proposed method consistently outperforms existing low-rank and cross-layer approaches at various compression ratios. Moreover, we find that the benefits of CommonKV are orthogonal to other quantization and eviction methods. By integrating these approaches, we can ultimately achieve a 98\% compression ratio without significant performance loss.

cs.LG↗

arXiv2025

Multi-Layer Attention is the Amplifier of Demonstration Effectiveness

Numerous studies have investigated the underlying mechanisms of in-context learning (ICL) effectiveness to inspire the design of related methods. However, existing work predominantly assumes the effectiveness of the demonstrations provided within ICL, while many research indicates that not all demonstrations are effective, failing to yielding any performance improvement during ICL. Therefore, in this paper, we investigate the reasons behind demonstration ineffectiveness. Our analysis is based on gradient flow and linear self-attention models. By setting the gradient flow to zero, we deduce that a demonstration becomes ineffective if its information has either been learned by the model or is irrelevant to the user query. Furthermore, we demonstrate that in multi-layer models, the disparity in effectiveness among demonstrations is amplified with layer increasing, causing the model to focus more on effective ones. Considering that current demonstration selection methods primarily focus on the relevance to the user query while overlooking the information that the model has already assimilated, we propose a novel method called GradS, which leverages gradient flow for demonstration selection. We use the magnitude of the gradient flow of the demonstration with respect to a given user query as the criterion, thereby ensuring the effectiveness of the chosen ones. We validate our derivation and GradS on four prominent LLMs across five mainstream datasets. The experimental results confirm that the disparity in effectiveness among demonstrations is magnified as the model layer increases, substantiating our derivations. Moreover, GradS achieves a relative improvement of $6.8\%$ on average over the strongest baselines, demonstrating its effectiveness.

cs.CL↗

arXiv2025

Learning-to-Context Slope: Evaluating In-Context Learning Effectiveness Beyond Performance Illusions

In-context learning (ICL) has emerged as an effective approach to enhance the performance of large language models (LLMs). However, its effectiveness varies significantly across models and tasks, posing challenges for practitioners to determine when ICL reliably improves performance. Current evaluation approaches, reliant on performance change after applying ICL, suffer from low reliability, poor attribution, and impracticality in data-insufficient scenarios. We propose the Learning-to-Context Slope (LCS), a novel metric that quantifies ICL effectiveness by modeling the slope between learning gain (loss decrease from demonstrations) and contextual relevance (demonstration-input relevance). LCS addresses key limitations of performance-based metrics: (1) it captures continuous loss changes even when outputs are incorrect, improving reliability; (2) its formulation attributes ICL failures to weak contextual alignment (inability to adapt inputs to demonstrations) or strong output calibration (self-verification of correctness); and (3) it minimizes reliance on labeled data via synthetic evaluation. Extensive experiments demonstrate that LCS strongly correlates with performance improvements in labeled settings and reliably reflects true effectiveness in biased or data-scarce scenarios. Further analysis reveals actionable thresholds for LCS and identifies model capabilities critical to ICL success.

cs.CL↗

arXiv2025

A Survey on Transformer Context Extension: Approaches and Evaluation

Large language models (LLMs) based on Transformer have been widely applied in the filed of natural language processing (NLP), demonstrating strong performance, particularly in handling short text tasks. However, when it comes to long context scenarios, the performance of LLMs degrades due to some challenges. To alleviate this phenomenon, there is a number of work proposed recently. In this survey, we first list the challenges of applying pre-trained LLMs to process long contexts. Then systematically review the approaches related to long context and propose our taxonomy categorizing them into four main types: positional encoding, context compression, retrieval augmented, and attention pattern. In addition to the approaches, we focus on the evaluation of long context, organizing relevant data, tasks, and metrics based on existing long context benchmarks. Finally, we summarize unresolved issues in the long context domain and put forward our views on future developments.

cs.CL↗

arXiv2025

Format-Adapter: Improving Reasoning Capability of LLMs by Adapting Suitable Format

Generating and voting multiple answers is an effective method to mitigate reasoning inconsistencies of large language models (LLMs). Prior works have shown that multiple reasoning formats outperform a single format when generating multiple answers. However, previous works using multiple formats rely on formats labeled by humans, which could be unsuitable for all tasks and have high labeling costs. To address this issue, we adapt suitable formats to the given tasks by generating and selecting formats. We first propose how to measure the reasoning error when generating multiple answers. Then, we introduce Format-Adapter, which utilizes LLMs to generate and select suitable reasoning formats by minimizing the error measurement we present. We conduct experiments on math and commonsense reasoning tasks, where Format-Adapter achieves a 4.3% performance improvement on average over previous works, demonstrating the effectiveness.

cs.CL↗

arXiv2025

Advancing Tool-Augmented Large Language Models via Meta-Verification and Reflection Learning

Empowering large language models (LLMs) with effective tool utilization capabilities is crucial for enabling AI agents to solve complex problems. However, current models face two major limitations: (1) unreliable tool planning and invocation due to low-quality instruction datasets (e.g., widespread hallucinated API calls), and (2) weak tool reflection abilities (over 90% of errors cannot be corrected) resulting from static imitation learning. To address these critical limitations, we propose Tool-MVR, a novel Tool-Augmented LLM that achieves comprehensive System 2 reasoning through two key innovations. Specifically, we first introduce Multi-Agent Meta-Verification (MAMV), a systematic pipeline that rigorously validates APIs, queries, and reasoning trajectories to construct ToolBench-V, a new high-quality instruction dataset that addresses the limitation of unreliable tool planning and invocation. Second, we propose Exploration-based Reflection Learning (EXPLORE), which enhances tool reflection capabilities by leveraging tool feedback through a dynamic "Error -> Reflection -> Correction" learning paradigm, resulting in our reflection dataset ToolBench-R and addressing the critical weakness in tool reflection. Finally, we obtain Tool-MVR by finetuning open-source LLMs (e.g., Qwen-7B) on both ToolBench-V and ToolBench-R. Our experiments demonstrate that Tool-MVR achieves state-of-the-art performance on StableToolBench, surpassing both ToolLLM (by 23.9%) and GPT-4 (by 15.3%) while reducing API calls by 31.4%, with strong generalization capabilities across unseen tools and scenarios. Additionally, on our proposed RefineToolBench, the first benchmark specifically designed to evaluate tool reflection capabilities, Tool-MVR achieves a 58.9% error correction rate, significantly outperforming ToolLLM's 9.1%.

cs.CL↗

arXiv2025

Tag-Evol: Achieving Efficient Instruction Evolving via Tag Injection

Evol-Instruct has made significant improvements as a data synthesis method in several areas. Existing methods typically rely on a fixed set of strategies to evolve, which require manual design and are monolithic in form. In addition, iterative evolution also makes the acquisition of hard samples expensive. In view of this, we propose the Tag-Evol framework, a more diverse and efficient instruction evolving method. Specifically, Tag-Evol uses diverse and specific knowledge tags as strategies to achieve controlled evolution by injecting different combinations of tags into the original instructions. Experiments with multiple backbones in diverse domain benchmarks show that the proposed method generates significantly better evolved data than other methods. Furthermore, we conduct a thorough analysis of the evolved data, demonstrating that Tag-Evol is not only efficient but also generates more diverse and challenging data.

cs.CL↗