Recent studies show that in supervised fine-tuning (SFT) of large language models (LLMs), data quality matters more than quantity. While most data cleaning methods concentrate on filtering entire samples, the quality of individual tokens within a sample can vary significantly. After pre-training, even in high-quality samples, patterns or phrases that are not task-related can be redundant or uninformative. Continuing to fine-tune on these patterns may offer limited benefit and even degrade downstream task performance. In this paper, we investigate token quality from a noisy-label perspective and propose a generic token cleaning pipeline for SFT tasks. Our method filters out uninformative tokens while preserving those carrying key task-specific information. Specifically, we first evaluate token quality by examining the influence of model updates on each token, then apply a threshold-based separation. The token influence can be measured in a single pass with a fixed reference model or iteratively with self-evolving reference models. The benefits and limitations of both methods are analyzed theoretically by error upper bounds. Extensive experiments show that our framework consistently improves performance across multiple downstream tasks.
相關內容
With the advancements in long-context inference capabilities of large language models (LLMs), the KV cache has become one of the foundational components. However, its substantial GPU memory consumption makes KV cache compression a key technique for enabling efficient LLM inference in industrial scenarios. While recent studies have focused on optimizing the memory occupied by the KV cache, they overlook two critical factors: preserving semantic coherence and considering task-specific characteristic during compression. To address these limitations, we propose a novel task-adaptive KV cache window selection method, WindowKV. WindowKV dynamically selects local semantic windows consisting of consecutive tokens, according to task-specific characteristics, ensuring the retained KV cache captures continuous, essential context. Additionally, we introduce an intra-group layer KV cache indices sharing strategy to reduce computational overhead, achieving a balance between performance and efficiency. We rigorously evaluate WindowKV on the LongBench benchmark, and the results demonstrate that it maintains a performance comparable to full KV cache retention while using only 12% of the original KV cache, significantly reducing memory requirements. Furthermore, our method also achieves state-of-the-art results in the Needle-in-a-Haystack evaluation, highlighting its effectiveness and robustness.
Multimodal large language models (MLLMs) have demonstrated remarkable capabilities in various tasks. However, effectively evaluating these MLLMs on face perception remains largely unexplored. To address this gap, we introduce FaceBench, a dataset featuring hierarchical multi-view and multi-level attributes specifically designed to assess the comprehensive face perception abilities of MLLMs. Initially, we construct a hierarchical facial attribute structure, which encompasses five views with up to three levels of attributes, totaling over 210 attributes and 700 attribute values. Based on the structure, the proposed FaceBench consists of 49,919 visual question-answering (VQA) pairs for evaluation and 23,841 pairs for fine-tuning. Moreover, we further develop a robust face perception MLLM baseline, Face-LLaVA, by training with our proposed face VQA data. Extensive experiments on various mainstream MLLMs and Face-LLaVA are conducted to test their face perception ability, with results also compared against human performance. The results reveal that, the existing MLLMs are far from satisfactory in understanding the fine-grained facial attributes, while our Face-LLaVA significantly outperforms existing open-source models with a small amount of training data and is comparable to commercial ones like GPT-4o and Gemini. The dataset will be released at //github.com/CVI-SZU/FaceBench.
Large language models (LLMs) have demonstrated potential in assisting scientific research, yet their ability to discover high-quality research hypotheses remains unexamined due to the lack of a dedicated benchmark. To address this gap, we introduce the first large-scale benchmark for evaluating LLMs with a near-sufficient set of sub-tasks of scientific discovery: inspiration retrieval, hypothesis composition, and hypothesis ranking. We develop an automated framework that extracts critical components - research questions, background surveys, inspirations, and hypotheses - from scientific papers across 12 disciplines, with expert validation confirming its accuracy. To prevent data contamination, we focus exclusively on papers published in 2024, ensuring minimal overlap with LLM pretraining data. Our evaluation reveals that LLMs perform well in retrieving inspirations, an out-of-distribution task, suggesting their ability to surface novel knowledge associations. This positions LLMs as "research hypothesis mines", capable of facilitating automated scientific discovery by generating innovative hypotheses at scale with minimal human intervention.
Large Language Models (LLMs) have revolutionized code generation, achieving exceptional results on various established benchmarking frameworks. However, concerns about data contamination - where benchmark data inadvertently leaks into pre-training or fine-tuning datasets - raise questions about the validity of these evaluations. While this issue is known, limiting the industrial adoption of LLM-driven software engineering, hardware coding has received little to no attention regarding these risks. For the first time, we analyze state-of-the-art (SOTA) evaluation frameworks for Verilog code generation (VerilogEval and RTLLM), using established methods for contamination detection (CCD and Min-K% Prob). We cover SOTA commercial and open-source LLMs (CodeGen2.5, Minitron 4b, Mistral 7b, phi-4 mini, LLaMA-{1,2,3.1}, GPT-{2,3.5,4o}, Deepseek-Coder, and CodeQwen 1.5), in baseline and fine-tuned models (RTLCoder and Verigen). Our study confirms that data contamination is a critical concern. We explore mitigations and the resulting trade-offs for code quality vs fairness (i.e., reducing contamination toward unbiased benchmarking).
Spurred by the demand for interpretable models, research on eXplainable AI for language technologies has experienced significant growth, with feature attribution methods emerging as a cornerstone of this progress. While prior work in NLP explored such methods for classification tasks and textual applications, explainability intersecting generation and speech is lagging, with existing techniques failing to account for the autoregressive nature of state-of-the-art models and to provide fine-grained, phonetically meaningful explanations. We address this gap by introducing Spectrogram Perturbation for Explainable Speech-to-text Generation (SPES), a feature attribution technique applicable to sequence generation tasks with autoregressive models. SPES provides explanations for each predicted token based on both the input spectrogram and the previously generated tokens. Extensive evaluation on speech recognition and translation demonstrates that SPES generates explanations that are faithful and plausible to humans.
Recent advancements in instruction tuning for large language models (LLMs) suggest that a small, high-quality dataset can significantly equip LLMs with instruction-following capabilities, outperforming large datasets often burdened by quality and redundancy issues. However, the challenge lies in automatically identifying valuable subsets from large datasets to boost both the effectiveness and efficiency of instruction tuning. In this paper, we first establish data selection criteria based on three distinct aspects of data value: diversity, difficulty, and dependability, and then propose the D3 method comprising two key steps of scoring and selection. Specifically, in the scoring step, we define the diversity function to measure sample distinctiveness and introduce the uncertainty-based prediction difficulty to evaluate sample difficulty by mitigating the interference of context-oriented generation diversity. Additionally, we integrate an external LLM for dependability assessment. In the selection step, we formulate the D3 weighted coreset objective, which jointly optimizes three aspects of data value to solve for the most valuable subset. The two steps of D3 can iterate multiple rounds, incorporating feedback to refine the selection focus adaptively. Experiments on three datasets demonstrate the effectiveness of D3 in endowing LLMs with competitive or even superior instruction-following capabilities using less than 10% of the entire dataset.
Traditional benchmarks struggle to evaluate increasingly sophisticated language models in multilingual and culturally diverse contexts. To address this gap, we introduce MMLU-ProX, a comprehensive multilingual benchmark covering 13 typologically diverse languages with approximately 11,829 questions per language. Building on the challenging reasoning-focused design of MMLU-Pro, our framework employs a semi-automatic translation process: translations generated by state-of-the-art large language models (LLMs) are rigorously evaluated by expert annotators to ensure conceptual accuracy, terminological consistency, and cultural relevance. We comprehensively evaluate 25 state-of-the-art LLMs using 5-shot chain-of-thought (CoT) and zero-shot prompting strategies, analyzing their performance across linguistic and cultural boundaries. Our experiments reveal consistent performance degradation from high-resource languages to lower-resource ones, with the best models achieving over 70% accuracy on English but dropping to around 40% for languages like Swahili, highlighting persistent gaps in multilingual capabilities despite recent advances. MMLU-ProX is an ongoing project; we are expanding our benchmark by incorporating additional languages and evaluating more language models to provide a more comprehensive assessment of multilingual capabilities.
Recent advancements of in-context learning (ICL) show language models can significantly improve their performance when demonstrations are provided. However, little attention has been paid to model calibration and prediction confidence of ICL in cross-lingual scenarios. To bridge this gap, we conduct a thorough analysis of ICL for cross-lingual sentiment classification. Our findings suggest that ICL performs poorly in cross-lingual scenarios, exhibiting low accuracy and presenting high calibration errors. In response, we propose a novel approach, N2C2, which employs a -nearest neighbors augmented classifier for prediction confidence calibration. N2C2 narrows the prediction gap by leveraging a datastore of cached few-shot instances. Specifically, N2C2 integrates the predictions from the datastore and incorporates confidence-aware distribution, semantically consistent retrieval representation, and adaptive neighbor combination modules to effectively utilize the limited number of supporting instances. Evaluation on two multilingual sentiment classification datasets demonstrates that N2C2 outperforms traditional ICL. It surpasses fine tuning, prompt tuning and recent state-of-the-art methods in terms of accuracy and calibration errors.
With the rapid development of large language models (LLMs), fully fine-tuning (FT) these models is becoming increasingly infeasible due to high computational demands. Moreover, FT also increases the risk of catastrophic forgetting. As an alternative, Low-Rank Adaptation (LoRA) has been proposed. By fine-tuning only a small subset of parameters, LoRA achieves performance similar to FT while significantly reducing resource requirements. However, since LoRA inherits FT's design, the issue of catastrophic forgetting still remains. To address these limitations, we propose SECURA: Sigmoid-Enhanced CUR Decomposition LoRA, a novel PEFT variant designed to mitigate catastrophic forgetting while improving fine-tuning performance. Our method introduces a novel normalization technique, Sigmoid-based Magnitude Norm (S-MagNorm), which enhances parameter retention and fine-tuning efficiency. SECURA has been evaluated on a diverse range of tasks, including mathematical problem-solving (GSM8K), complex question-answering (CNNDM), translation (NewsDE), and complex multiple-choice reasoning (LogiQA). Experimental results demonstrate that it achieves an average fine-tuning improvement of 3.59% across four MCQ tasks and 2.51% across five QA tasks on Gemma2 2B, Qwen2 1.5B, Qwen2 7B, Llama3 8B, and Llama3.1 8B, outperforming DoRA. Additionally, SECURA demonstrates superior knowledge retention capabilities, achieving state-of-the-art performance in 16 continual learning tests and maintaining more than 70% accuracy on LLMs' basic knowledge compared to Experience Replay (ER), sequential learning (SEQ), EWC, I-LoRA, and CUR-LoRA.
The emergence of large language models (LLMs) has substantially influenced natural language processing, demonstrating exceptional results across various tasks. In this study, we employ ``Introspective Tips" to facilitate LLMs in self-optimizing their decision-making. By introspectively examining trajectories, LLM refines its policy by generating succinct and valuable tips. Our method enhances the agent's performance in both few-shot and zero-shot learning situations by considering three essential scenarios: learning from the agent's past experiences, integrating expert demonstrations, and generalizing across diverse games. Importantly, we accomplish these improvements without fine-tuning the LLM parameters; rather, we adjust the prompt to generalize insights from the three aforementioned situations. Our framework not only supports but also emphasizes the advantage of employing LLM in in-contxt decision-making. Experiments involving over 100 games in TextWorld illustrate the superior performance of our approach.
小貼士
登錄享
相關主題
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191