Decoder-only language models, such as GPT and LLaMA, generally decode on the last layer. Motivated by human's hierarchical thinking capability, we propose that a hierarchical decoder architecture could be built with different layers decoding texts simultaneously. Due to limited time and computationally resources, we choose to adapt a pretrained language model into this form of hierarchical decoder. Language heads of the last layer are copied to different selected intermediate layers, and fine-tuned with different task inputs. By thorough experiments, we validate that these selective intermediate layers could be adapted to speak meaningful and reasonable contents, and this paradigm of hierarchical decoder can obtain state-of-the-art performances on multiple tasks such as hierarchical text classification, classification-guided generation, and hierarchical text generation. This study suggests the possibility of a generalized hierarchical reasoner, pretraining from scratch.
We extend BeamAttack, an adversarial attack algorithm designed to evaluate the robustness of text classification systems through word-level modifications guided by beam search. Our extensions include support for word deletions and the option to skip substitutions, enabling the discovery of minimal modifications that alter model predictions. We also integrate LIME to better prioritize word replacements. Evaluated across multiple datasets and victim models (BiLSTM, BERT, and adversarially trained RoBERTa) within the BODEGA framework, our approach achieves over a 99\% attack success rate while preserving the semantic and lexical similarity of the original texts. Through both quantitative and qualitative analysis, we highlight BeamAttack's effectiveness and its limitations. Our implementation is available at https://github.com/LucK1Y/BeamAttack
Recently, Large Language Models (LLMs) have demonstrated significant potential for data annotation, markedly reducing the labor costs associated with downstream applications. However, existing methods mostly adopt an aggressive strategy by prompting LLM to determine a single gold label for each unlabeled sample. Due to the inherent uncertainty within LLMs, they often produce incorrect labels for difficult samples, severely compromising the data quality for downstream applications. Motivated by ambiguity aversion in human behaviors, we propose a novel candidate annotation paradigm wherein large language models are encouraged to output all possible labels when incurring uncertainty. To ensure unique labels are provided for downstream tasks, we develop a teacher-student framework CanDist that distills candidate annotations with a Small Language Model (SLM). We further provide a rigorous justification demonstrating that distilling candidate annotations from the teacher LLM offers superior theoretical guarantees compared to directly using single annotations. Extensive experiments across six text classification tasks validate the effectiveness of our proposed method. The source code is available at https://github.com/MingxuanXia/CanDist.
We often attribute human characteristics to large language models (LLMs) and claim that they "know" certain things. LLMs have an internal probabilistic knowledge that represents information retained during training. How can we assess the veracity of this knowledge? We examine two common methods for probing the veracity of LLMs and discover several assumptions that are flawed. To address these flawed assumptions, we introduce sAwMIL (short for Sparse Aware Multiple-Instance Learning), a probing method that utilizes the internal activations of LLMs to separate statements into true, false, and neither. sAwMIL is based on multiple-instance learning and conformal prediction. We evaluate sAwMIL on 5 validity criteria across 16 open-source LLMs, including both default and chat-based variants, as well as on 3 new datasets. Among the insights we provide are: (1) the veracity signal is often concentrated in the third quarter of an LLM's depth; (2) truth and falsehood signals are not always symmetric; (3) linear probes perform better on chat models than on default models; (4) nonlinear probes may be required to capture veracity signals for some LLMs with reinforcement learning from human feedback or knowledge distillation; and (5) LLMs capture a third type of signal that is distinct from true and false and is neither true nor false. These findings provide a reliable method for verifying what LLMs "know" and how certain they are of their probabilistic internal knowledge.
Recent advances in LegalAI have primarily focused on individual case judgment analysis, often overlooking the critical appellate process within the judicial system. Appeals serve as a core mechanism for error correction and ensuring fair trials, making them highly significant both in practice and in research. To address this gap, we present the AppealCase dataset, consisting of 10,000 pairs of real-world, matched first-instance and second-instance documents across 91 categories of civil cases. The dataset also includes detailed annotations along five dimensions central to appellate review: judgment reversals, reversal reasons, cited legal provisions, claim-level decisions, and whether there is new information in the second instance. Based on these annotations, we propose five novel LegalAI tasks and conduct a comprehensive evaluation across 20 mainstream models. Experimental results reveal that all current models achieve less than 50% F1 scores on the judgment reversal prediction task, highlighting the complexity and challenge of the appeal scenario. We hope that the AppealCase dataset will spur further research in LegalAI for appellate case analysis and contribute to improving consistency in judicial decision-making.
Query routing, the task to route user queries to different large language model (LLM) endpoints, can be considered as a text classification problem. However, out-of-distribution queries must be handled properly, as those could be questions about unrelated domains, queries in other languages, or even contain unsafe text. Here, we thus study a \emph{guarded} query routing problem, for which we first introduce the Guarded Query Routing Benchmark (GQR-Bench), which covers three exemplary target domains (law, finance, and healthcare), and seven datasets to test robustness against out-of-distribution queries. We then use GQR-Bench to contrast the effectiveness and efficiency of LLM-based routing mechanisms (GPT-4o-mini, Llama-3.2-3B, and Llama-3.1-8B), standard LLM-based guardrail approaches (LlamaGuard and NVIDIA NeMo Guardrails), continuous bag-of-words classifiers (WideMLP, fastText), and traditional machine learning models (SVM, XGBoost). Our results show that WideMLP, enhanced with out-of-domain detection capabilities, yields the best trade-off between accuracy (88\%) and speed (<4ms). The embedding-based fastText excels at speed (<1ms) with acceptable accuracy (80\%), whereas LLMs yield the highest accuracy (91\%) but are comparatively slow (62ms for local Llama-3.1:8B and 669ms for remote GPT-4o-mini calls). Our findings challenge the automatic reliance on LLMs for (guarded) query routing and provide concrete recommendations for practical applications. GQR-Bench will be released as a Python package -- \texttt{gqr}.
Effective prompt engineering remains a central challenge in fully harnessing the capabilities of LLMs. While well-designed prompts can dramatically enhance performance, crafting them typically demands expert intuition and a nuanced understanding of the task. Moreover, the most impactful prompts often hinge on subtle semantic cues, ones that may elude human perception but are crucial for guiding LLM behavior. In this paper, we introduce PRL (Prompts from Reinforcement Learning), a novel RL-based approach for automatic prompt generation. Unlike previous methods, PRL can produce novel few-shot examples that were not seen during training. Our approach achieves state-of-the-art performance across a range of benchmarks, including text classification, simplification, and summarization. On the classification task, it surpasses prior methods by 2.58% over APE and 1.00% over EvoPrompt. Additionally, it improves the average ROUGE scores on the summarization task by 4.32 over APE and by 2.12 over EvoPrompt and the SARI score on simplification by 6.93 over APE and by 6.01 over EvoPrompt. Our code is available at https://github.com/Batorskq/prl .
Recent research highlights concerns about the trustworthiness of third-party Pre-Trained Language Models (PTLMs) due to potential backdoor attacks. These backdoored PTLMs, however, are effective only for specific pre-defined downstream tasks. In reality, these PTLMs can be adapted to many other unrelated downstream tasks. Such adaptation may lead to unforeseen consequences in downstream model outputs, consequently raising user suspicion and compromising attack stealthiness. We refer to this phenomenon as backdoor complications. In this paper, we undertake the first comprehensive quantification of backdoor complications. Through extensive experiments using 4 prominent PTLMs and 16 text classification benchmark datasets, we demonstrate the widespread presence of backdoor complications in downstream models fine-tuned from backdoored PTLMs. The output distribution of triggered samples significantly deviates from that of clean samples. Consequently, we propose a backdoor complication reduction method leveraging multi-task learning to mitigate complications without prior knowledge of downstream tasks. The experimental results demonstrate that our proposed method can effectively reduce complications while maintaining the efficacy and consistency of backdoor attacks. Our code is available at https://github.com/zhangrui4041/Backdoor_Complications.
Robustness to label noise within data is a significant challenge in federated learning (FL). From the data-centric perspective, the data quality of distributed datasets can not be guaranteed since annotations of different clients contain complicated label noise of varying degrees, which causes the performance degradation. There have been some early attempts to tackle noisy labels in FL. However, there exists a lack of benchmark studies on comprehensively evaluating their practical performance under unified settings. To this end, we propose the first benchmark study FNBench to provide an experimental investigation which considers three diverse label noise patterns covering synthetic label noise, imperfect human-annotation errors and systematic errors. Our evaluation incorporates eighteen state-of-the-art methods over five image recognition datasets and one text classification dataset. Meanwhile, we provide observations to understand why noisy labels impair FL, and additionally exploit a representation-aware regularization method to enhance the robustness of existing methods against noisy labels based on our observations. Finally, we discuss the limitations of this work and propose three-fold future directions. To facilitate related communities, our source code is open-sourced at https://github.com/Sprinter1999/FNBench.
Hierarchical Text Classification (HTC) involves assigning documents to labels organized within a taxonomy. Most previous research on HTC has focused on supervised methods. However, in real-world scenarios, employing supervised HTC can be challenging due to a lack of annotated data. Moreover, HTC often faces issues with large label spaces and long-tail distributions. In this work, we present Knowledge Graphs for zero-shot Hierarchical Text Classification (KG-HTC), which aims to address these challenges of HTC in applications by integrating knowledge graphs with Large Language Models (LLMs) to provide structured semantic context during classification. Our method retrieves relevant subgraphs from knowledge graphs related to the input text using a Retrieval-Augmented Generation (RAG) approach. Our KG-HTC can enhance LLMs to understand label semantics at various hierarchy levels. We evaluate KG-HTC on three open-source HTC datasets: WoS, DBpedia, and Amazon. Our experimental results show that KG-HTC significantly outperforms three baselines in the strict zero-shot setting, particularly achieving substantial improvements at deeper levels of the hierarchy. This evaluation demonstrates the effectiveness of incorporating structured knowledge into LLMs to address HTC's challenges in large label spaces and long-tailed label distributions. Our code is available at: https://github.com/QianboZang/KG-HTC.
Reliance on spurious correlations (shortcuts) has been shown to underlie many of the successes of language models. Previous work focused on identifying the input elements that impact prediction. We investigate how shortcuts are actually processed within the model's decision-making mechanism. We use actor names in movie reviews as controllable shortcuts with known impact on the outcome. We use mechanistic interpretability methods and identify specific attention heads that focus on shortcuts. These heads gear the model towards a label before processing the complete input, effectively making premature decisions that bypass contextual analysis. Based on these findings, we introduce Head-based Token Attribution (HTA), which traces intermediate decisions back to input tokens. We show that HTA is effective in detecting shortcuts in LLMs and enables targeted mitigation by selectively deactivating shortcut-related attention heads.
Large open-source projects receive a large number of issues (known as bugs), including software defect (i.e., bug) reports and new feature requests from their user and developer communities at a fast rate. The often limited project resources do not allow them to deal with all issues. Instead, they have to prioritize them according to the project's priorities and the issues' severities. In this paper, we propose a novel approach to automated bug prioritization based on the natural language text of the bug reports that are stored in the open bug repositories of the issue-tracking systems. We conduct topic modeling using a variant of LDA called TopicMiner-MTM and text classification with the BERT large language model to achieve a higher performance level compared to the state-of-the-art. Experimental results using an existing reference dataset containing 85,156 bug reports of the Eclipse Platform project indicate that we outperform existing approaches in terms of Accuracy, Precision, Recall, and F1-measure of the bug report priority prediction.
This study investigates the self-rationalization framework constructed with a cooperative game, where a generator initially extracts the most informative segment from raw input, and a subsequent predictor utilizes the selected subset for its input. The generator and predictor are trained collaboratively to maximize prediction accuracy. In this paper, we first uncover a potential caveat: such a cooperative game could unintentionally introduce a sampling bias during rationale extraction. Specifically, the generator might inadvertently create an incorrect correlation between the selected rationale candidate and the label, even when they are semantically unrelated in the original dataset. Subsequently, we elucidate the origins of this bias using both detailed theoretical analysis and empirical evidence. Our findings suggest a direction for inspecting these correlations through attacks, based on which we further introduce an instruction to prevent the predictor from learning the correlations. Through experiments on six text classification datasets and two graph classification datasets using three network architectures (GRUs, BERT, and GCN), we show that our method not only significantly outperforms recent rationalization methods, but also achieves comparable or even better results than a representative LLM (llama3.1-8b-instruct).
With the advance of large language models (LLMs), LLMs have been utilized for the various tasks. However, the issues of variability and reproducibility of results from each trial of LLMs have been largely overlooked in existing literature while actual human annotation uses majority voting to resolve disagreements among annotators. Therefore, this study introduces the straightforward ensemble strategy to a sentiment analysis using LLMs. As the results, we demonstrate that the ensemble of multiple inference using medium-sized LLMs produces more robust and accurate results than using a large model with a single attempt with reducing RMSE by 18.6%.
Large language models (LLMs) have revolutionized natural language processing by solving a wide range of tasks simply guided by a prompt. Yet their performance is highly sensitive to prompt formulation. While automated prompt optimization addresses this challenge by finding optimal prompts, current methods require a substantial number of LLM calls and input tokens, making prompt optimization expensive. We introduce CAPO (Cost-Aware Prompt Optimization), an algorithm that enhances prompt optimization efficiency by integrating AutoML techniques. CAPO is an evolutionary approach with LLMs as operators, incorporating racing to save evaluations and multi-objective optimization to balance performance with prompt length. It jointly optimizes instructions and few-shot examples while leveraging task descriptions for improved robustness. Our extensive experiments across diverse datasets and LLMs demonstrate that CAPO outperforms state-of-the-art discrete prompt optimization methods in 11/15 cases with improvements up to 21%p. Our algorithm achieves better performances already with smaller budgets, saves evaluations through racing, and decreases average prompt length via a length penalty, making it both cost-efficient and cost-aware. Even without few-shot examples, CAPO outperforms its competitors and generally remains robust to initial prompts. CAPO represents an important step toward making prompt optimization more powerful and accessible by improving cost-efficiency.
Language models based on the Transformer architecture achieve excellent results in many language-related tasks, such as text classification or sentiment analysis. However, despite the architecture of these models being well-defined, little is known about how their internal computations help them achieve their results. This renders these models, as of today, a type of 'black box' systems. There is, however, a line of research -- 'interpretability' -- aiming to learn how information is encoded inside these models. More specifically, there is work dedicated to studying whether Transformer-based models possess knowledge of linguistic phenomena similar to human speakers -- an area we call 'linguistic interpretability' of these models. In this survey we present a comprehensive analysis of 160 research works, spread across multiple languages and models -- including multilingual ones -- that attempt to discover linguistic information from the perspective of several traditional Linguistics disciplines: Syntax, Morphology, Lexico-Semantics and Discourse. Our survey fills a gap in the existing interpretability literature, which either not focus on linguistic knowledge in these models or present some limitations -- e.g. only studying English-based models. Our survey also focuses on Pre-trained Language Models not further specialized for a downstream task, with an emphasis on works that use interpretability techniques that explore models' internal representations.
This study presents a multi-stage approach to mental health classification by leveraging traditional machine learning algorithms, deep learning architectures, and transformer-based models. A novel data set was curated and utilized to evaluate the performance of various methods, starting with conventional classifiers and advancing through neural networks. To broaden the architectural scope, recurrent neural networks (RNNs) such as LSTM and GRU were also evaluated to explore their effectiveness in modeling sequential patterns in the data. Subsequently, transformer models such as BERT were fine-tuned to assess the impact of contextual embeddings in this domain. Beyond these baseline evaluations, the core contribution of this study lies in a novel training strategy involving a dual-model architecture composed of a teacher and a student network. Unlike standard distillation techniques, this method does not rely on soft label transfer; instead, it facilitates information flow through both the teacher model's output and its latent representations by modifying the loss function. The experimental results highlight the effectiveness of each modeling stage and demonstrate that the proposed loss function and teacher-student interaction significantly enhance the model's learning capacity in mental health prediction tasks.
Zero-shot text classification typically relies on prompt engineering, but the inherent prompt brittleness of large language models undermines its reliability. Minor changes in prompt can cause significant discrepancies in model performance. We attribute this prompt brittleness largely to the narrow focus on nexttoken probabilities in existing methods. To address this, we propose Placeholding Parallel Prediction (P3), a novel approach that predicts token probabilities across multiple positions and simulates comprehensive sampling of generation paths in a single run of a language model. Experiments show improved accuracy and up to 98% reduction in the standard deviation across prompts, boosting robustness. Even without a prompt, P3 maintains comparable performance, reducing the need for prompt engineering.
Automatic text classification (ATC) has experienced remarkable advancements in the past decade, best exemplified by recent small and large language models (SLMs and LLMs), leveraged by Transformer architectures. Despite recent effectiveness improvements, a comprehensive cost-benefit analysis investigating whether the effectiveness gains of these recent approaches compensate their much higher costs when compared to more traditional text classification approaches such as SVMs and Logistic Regression is still missing in the literature. In this context, this work's main contributions are twofold: (i) we provide a scientifically sound comparative analysis of the cost-benefit of twelve traditional and recent ATC solutions including five open LLMs, and (ii) a large benchmark comprising {22 datasets}, including sentiment analysis and topic classification, with their (train-validation-test) partitions based on folded cross-validation procedures, along with documentation, and code. The release of code, data, and documentation enables the community to replicate experiments and advance the field in a more scientifically sound manner. Our comparative experimental results indicate that LLMs outperform traditional approaches (up to 26%-7.1% on average) and SLMs (up to 4.9%-1.9% on average) in terms of effectiveness. However, LLMs incur significantly higher computational costs due to fine-tuning, being, on average 590x and 8.5x slower than traditional methods and SLMs, respectively. Results suggests the following recommendations: (1) LLMs for applications that require the best possible effectiveness and can afford the costs; (2) traditional methods such as Logistic Regression and SVM for resource-limited applications or those that cannot afford the cost of tuning large LLMs; and (3) SLMs like Roberta for near-optimal effectiveness-efficiency trade-off.
Tibetan, a minority language in China, features a highly intricate grammatical structure, characterized by four verb tenses and a tense system with frequent irregularities, contributing to its extensive inflectional diversity. Recently, advances in Large Language Models (LLMs) have transformed the paradigm in many domains. Despite the success in other fields, current LLMs often fall short in catering to the needs of domain experts like Tibetans, and the potential of LLMs for Tibetan culture is under-explored. The intrinsic reasons are the immense and intricate nature of Tibetan culture as well as the necessity for higher granularity and richness in knowledge. Simultaneously, the complexity and uniqueness of its grammatical structure, coupled with its status as a minority ethnic language, contribute to data scarcity, which remains a fundamental challenge. To alleviate these issues, we introduce Llama-Sunshine (Sun-Shine), the first large language model for Tibetan culture, which is expert in various Tibetan language processing tasks. Sun-Shine incorporates state-of-the-art model architectures optimized for Tibetan's linguistic features. We also propose TIB-STC, a comprehensive dataset comprising diverse Tibetan texts such as literature, religious scripts, news, and conversational data, which is also the first large-scale dataset for Tibetan culture. Though comprehensive experiments, Sun-Shine not only demonstrates a higher level of knowledge expertise for Tibetan culture but also gains preliminary embodied intelligence capabilities in Tibetan language processing tasks, like language modeling, text classification, machine translation, and syntactic analysis. Moreover, it excels in low-resource scenarios, showcasing strong generalization capabilities.