Large language models (LLMs) have shown remarkable performances across a wide range of tasks. However, the mechanisms by which these models encode tasks of varying complexities remain poorly understood. In this paper, we explore the hypothesis that LLMs process concepts of varying complexities in different layers, introducing the idea of “Concept Depth” to suggest that more complex concepts are typically acquired in deeper layers. Specifically, we categorize concepts based on their level of abstraction, defining them in the order of increasing complexity within factual, emotional, and inferential tasks. We conduct extensive probing experiments using layer-wise representations across various LLM families (Gemma, LLaMA, Qwen) on various datasets spanning the three domains of tasks. Our findings reveal that models could efficiently conduct probing for simpler tasks in shallow layers, and more complex tasks typically necessitate deeper layers for accurate understanding. Additionally, we examine how external factors, such as adding noise to the input and quantizing the model weights, might affect layer-wise representations. Our findings suggest that these factors can impede the development of a conceptual understanding of LLMs until deeper layers are explored. We hope that our proposed concept and experimental insights will enhance the understanding of the mechanisms underlying LLMs. Our codes are available at https://github.com/Luckfort/CD.

Sparse Autoencoders (SAEs) have recently emerged as powerful tools for interpreting and steering the internal representations of large language models (LLMs). However, conventional approaches to analyzing SAEs typically rely solely on input-side activations, without considering the influence between each latent feature and the model’s output. This work is built on two key hypotheses: (1) activated latents do not contribute equally to the construction of the model’s output, and (2) only latents with high influence are effective for model steering. To validate these hypotheses, we propose Gradient Sparse Autoencoder (GradSAE), a simple yet effective method that identifies the most influential latents by incorporating output-side gradient information.

Large language models (LLMs) have demonstrated impressive capabilities in natural language understanding and generation, but controlling their behavior reliably remains challenging, especially in open-ended generation settings. This paper introduces a novel supervised steering approach that operates in sparse, interpretable representation spaces. We employ sparse autoencoders (SAEs) to obtain sparse latent representations that aim to disentangle semantic attributes from model activations. Then we train linear classifiers to identify a small subspace of task-relevant dimensions in latent representations. Finally, we learn supervised steering vectors constrained to this subspace, optimized to align with target behaviors. Experiments across sentiment, truthfulness, and politics polarity steering tasks with multiple LLMs demonstrate that our supervised steering vectors achieve higher success rates with minimal degradation in generation quality compared to existing methods. Further analysis reveals that a notably small subspace is sufficient for effective steering, enabling more targeted and interpretable interventions.

Large Language Models (LLMs) demonstrate the ability to solve reasoning and mathematical problems using the Chain-of-Thought (CoT) technique. Expanding CoT length, as seen in models such as DeepSeek-R1, significantly enhances this reasoning for complex problems, but requires costly and high-quality long CoT data and fine-tuning. This work, inspired by the deep thinking paradigm of DeepSeek-R1, utilizes a steering technique to enhance the reasoning ability of an LLM without external datasets. Our method first employs Sparse Autoencoders (SAEs) to extract interpretable features from vanilla CoT. These features are then used to steer the LLM’s internal states during generation. Recognizing that many LLMs do not have corresponding pre-trained SAEs, we further introduce a novel SAE-free steering algorithm, which directly computes steering directions from the residual activations of an LLM, obviating the need for an explicit SAE. Experimental results demonstrate that both our SAE-based and subsequent SAE-free steering algorithms significantly enhance the reasoning capabilities of LLMs.

Backdoor attack is a severe threat to the trustworthiness of DNN-based language models. In this paper, we first extend the definition of memorization of language models from sample-wise to more fine-grained sentence element-wise (e.g., word, phrase, structure, and style), and then point out that language model backdoors are a type of element-wise memorization. Through further analysis, we find that the strength of such memorization is positively correlated to the frequency of duplicated elements in the training dataset. In conclusion, duplicated sentence elements are necessary for successful backdoor attacks. Based on this, we propose a data-centric defense. We first detect trigger candidates in training data by finding memorizable elements, i.e., duplicated elements, and then confirm real triggers by testing if the candidates can activate backdoor behaviors (i.e., malicious elements). Results show that our method outperforms state-of-the-art defenses in defending against different types of NLP backdoors.

Role-playing has emerged as an effective technique for enhancing the reasoning capabilities of large language models (LLMs). However, existing methods primarily rely on prompt engineering, which often lacks stability and interpretability. In this paper, we introduce Sparse Autoencoder Role-Playing Steering (SRPS), a novel framework that identifies and manipulates internal model features associated with role-playing behavior. Our approach extracts latent representations from role-play prompts, selects the most relevant features based on activation patterns, and constructs a steering vector that can be injected into the model’s residual stream with controllable intensity. Our method enables fine-grained control over role-specific behavior and offers insights into how role information influences internal model activations. Extensive experiments across various reasoning benchmarks and model sizes demonstrate consistent performance gains. Notably, in the zero-shot chain-of-thought (CoT) setting, the accuracy of Llama3.1-8B on CSQA improves from 31.86% to 39.80%, while Gemma2-9B on SVAMP increases from 37.50% to 45.10%. These results highlight the potential of SRPS to enhance reasoning ability in LLMs, providing better interpretability and stability compared to traditional prompt-based role-playing.

Large Language Models (LLMs) have transformed natural language processing, yet their internal mechanisms remain largely opaque. Recently, mechanistic interpretability has attracted significant attention from the research community as a means to understand the inner workings of LLMs. Among various mechanistic interpretability approaches, Sparse Autoencoders (SAEs) have emerged as a promising method due to their ability to disentangle the complex, superimposed features within LLMs into more interpretable components. This paper presents a comprehensive survey of SAEs for interpreting and understanding the internal workings of LLMs. Our major contributions include: (1) exploring the technical framework of SAEs, covering basic architecture, design improvements, and effective training strategies; (2) examining different approaches to explaining SAE features, categorized into input-based and output-based explanation methods; (3) discussing evaluation methods for assessing SAE performance, covering both structural and functional metrics; and (4) investigating real-world applications of SAEs in understanding and manipulating LLM behaviors.

Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities in processing both visual and textual information. However, the critical challenge of alignment between visual and textual representations is not fully understood. This survey presents a comprehensive examination of alignment and misalignment in LVLMs through an explainability lens. We first examine the fundamentals of alignment, exploring its representational and behavioral aspects, training methodologies, and theoretical foundations. We then analyze misalignment phenomena across three semantic levels: object, attribute, and relational misalignment. Our investigation reveals that misalignment emerges from challenges at multiple levels: the data level, the model level, and the inference level. We provide a comprehensive review of existing mitigation strategies, categorizing them into parameter-frozen and parameter-tuning approaches. Finally, we outline promising future research directions, emphasizing the need for standardized evaluation protocols and in-depth explainability studies.

The potential of large language models (LLMs) in contract legal risk analysis remains underexplored. In response, this paper introduces ContractEval, the first benchmark to thoroughly evaluate whether open-source LLMs could match proprietary LLMs in identifying clause-level legal risks in commercial contracts. Using the Contract Understanding Atticus Dataset (CUAD), we assess 4 proprietary and 15 open-source LLMs. Our results highlight five key findings: (1) Proprietary models outperform open-source models in both correctness and output effectiveness. (2) Larger open-source models generally perform better, though the improvement slows down as models get bigger. (3) Reasoning (“thinking”) mode improves output effectiveness but reduces correctness, likely due to over-complicating simpler tasks. (4) Open-source models generate “no related clause” responses more frequently even when relevant clauses are present. (5) Model quantization speed up inference but at the cost of performance drop, showing the tradeoff between efficiency and accuracy. These findings suggest that while most LLMs perform at a level comparable to junior legal assistants, open-source models require targeted fine-tuning to ensure correctness and effectiveness in high-stakes legal settings. ContractEval offers a solid benchmark to guide future development of legal-domain LLMs.

Probing techniques for large language models (LLMs) have primarily focused on English, overlooking the vast majority of other world’s languages. In this paper, we extend these probing methods to a multilingual context, investigating how LLMs encode linguistic structures across diverse languages. We conduct experiments on several open-source LLM models, analyzing probing accuracy, trends across layers, and similarities between probing vectors for multiple languages. Our key findings reveal: (1) a consistent performance gap between high-resource and low-resource languages, with high-resource languages achieving significantly higher probing accuracy; (2) divergent layer-wise accuracy trends, where high-resource languages show substantial improvement in deeper layers similar to English; and (3) higher representational similarities among high-resource languages, with low-resource languages demonstrating lower similarities both among themselves and with high-resource languages. These results provide insights into how linguistic structures are represented differently across languages in LLMs and emphasize the need for improved structure modeling for low-resource languages.

Recent studies highlight the effectiveness of using in-context learning (ICL) to steer large language models (LLMs) in processing tabular data, a challenging task given the structured nature of such data. Despite advancements in performance, the fairness implications of these methods are less understood. This study investigates how varying demonstrations within ICL prompts influence the fairness outcomes of LLMs. Our findings reveal that deliberately including minority group samples in prompts significantly boosts fairness without sacrificing predictive accuracy. Further experiments demonstrate that the proportion of minority to majority samples in demonstrations affects the trade-off between fairness and prediction accuracy. Based on these insights, we introduce a mitigation technique that employs clustering and evolutionary strategies to curate a diverse and representative sample set from the training data. This approach aims to enhance both predictive performance and fairness in ICL applications. Experimental results validate that our proposed method dramatically improves fairness across various metrics, showing its efficacy in real-world scenarios.

Large language models (LLMs) have notably revolutionized many domains within natural language processing due to their exceptional performance. Their security has become increasingly vital. This study is centered on protecting LLMs against unauthorized access and potential theft. We propose a simple yet effective protective measure wherein a unique key prompt is embedded within the LLM. This mechanism enables the model to respond only when presented with the correct key prompt; otherwise, LLMs will refuse to react to any input instructions. This key prompt protection offers a robust solution to prevent the unauthorized use of LLMs, as the model becomes unusable without the correct key. We evaluated the proposed protection on multiple LLMs and NLP tasks. Results demonstrate that our method can successfully protect the LLM without significantly impacting the model’s original function. Moreover, we demonstrate potential attacks that attempt to bypass the protection mechanism will adversely affect the model’s performance, further emphasizing the effectiveness of the proposed protection method.

Chain of Thought (CoT) is significant in improving the reasoning abilities of large language models (LLMs). However, the correlation between the effectiveness of CoT and the length of reasoning steps in prompts remains largely unknown. To shed light on this, we have conducted several empirical experiments to explore the relations. Specifically, we design experiments that expand and compress the rationale reasoning steps within CoT demonstrations, while keeping all other factors constant. We have the following key findings. First, the results indicate that lengthening the reasoning steps in prompts, even without adding new information into the prompt, considerably enhances LLMs’ reasoning abilities across multiple datasets. Alternatively, shortening the reasoning steps, even while preserving the key information, significantly diminishes the reasoning abilities of models. This finding highlights the importance of the number of steps in CoT prompts and provides practical guidance to make better use of LLMs’ potential in complex problem-solving scenarios. Second, we also investigated the relationship between the performance of CoT and the rationales used in demonstrations. Surprisingly, the result shows that even incorrect rationales can yield favorable outcomes if they maintain the requisite length of inference. Third, we observed that the advantages of increasing reasoning steps are task-dependent: simpler tasks require fewer steps, whereas complex tasks gain significantly from longer inference sequences.

Human-like social bias of pre-trained language models (PLMs) on downstream tasks have attracted increasing attention. The potential flaws in the training data are the main factor that causes unfairness in PLMs. Existing data-centric debiasing strategies mainly leverage explicit bias words (defined as sensitive attribute words specific to demographic groups) for counterfactual data augmentation to balance the training data. However, they lack consideration of implicit bias words potentially associated with explicit bias words in complex distribution data, which indirectly harms the fairness of PLMs. To this end, we propose a **Data**-Centric **Debias**ing method (named Data-Debias), which uses an explainability method to search for implicit bias words to assist in debiasing PLMs. Specifically, we compute the feature attributions of all tokens using the Integrated Gradients method, and then treat the tokens that have a large impact on the model’s decision as implicit bias words. To make the search results more precise, we iteratively train a biased model to amplify the bias with each iteration. Finally, we use the implicit bias words searched in the last iteration to assist in debiasing PLMs. Extensive experimental results on multiple PLMs debiasing on three different classification tasks demonstrate that Data-Debias achieves state-of-the-art debiasing performance and strong generalization while maintaining predictive abilities.

Recent research has shown that large language models rely on spurious correlations in the data for natural language understanding (NLU) tasks. In this work, we aim to answer the following research question: Can we reduce spurious correlations by modifying the ground truth labels of the training data? Specifically, we propose a simple yet effective debiasing framework, named Soft Label Encoding (SoftLE). First, we train a teacher model to quantify each sample’s degree of relying on shortcuts. Then, we encode this shortcut degree into a dummy class and use it to smooth the original ground truth labels, generating soft labels. These soft labels are used to train a more robust student model that reduces spurious correlations between shortcut features and certain classes. Extensive experiments on two NLU benchmark tasks via two language models demonstrate that SoftLE significantly improves out-of-distribution generalization while maintaining satisfactory in-distribution accuracy. Our code is available at https://github.com/ZiruiHE99/sle

Deep learning has introduced significant improvements in many software analysis tasks. Although the Large Language Models (LLMs) based neural code models demonstrate commendable performance when trained and tested within the intra-project independent and identically distributed (IID) setting, they often struggle to generalize effectively to real-world inter-project out-of-distribution (OOD) data. In this work, we show that this phenomenon is caused by the heavy reliance on project-specific shortcuts for prediction instead of ground-truth evidence. We propose a Cond-Idf measurement to interpret this behavior, which quantifies the relatedness of a token with a label and its project-specificness. The strong correlation between model behavior and the proposed measurement indicates that without proper regularization, models tend to leverage spurious statistical cues for prediction. Equipped with these observations, we propose a novel bias mitigation mechanism that regularizes the model’s learning behavior by leveraging latent logic relations among samples. Experimental results on two representative program analysis tasks indicate that our mitigation framework can improve both inter-project OOD generalization and adversarial robustness, while not sacrificing accuracy on intra-project IID data.

As the representation capability of Pre-trained Language Models (PLMs) improve, there is growing concern that they will inherit social biases from unprocessed corpora. Most previous debiasing techniques used Counterfactual Data Augmentation (CDA) to balance the training corpus. However, CDA slightly modifies the original corpus, limiting the representation distance between different demographic groups to a narrow range. As a result, the debiasing model easily fits the differences between counterfactual pairs, which affects its debiasing performance with limited text resources. In this paper, we propose an adversarial training-inspired two-stage debiasing model using Contrastive learning with Continuous Prompt Augmentation (named CCPA) to mitigate social biases in PLMs’ encoding. In the first stage, we propose a data augmentation method based on continuous prompt tuning to push farther the representation distance between sample pairs along different demographic groups. In the second stage, we utilize contrastive learning to pull closer the representation distance between the augmented sample pairs and then fine-tune PLMs’ parameters to get debiased encoding. Our approach guides the model to achieve stronger debiasing performance by adding difficulty to the training process. Extensive experiments show that CCPA outperforms baselines in terms of debiasing performance. Meanwhile, experimental results on the GLUE benchmark show that CCPA retains the language modeling capability of PLMs.

Recent work has focused on compressing pre-trained language models (PLMs) like BERT where the major focus has been to improve the in-distribution performance for downstream tasks. However, very few of these studies have analyzed the impact of compression on the generalizability and robustness of compressed models for out-of-distribution (OOD) data. Towards this end, we study two popular model compression techniques including knowledge distillation and pruning and show that the compressed models are significantly less robust than their PLM counterparts on OOD test sets although they obtain similar performance on in-distribution development sets for a task. Further analysis indicates that the compressed models overfit on the shortcut samples and generalize poorly on the hard ones. We further leverage this observation to develop a regularization strategy for robust model compression based on sample uncertainty.

Recent studies indicate that NLU models are prone to rely on shortcut features for prediction, without achieving true language understanding. As a result, these models fail to generalize to real-world out-of-distribution data. In this work, we show that the words in the NLU training set can be modeled as a long-tailed distribution. There are two findings: 1) NLU models have strong preference for features located at the head of the long-tailed distribution, and 2) Shortcut features are picked up during very early few iterations of the model training. These two observations are further employed to formulate a measurement which can quantify the shortcut degree of each training sample. Based on this shortcut measurement, we propose a shortcut mitigation framework LGTR, to suppress the model from making overconfident predictions for samples with large shortcut degree. Experimental results on three NLU benchmarks demonstrate that our long-tailed distribution explanation accurately reflects the shortcut learning behavior of NLU models. Experimental analysis further indicates that LGTR can improve the generalization accuracy on OOD data, while preserving the accuracy on in-distribution data.