Xuansheng Wu

Yun Wang, Xin Xia, Xuansheng Wu et al.

LLM-based automated scoring approaches near-human performance, but scaling to new tasks remains bottlenecked by the per-item human configuration of upstream stages such as rubric construction. Human experts bypass this bottleneck through evaluation heuristics developed over extensive practice. We ask whether LLMs can learn similar heuristics directly from scoring experience, and formalize this as the concept of assessment skills: item-independent natural-language procedural knowledge that guides LLMs through specific stages of the scoring workflow. Focusing on rubric construction as a first instantiation, we propose an iterative framework that decomposes a skill into a fixed scaffold and learnable item-agnostic rules, refining the rules through LLM-driven diagnosis of scoring errors and validation-gated selection. The framework requires no expert-written rubric. On all ten ASAP-SAS items, optimized skills substantially improve LLM-based scoring and frequently surpass the dataset-provided expert rubric. Cross-item transfer experiments further reveal that learned skills capture both generalizable and item-specific patterns.

Yucheng Shi, Mengnan Du, Xuansheng Wu et al.

Backdoor attacks inject poisoned samples into the training data, resulting in the misclassification of the poisoned input during a model's deployment. Defending against such attacks is challenging, especially for real-world black-box models where only query access is permitted. In this paper, we propose a novel defense framework against backdoor attacks through Zero-shot Image Purification (ZIP). Our framework can be applied to poisoned models without requiring internal information about the model or any prior knowledge of the clean/poisoned samples. Our defense framework involves two steps. First, we apply a linear transformation (e.g., blurring) on the poisoned image to destroy the backdoor pattern. Then, we use a pre-trained diffusion model to recover the missing semantic information removed by the transformation. In particular, we design a new reverse process by using the transformed image to guide the generation of high-fidelity purified images, which works in zero-shot settings. We evaluate our ZIP framework on multiple datasets with different types of attacks. Experimental results demonstrate the superiority of our ZIP framework compared to state-of-the-art backdoor defense baselines. We believe that our results will provide valuable insights for future defense methods for black-box models. Our code is available at https://github.com/sycny/ZIP.

Xuansheng Wu, Huachi Zhou, Yucheng Shi et al.

Recommendation systems help users find matched items based on their previous behaviors. Personalized recommendation becomes challenging in the absence of historical user-item interactions, a practical problem for startups known as the system cold-start recommendation. While existing research addresses cold-start issues for either users or items, we still lack solutions for system cold-start scenarios. To tackle the problem, we propose PromptRec, a simple but effective approach based on in-context learning of language models, where we transform the recommendation task into the sentiment analysis task on natural language containing user and item profiles. However, this naive approach heavily relies on the strong in-context learning ability emerged from large language models, which could suffer from significant latency for online recommendations. To solve the challenge, we propose to enhance small language models for recommender systems with a data-centric pipeline, which consists of: (1) constructing a refined corpus for model pre-training; (2) constructing a decomposed prompt template via prompt pre-training. They correspond to the development of training data and inference data, respectively. The pipeline is supported by a theoretical framework that formalizes the connection between in-context recommendation and language modeling. To evaluate our approach, we introduce a cold-start recommendation benchmark, and the results demonstrate that the enhanced small language models can achieve comparable cold-start recommendation performance to that of large models with only $17\%$ of the inference time. To the best of our knowledge, this is the first study to tackle the system cold-start recommendation problem. We believe our findings will provide valuable insights for future works. The benchmark and implementations are available at https://github.com/JacksonWuxs/PromptRec.

Xuansheng Wu, Wenlin Yao, Jianshu Chen et al.

Large Language Models (LLMs) have achieved remarkable success, where instruction tuning is the critical step in aligning LLMs with user intentions. In this work, we investigate how the instruction tuning adjusts pre-trained models with a focus on intrinsic changes. Specifically, we first develop several local and global explanation methods, including a gradient-based method for input-output attribution, and techniques for interpreting patterns and concepts in self-attention and feed-forward layers. The impact of instruction tuning is then studied by comparing the explanations derived from the pre-trained and instruction-tuned models. This approach provides an internal perspective of the model shifts on a human-comprehensible level. Our findings reveal three significant impacts of instruction tuning: 1) It empowers LLMs to recognize the instruction parts of user prompts, and promotes the response generation constantly conditioned on the instructions. 2) It encourages the self-attention heads to capture more word-word relationships about instruction verbs. 3) It encourages the feed-forward networks to rotate their pre-trained knowledge toward user-oriented tasks. These insights contribute to a more comprehensive understanding of instruction tuning and lay the groundwork for future work that aims at explaining and optimizing LLMs for various applications. Our code and data are publicly available at https://github.com/JacksonWuxs/Interpret_Instruction_Tuning_LLMs.

Ehsan Latif, Gengchen Mai, Matthew Nyaaba et al.

Artificial general intelligence (AGI) has gained global recognition as a future technology due to the emergence of breakthrough large language models and chatbots such as GPT-4 and ChatGPT, respectively. Compared to conventional AI models, typically designed for a limited range of tasks, demand significant amounts of domain-specific data for training and may not always consider intricate interpersonal dynamics in education. AGI, driven by the recent large pre-trained models, represents a significant leap in the capability of machines to perform tasks that require human-level intelligence, such as reasoning, problem-solving, decision-making, and even understanding human emotions and social interactions. This position paper reviews AGI's key concepts, capabilities, scope, and potential within future education, including achieving future educational goals, designing pedagogy and curriculum, and performing assessments. It highlights that AGI can significantly improve intelligent tutoring systems, educational assessment, and evaluation procedures. AGI systems can adapt to individual student needs, offering tailored learning experiences. They can also provide comprehensive feedback on student performance and dynamically adjust teaching methods based on student progress. The paper emphasizes that AGI's capabilities extend to understanding human emotions and social interactions, which are critical in educational settings. The paper discusses that ethical issues in education with AGI include data bias, fairness, and privacy and emphasizes the need for codes of conduct to ensure responsible AGI use in academic settings like homework, teaching, and recruitment. We also conclude that the development of AGI necessitates interdisciplinary collaborations between educators and AI engineers to advance research and application efforts.

Xuansheng Wu, Zhiyi Zhao, Ninghao Liu

We propose a novel non-parametric/un-trainable language model, named Non-Parametric Pairwise Attention Random Walk Model (NoPPA), to generate sentence embedding only with pre-trained word embedding and pre-counted word frequency. To the best we know, this study is the first successful attempt to break the constraint on bag-of-words assumption with a non-parametric attention mechanism. We evaluate our method on eight different downstream classification tasks. The experiment results show that NoPPA outperforms all kinds of bag-of-words-based methods in each dataset and provides a comparable or better performance than the state-of-the-art non-parametric methods on average. Furthermore, visualization supports that NoPPA can understand contextual topics, common phrases, and word causalities. Our model is available at https://github.com/JacksonWuxs/NoPPA.

Xuansheng Wu, Xinyu He, Tianming Liu et al.

Developing models to automatically score students' written responses to science problems is critical for science education. However, collecting and labeling sufficient student responses for training models is time and cost-consuming. Recent studies suggest that pre-trained language models can be adapted to downstream tasks without fine-tuning with prompts. However, no research has employed such a prompt approach in science education. As student responses are presented with natural language, aligning the scoring procedure as the next sentence prediction task using prompts can skip the costly fine-tuning stage. In this study, we developed a zero-shot approach to automatically score student responses via Matching Exemplars as Next Sentence Prediction (MeNSP). This approach employs no training samples. We first apply MeNSP in scoring three assessment tasks of scientific argumentation and found machine-human scoring agreements, Cohen's Kappa ranges from 0.30 to 0.57, and F1 score ranges from 0.54 to 0.81. To improve the performance, we extend our research to the few-shots setting, either randomly selecting labeled student responses or manually constructing responses to fine-tune the models. We find that one task's performance is improved with more samples, Cohen's Kappa from 0.30 to 0.38, and F1 score from 0.54 to 0.59; for the two others, scoring performance is not improved. We also find that randomly selected few-shots perform better than the human expert-crafted approach. This study suggests that MeNSP can yield referable automatic scoring for student responses while significantly reducing the cost of model training. This method can benefit low-stakes classroom assessment practices in science education. Future research should further explore the applicability of the MeNSP in different types of assessment tasks in science education and improve the model performance.

Gyeong-Geon Lee, Ehsan Latif, Xuansheng Wu et al.

This study investigates the application of large language models (LLMs), specifically GPT-3.5 and GPT-4, with Chain-of-Though (CoT) in the automatic scoring of student-written responses to science assessments. We focused on overcoming the challenges of accessibility, technical complexity, and lack of explainability that have previously limited the use of artificial intelligence-based automatic scoring tools among researchers and educators. With a testing dataset comprising six assessment tasks (three binomial and three trinomial) with 1,650 student responses, we employed six prompt engineering strategies to automatically score student responses. The six strategies combined zero-shot or few-shot learning with CoT, either alone or alongside item stem and scoring rubrics. Results indicated that few-shot (acc = .67) outperformed zero-shot learning (acc = .60), with 12.6% increase. CoT, when used without item stem and scoring rubrics, did not significantly affect scoring accuracy (acc = .60). However, CoT prompting paired with contextual item stems and rubrics proved to be a significant contributor to scoring accuracy (13.44% increase for zero-shot; 3.7% increase for few-shot). We found a more balanced accuracy across different proficiency categories when CoT was used with a scoring rubric, highlighting the importance of domain-specific reasoning in enhancing the effectiveness of LLMs in scoring tasks. We also found that GPT-4 demonstrated superior performance over GPT -3.5 in various scoring tasks when combined with the single-call greedy sampling or ensemble voting nucleus sampling strategy, showing 8.64% difference. Particularly, the single-call greedy sampling strategy with GPT-4 outperformed other approaches.

Xuansheng Wu, Kaixiong Zhou, Mingchen Sun et al.

The recent "pre-train, prompt, predict training" paradigm has gained popularity as a way to learn generalizable models with limited labeled data. The approach involves using a pre-trained model and a prompting function that applies a template to input samples, adding indicative context and reformulating target tasks as the pre-training task. However, the design of prompts could be a challenging and time-consuming process in complex tasks. The limitation can be addressed by using graph data, as graphs serve as structured knowledge repositories by explicitly modeling the interaction between entities. In this survey, we review prompting methods from the graph perspective, where prompting functions are augmented with graph knowledge. In particular, we introduce the basic concepts of graph prompt learning, organize the existing work of designing graph prompting functions, and describe their applications and future challenges. This survey will bridge the gap between graphs and prompt design to facilitate future methodology development.

Xuansheng Wu, Padmaja Pravin Saraf, Gyeonggeon Lee et al.

Large language models (LLMs) have demonstrated strong potential in performing automatic scoring for constructed response assessments. While constructed responses graded by humans are usually based on given grading rubrics, the methods by which LLMs assign scores remain largely unclear. It is also uncertain how closely AI's scoring process mirrors that of humans or if it adheres to the same grading criteria. To address this gap, this paper uncovers the grading rubrics that LLMs used to score students' written responses to science tasks and their alignment with human scores. We also examine whether enhancing the alignments can improve scoring accuracy. Specifically, we prompt LLMs to generate analytic rubrics that they use to assign scores and study the alignment gap with human grading rubrics. Based on a series of experiments with various configurations of LLM settings, we reveal a notable alignment gap between human and LLM graders. While LLMs can adapt quickly to scoring tasks, they often resort to shortcuts, bypassing deeper logical reasoning expected in human grading. We found that incorporating high-quality analytical rubrics designed to reflect human grading logic can mitigate this gap and enhance LLMs' scoring accuracy. These results underscore the need for a nuanced approach when applying LLMs in science education and highlight the importance of aligning LLM outputs with human expectations to ensure efficient and accurate automatic scoring.

Xuansheng Wu, Haiyan Zhao, Yaochen Zhu et al.

Explainable AI (XAI) refers to techniques that provide human-understandable insights into the workings of AI models. Recently, the focus of XAI is being extended toward explaining Large Language Models (LLMs). This extension calls for a significant transformation in the XAI methodologies for two reasons. First, many existing XAI methods cannot be directly applied to LLMs due to their complexity and advanced capabilities. Second, as LLMs are increasingly deployed in diverse applications, the role of XAI shifts from merely opening the ``black box'' to actively enhancing the productivity and applicability of LLMs in real-world settings. Meanwhile, the conversation and generation abilities of LLMs can reciprocally enhance XAI. Therefore, in this paper, we introduce Usable XAI in the context of LLMs by analyzing (1) how XAI can explain and improve LLM-based AI systems and (2) how XAI techniques can be improved by using LLMs. We introduce 10 strategies, introducing the key techniques for each and discussing their associated challenges. We also provide case studies to demonstrate how to obtain and leverage explanations. The code used in this paper can be found at: https://github.com/JacksonWuxs/UsableXAI_LLM.

Yucheng Shi, Qiaoyu Tan, Xuansheng Wu et al.

Large Language Models (LLMs) have shown proficiency in question-answering tasks but often struggle to integrate real-time knowledge, leading to potentially outdated or inaccurate responses. This problem becomes even more challenging when dealing with multi-hop questions, since they require LLMs to update and integrate multiple knowledge pieces relevant to the questions. To tackle the problem, we propose the Retrieval-Augmented model Editing (RAE) framework for multi-hop question answering. RAE first retrieves edited facts and then refines the language model through in-context learning. Specifically, our retrieval approach, based on mutual information maximization, leverages the reasoning abilities of LLMs to identify chain facts that traditional similarity-based searches might miss. In addition, our framework includes a pruning strategy to eliminate redundant information from the retrieved facts, which enhances the editing accuracy and mitigates the hallucination problem. Our framework is supported by theoretical justification for its fact retrieval efficacy. Finally, comprehensive evaluation across various LLMs validates RAE's ability in providing accurate answers with updated knowledge. Our code is available at: https://github.com/sycny/RAE.

Xuansheng Wu, Jiayi Yuan, Wenlin Yao et al.

Large language models (LLMs) excel at handling human queries, but they can occasionally generate flawed or unexpected responses. Understanding their internal states is crucial for understanding their successes, diagnosing their failures, and refining their capabilities. Although sparse autoencoders (SAEs) have shown promise for interpreting LLM internal representations, limited research has explored how to better explain SAE features, i.e., understanding the semantic meaning of features learned by SAE. Our theoretical analysis reveals that existing explanation methods suffer from the frequency bias issue, where they emphasize linguistic patterns over semantic concepts, while the latter is more critical to steer LLM behaviors. To address this, we propose using a fixed vocabulary set for feature interpretations and designing a mutual information-based objective, aiming to better capture the semantic meaning behind these features. We further propose two runtime steering strategies that adjust the learned feature activations based on their corresponding explanations. Empirical results show that, compared to baselines, our method provides more discourse-level explanations and effectively steers LLM behaviors to defend against jailbreak attacks. These findings highlight the value of explanations for steering LLM behaviors in downstream applications. We will release our code and data once accepted.

Xuansheng Wu, Wenhao Yu, Xiaoming Zhai et al.

Modern text classification methods heavily rely on contextual embeddings from large language models (LLMs). Compared to human-engineered features, these embeddings provide automatic and effective representations for classification model training. However, they also introduce a challenge: we lose the ability to manually remove unintended features, such as sensitive or task-irrelevant features, to guarantee regulatory compliance or improve the generalizability of classification models. This limitation arises because LLM embeddings are opaque and difficult to interpret. In this paper, we propose a novel framework to identify and regularize unintended features in the LLM latent space. Specifically, we first pre-train a sparse autoencoder (SAE) to extract interpretable features from LLM latent spaces. To ensure the SAE can capture task-specific features, we further fine-tune it on task-specific datasets. In training the classification model, we propose a simple and effective regularizer, by minimizing the similarity between the classifier weights and the identified unintended feature, to remove the impact of these unintended features on classification. We evaluate the proposed framework on three real-world tasks, including toxic chat detection, reward modeling, and disease diagnosis. Results show that the proposed self-regularization framework can improve the classifier's generalizability by regularizing those features that are not semantically correlated to the task. This work pioneers controllable text classification on LLM latent spaces by leveraging interpreted features to address generalizability, fairness, and privacy challenges. The code and data are publicly available at https://github.com/JacksonWuxs/Controllable_LLM_Classifier.

Shu Yang, Junchao Wu, Xilin Gong et al.

Large Reasoning Models (LRMs) have demonstrated remarkable performance on complex tasks by engaging in extended reasoning before producing final answers. Beyond improving abilities, these detailed reasoning traces also create a new opportunity for AI safety, CoT Monitorability: monitoring potential model misbehavior, such as the use of shortcuts or sycophancy, through their chain-of-thought (CoT) during decision-making. However, two key fundamental challenges arise when attempting to build more effective monitors through CoT analysis. First, as prior research on CoT faithfulness has pointed out, models do not always truthfully represent their internal decision-making in the generated reasoning. Second, monitors themselves may be either overly sensitive or insufficiently sensitive, and can potentially be deceived by models' long, elaborate reasoning traces. In this paper, we present the first systematic investigation of the challenges and potential of CoT monitorability. Motivated by two fundamental challenges we mentioned before, we structure our study around two central perspectives: (i) verbalization: to what extent do LRMs faithfully verbalize the true factors guiding their decisions in the CoT, and (ii) monitor reliability: to what extent can misbehavior be reliably detected by a CoT-based monitor? Specifically, we provide empirical evidence and correlation analyses between verbalization quality, monitor reliability, and LLM performance across mathematical, scientific, and ethical domains. Then we further investigate how different CoT intervention methods, designed to improve reasoning efficiency or performance, will affect monitoring effectiveness. Finally, we propose MoME, a new paradigm in which LLMs monitor other models' misbehavior through their CoT and provide structured judgments along with supporting evidence.

Shu Yang, Junchao Wu, Xuansheng Wu et al.

Large Reasoning Models (LRMs) have achieved remarkable performance on complex tasks by engaging in extended reasoning before producing final answers, yet this strength introduces the risk of overthinking, where excessive token generation occurs even for simple tasks. While recent work in efficient reasoning seeks to reduce reasoning length while preserving accuracy, it remains unclear whether such optimization is truly a free lunch. Drawing on the intuition that compressing reasoning may reduce the robustness of model responses and lead models to omit key reasoning steps, we investigate whether efficient reasoning strategies introduce behavioral inconsistencies. To systematically assess this, we introduce $ICBENCH$, a benchmark designed to measure inconsistency in LRMs across three dimensions: inconsistency across task settings (ITS), inconsistency between training objectives and learned behavior (TR-LB), and inconsistency between internal reasoning and self-explanations (IR-SE). Applying $ICBENCH$ to a range of open-source LRMs, we find that while larger models generally exhibit greater consistency than smaller ones, they all display widespread "scheming" behaviors, including self-disagreement, post-hoc rationalization, and the withholding of reasoning cues. Crucially, our results demonstrate that efficient reasoning strategies such as No-Thinking and Simple Token-Budget consistently increase all three defined types of inconsistency. These findings suggest that although efficient reasoning enhances token-level efficiency, further investigation is imperative to ascertain whether it concurrently introduces the risk of models evading effective supervision.

Tianze Yang, Yucheng Shi, Mengnan Du et al.

Vector-Quantized Generative Models (VQGMs) have emerged as powerful tools for image generation. However, the key component of VQGMs -- the codebook of discrete tokens -- is still not well understood, e.g., which tokens are critical to generate an image of a certain concept? This paper introduces Concept-Oriented Token Explanation (CORTEX), a novel approach for interpreting VQGMs by identifying concept-specific token combinations. Our framework employs two methods: (1) a sample-level explanation method that analyzes token importance scores in individual images, and (2) a codebook-level explanation method that explores the entire codebook to find globally relevant tokens. Experimental results demonstrate CORTEX's efficacy in providing clear explanations of token usage in the generative process, outperforming baselines across multiple pretrained VQGMs. Besides enhancing VQGMs transparency, CORTEX is useful in applications such as targeted image editing and shortcut feature detection. Our code is available at https://github.com/YangTianze009/CORTEX.

Yun Wang, Zhaojun Ding, Xuansheng Wu et al.

Automated scoring plays a crucial role in education by reducing the reliance on human raters, offering scalable and immediate evaluation of student work. While large language models (LLMs) have shown strong potential in this task, their use as end-to-end raters faces challenges such as low accuracy, prompt sensitivity, limited interpretability, and rubric misalignment. These issues hinder the implementation of LLM-based automated scoring in assessment practice. To address the limitations, we propose AutoSCORE, a multi-agent LLM framework enhancing automated scoring via rubric-aligned Structured COmponent REcognition. With two agents, AutoSCORE first extracts rubric-relevant components from student responses and encodes them into a structured representation (i.e., Scoring Rubric Component Extraction Agent), which is then used to assign final scores (i.e., Scoring Agent). This design ensures that model reasoning follows a human-like grading process, enhancing interpretability and robustness. We evaluate AutoSCORE on four benchmark datasets from the ASAP benchmark, using both proprietary and open-source LLMs (GPT-4o, LLaMA-3.1-8B, and LLaMA-3.1-70B). Across diverse tasks and rubrics, AutoSCORE consistently improves scoring accuracy, human-machine agreement (QWK, correlations), and error metrics (MAE, RMSE) compared to single-agent baselines, with particularly strong benefits on complex, multi-dimensional rubrics, and especially large relative gains on smaller LLMs. These results demonstrate that structured component recognition combined with multi-agent design offers a scalable, reliable, and interpretable solution for automated scoring.

Yiwei Qin, Kaiqiang Song, Yebowen Hu et al. · tencent-ai

This paper introduces the Decomposed Requirements Following Ratio (DRFR), a new metric for evaluating Large Language Models' (LLMs) ability to follow instructions. Addressing a gap in current methodologies, DRFR breaks down complex instructions into simpler criteria, facilitating a detailed analysis of LLMs' compliance with various aspects of tasks. Alongside this metric, we present InFoBench, a benchmark comprising 500 diverse instructions and 2,250 decomposed questions across multiple constraint categories. Our experiments compare DRFR with traditional scoring methods and explore annotation sources, including human experts, crowd-sourced workers, and GPT-4. The findings demonstrate DRFR's higher reliability and the effectiveness of using GPT-4 as a cost-efficient annotator. The evaluation of several advanced LLMs using this framework reveals their strengths and areas needing improvement, particularly in complex instruction-following. This study contributes a novel metric and benchmark, offering insights for future LLM development and evaluation.

Xinquan Chen, Zhenyun Yin, Shan He et al.

As large models evolve from conversational assistants into autonomous agents, challenges increasingly arise from long-horizon decision making, tool use, and real environment interaction. Existing agenticinfrastructure remain fragmented across evaluation, data management, and agent evolution, making it difficult to discover risks systematically and improve models in a continuous closed loop. In this report, we present \textbf{Safactory}, a scalable agent factory for trustworthy autonomous intelligence. Safactory integrates three tightly coupled platforms: a \textbf{Parallel Simulation Platform} for trajectory generation, a \textbf{Trustworthy Data Platform} for trajectory storage and experience extraction, and an \textbf{Autonomous Evolution Platform} for asynchronous reinforcement learning and on-policy distillation. As far as we know, Safactory is the first framework to propose a unified evolutionary pipeline for next-generation trustworthy autonomous intelligence.

Dong Shu, Xuansheng Wu, Haiyan Zhao et al.

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.

Haiyan Zhao, Xuansheng Wu, Fan Yang et al.

Linear concept vectors effectively steer LLMs, but existing methods suffer from noisy features in diverse datasets that undermine steering robustness. We propose Sparse Autoencoder-Denoised Concept Vectors (SDCV), which selectively keep the most discriminative SAE latents while reconstructing hidden representations. Our key insight is that concept-relevant signals can be explicitly separated from dataset noise by scaling up activations of top-k latents that best differentiate positive and negative samples. Applied to linear probing and difference-in-mean, SDCV consistently improves steering success rates by 4-16\% across six challenging concepts, while maintaining topic relevance.

Shuchen Guo, Yun Wang, Jichao Yu et al.

This study investigated potential scoring biases and disparities toward English Language Learners (ELLs) when using automatic scoring systems for middle school students' written responses to science assessments. We specifically focus on examining how unbalanced training data with ELLs contributes to scoring bias and disparities. We fine-tuned BERT with four datasets: responses from (1) ELLs, (2) non-ELLs, (3) a mixed dataset reflecting the real-world proportion of ELLs and non-ELLs (unbalanced), and (4) a balanced mixed dataset with equal representation of both groups. The study analyzed 21 assessment items: 10 items with about 30,000 ELL responses, five items with about 1,000 ELL responses, and six items with about 200 ELL responses. Scoring accuracy (Acc) was calculated and compared to identify bias using Friedman tests. We measured the Mean Score Gaps (MSGs) between ELLs and non-ELLs and then calculated the differences in MSGs generated through both the human and AI models to identify the scoring disparities. We found that no AI bias and distorted disparities between ELLs and non-ELLs were found when the training dataset was large enough (ELL = 30,000 and ELL = 1,000), but concerns could exist if the sample size is limited (ELL = 200).

Dong Shu, Xuansheng Wu, Haiyan Zhao et al.

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 causal 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 causal 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.

Zhongzhi Li, Xuansheng Wu, Yijiang Li et al.

The diversity of post-training data is critical for effective downstream performance in large language models (LLMs). Many existing approaches to constructing post-training data quantify diversity using text-based metrics that capture linguistic variation, but such metrics provide only weak signals for the task-relevant features that determine downstream performance. In this work, we introduce Feature Activation Coverage (FAC) which measures data diversity in an interpretable feature space. Building upon this metric, we further propose a diversity-driven data synthesis framework, named FAC Synthesis, that first uses a sparse autoencoder to identify missing features from a seed dataset, and then generates synthetic samples that explicitly reflect these features. Experiments show that our approach consistently improves both data diversity and downstream performance on various tasks, including instruction following, toxicity detection, reward modeling, and behavior steering. Interestingly, we identify a shared, interpretable feature space across model families (i.e., LLaMA, Mistral, and Qwen), enabling cross-model knowledge transfer. Our work provides a solid and practical methodology for exploring data-centric optimization of LLMs.

Xuansheng Wu, Xiaoman Pan, Wenlin Yao et al.

Reinforcement learning with verifiable rewards (RLVR) can elicit strong reasoning in large language models (LLMs), while their performance after RLVR varies dramatically across different base models. This raises a fundamental question: what microscopic property of pre-trained models leads to this variation? To investigate, we formalize reasoning as chains of Horn clauses ("if-then" rules) built from features extracted from the LLM's latent space via cross-layer sparse autoencoders (SAEs). We estimate the transition probabilities between its features, and further categorize each rule by its semantic soundness level (e.g., strict, plausible, noisy) with an LLM. Our key discovery is that high-potential models are inherently soundness-aware: their internal probability distributions systematically shift across rules' soundness levels, becoming highly distinct for "strict" versus "noisy" rules. In contrast, weaker models are soundness-agnostic, collapsing to one distribution regardless of soundness levels. To quantify this, we introduce the Soundness-Aware Level (SAL), a microscopic metric using the Jensen-Shannon Divergence to measure the separation between these distributions. We show that SAL's predictions of post-RLVR reasoning performance follow a precise empirical law (R^2=0.87) across diverse model families (Qwen, Mistral, Llama, DeepSeek) and scales (0.5B-14B). This reveals that a model's reasoning potential is tied to its intrinsic, pre-trained ability to distinguish sound knowledge from unsound ones. These findings underscore the critical role of model pre-training in shaping reasoning and offer a practical metric grounded in the model's internal mechanisms for selecting/designing stronger base models.

Zhenyue Qin, Yang Liu, Yu Yin et al.

Vision-threatening eye diseases pose a major global health burden, with timely diagnosis limited by workforce shortages and restricted access to specialized care. While multimodal large language models (MLLMs) show promise for medical image interpretation, advancing MLLMs for ophthalmology is hindered by the lack of comprehensive benchmark datasets suitable for evaluating generative models. We present a large-scale multimodal ophthalmology benchmark comprising 32,633 instances with multi-granular annotations across 12 common ophthalmic conditions and 5 imaging modalities. The dataset integrates imaging, anatomical structures, demographics, and free-text annotations, supporting anatomical structure recognition, disease screening, disease staging, and demographic prediction for bias evaluation. This work extends our preliminary LMOD benchmark with three major enhancements: (1) nearly 50% dataset expansion with substantial enlargement of color fundus photography; (2) broadened task coverage including binary disease diagnosis, multi-class diagnosis, severity classification with international grading standards, and demographic prediction; and (3) systematic evaluation of 24 state-of-the-art MLLMs. Our evaluations reveal both promise and limitations. Top-performing models achieved ~58% accuracy in disease screening under zero-shot settings, and performance remained suboptimal for challenging tasks like disease staging. We will publicly release the dataset, curation pipeline, and leaderboard to potentially advance ophthalmic AI applications and reduce the global burden of vision-threatening diseases.

Anyi Wang, Xuansheng Wu, Dong Shu et al.

Steering has emerged as a promising approach in controlling large language models (LLMs) without modifying model parameters. However, most existing steering methods rely on large-scale datasets to learn clear behavioral information, which limits their applicability in many real-world scenarios. The steering vectors extracted from small dataset often contain task-irrelevant noising features, which degrades their effectiveness. To refine the steering vectors learned from limited data, we introduce Refinement of Steering Vector via Sparse Autoencoder (SAE-RSV) that leverages SAEs to semantically denoise and augment the steering vectors. In our framework, we first remove task-irrelevant features according to their semantics provided by SAEs, and then enrich task-relevant features missing from the small dataset through their semantic similarity to the identified relevant features. Extensive experiments demonstrate that the proposed SAE-RSV substantially outperforms all the baseline methods including supervised fine-tuning. Our findings show that effective steering vector can be constructed from limited training data by refining the original steering vector through SAEs.