arXiv subjects

Lun Wang

Publications and source records attributed to Lun Wang.

At least 19 recordsLinked to original sources

arXiv2026

On the Stability of Prompt Ranking in Large Language Model Evaluation

Prompt-based interaction has become a dominant paradigm for using large language models (LLMs), where multiple candidate prompts are evaluated and the top-ranked one is selected for downstream use. This workflow implicitly assumes that prompt rankings are stable under minor variations in evaluation conditions. In this paper, we systematically study prompt ranking stability under common sources of variability, including random seeds and limited evaluation subsets. Across three open-weight LLMs and two benchmark tasks, we find that while overall rank correlations are often moderate to high, the identity of the top-performing prompt frequently changes, leading to unreliable selection decisions. To address this issue, we propose a simple stability-aware selection strategy based on a lower confidence bound, which accounts for both performance and variance. Our results show that this approach improves robustness in unstable settings while remaining competitive in more stable regimes. These findings highlight the importance of accounting for evaluation uncertainty in prompt selection and LLM benchmarking.

cs.CL↗

arXiv2026

Steering LLMs for Culturally Localized Generation

LLMs are deployed globally, yet produce responses biased towards cultures with abundant training data. Existing cultural localization approaches such as prompting or post-training alignment are black-box, hard to control, and do not reveal whether failures reflect missing knowledge or poor elicitation. In this paper, we address these gaps using mechanistic interpretability to uncover and manipulate cultural representations in LLMs. Leveraging sparse autoencoders, we identify interpretable features that encode culturally salient information and aggregate them into Cultural Embeddings (CuE). We use CuE both to analyze implicit cultural biases under underspecified prompts and to construct white-box steering interventions. Across multiple models, we show that CuE-based steering increases cultural faithfulness and elicits significantly rarer, long-tail cultural concepts than prompting alone. Notably, CuE-based steering is complementary to black-box localization methods, offering gains when applied on top of prompt-augmented inputs. This also suggests that models do benefit from better elicitation strategies, and don't necessarily lack long-tail knowledge representation, though this varies across cultures. Our results provide both diagnostic insight into cultural representations in LLMs and a controllable method to steer towards desired cultures.

cs.CL↗

arXiv2026

OpenSage: Self-programming Agent Generation Engine

Agent development kits (ADKs) provide effective platforms and tooling for constructing agents, and their designs are critical to the constructed agents' performance, especially the functionality for agent topology, tools, and memory. However, current ADKs either lack sufficient functional support or rely on humans to manually design these components, limiting agents' generalizability and overall performance. We propose OpenSage, the first ADK that enables LLMs to automatically create agents with self-generated topology and toolsets while providing comprehensive and structured memory support. OpenSage offers effective functionality for agents to create and manage their own sub-agents and toolkits. It also features a hierarchical, graph-based memory system for efficient management and a specialized toolkit tailored to software engineering tasks. Extensive experiments across three state-of-the-art benchmarks with various backbone models demonstrate the advantages of OpenSage over existing ADKs. We also conduct rigorous ablation studies to demonstrate the effectiveness of our design for each component. We believe OpenSage can pave the way for the next generation of agent development, shifting the focus from human-centered to AI-centered paradigms.

cs.AI↗

arXiv2026

rePIRL: Learn PRM with Inverse RL for LLM Reasoning

Process rewards have been widely used in deep reinforcement learning to improve training efficiency, reduce variance, and prevent reward hacking. In LLM reasoning, existing works also explore various solutions for learning effective process reward models (PRM) with or without the help of an expert policy. However, existing methods either rely on strong assumptions about the expert policies (e.g., requiring their reward functions) or suffer intrinsic limitations (e.g., entropy collapse), resulting in weak PRMs or limited generalizability. In this paper, we introduce rePIRL, an inverse RL-inspired framework that learns effective PRMs with minimal assumptions about expert policies. Specifically, we design a dual learning process that updates the policy and the PRM interchangeably. Our learning algorithm has customized techniques to address the challenges of scaling traditional inverse RL to LLMs. We theoretically show that our proposed learning framework can unify both online and offline PRM learning methods, justifying that rePIRL can learn PRMs with minimal assumptions. Empirical evaluations on standardized math and coding reasoning datasets demonstrate the effectiveness of rePIRL over existing methods. We further show the application of our trained PRM in test-time training, test-time scaling, and providing an early signal for training hard problems. Finally, we validate our training recipe and key design choices via a detailed ablation study.

cs.LG↗

arXiv2026

TermiGen: High-Fidelity Environment and Robust Trajectory Synthesis for Terminal Agents

Executing complex terminal tasks remains a significant challenge for open-weight LLMs, constrained by two fundamental limitations. First, high-fidelity, executable training environments are scarce: environments synthesized from real-world repositories are not diverse and scalable, while trajectories synthesized by LLMs suffer from hallucinations. Second, standard instruction tuning uses expert trajectories that rarely exhibit simple mistakes common to smaller models. This creates a distributional mismatch, leaving student models ill-equipped to recover from their own runtime failures. To bridge these gaps, we introduce TermiGen, an end-to-end pipeline for synthesizing verifiable environments and resilient expert trajectories. Termi-Gen first generates functionally valid tasks and Docker containers via an iterative multi-agent refinement loop. Subsequently, we employ a Generator-Critic protocol that actively injects errors during trajectory collection, synthesizing data rich in error-correction cycles. Fine-tuned on this TermiGen-generated dataset, our TermiGen-Qwen2.5-Coder-32B achieves a 31.3% pass rate on TerminalBench. This establishes a new open-weights state-of-the-art, outperforming existing baselines and notably surpassing capable proprietary models such as o4-mini. Dataset is avaiable at https://github.com/ucsb-mlsec/terminal-bench-env.

cs.AI↗

arXiv2026

DevOps-Gym: Benchmarking AI Agents in Software DevOps Cycle

Even though demonstrating extraordinary capabilities in code generation and software issue resolving, AI agents' capabilities in the full software DevOps cycle are still unknown. Different from pure code generation, handling the DevOps cycle in real-world software, including developing, deploying, and managing, requires analyzing large-scale projects, understanding dynamic program behaviors, leveraging domain-specific tools, and making sequential decisions. However, existing benchmarks focus on isolated problems and lack environments and tool interfaces for DevOps. We introduce DevOps-Gym, the first end-to-end benchmark for evaluating AI agents across core DevOps workflows: build and configuration, monitoring, issue resolving, and test generation. DevOps-Gym includes 700+ real-world tasks collected from 30+ projects in Java and Go. We develop a semi-automated data collection mechanism with rigorous and non-trivial expert efforts in ensuring the task coverage and quality. Our evaluation of state-of-the-art models and agents reveals fundamental limitations: they struggle with issue resolving and test generation in Java and Go, and remain unable to handle new tasks such as monitoring and build and configuration. These results highlight the need for essential research in automating the full DevOps cycle with AI agents.

cs.SE↗

arXiv2025

MUSIC: MUlti-Step Instruction Contrast for Multi-Turn Reward Models

Evaluating the quality of multi-turn conversations is crucial for developing capable Large Language Models (LLMs), yet remains a significant challenge, often requiring costly human evaluation. Multi-turn reward models (RMs) offer a scalable alternative and can provide valuable signals for guiding LLM training. While recent work has advanced multi-turn \textit{training} techniques, effective automated \textit{evaluation} specifically for multi-turn interactions lags behind. We observe that standard preference datasets, typically contrasting responses based only on the final conversational turn, provide insufficient signal to capture the nuances of multi-turn interactions. Instead, we find that incorporating contrasts spanning \textit{multiple} turns is critical for building robust multi-turn RMs. Motivated by this finding, we propose \textbf{MU}lti-\textbf{S}tep \textbf{I}nstruction \textbf{C}ontrast (MUSIC), an unsupervised data augmentation strategy that synthesizes contrastive conversation pairs exhibiting differences across multiple turns. Leveraging MUSIC on the Skywork preference dataset, we train a multi-turn RM based on the Gemma-2-9B-Instruct model. Empirical results demonstrate that our MUSIC-augmented RM outperforms baseline methods, achieving higher alignment with judgments from advanced proprietary LLM judges on multi-turn conversations, crucially, without compromising performance on standard single-turn RM benchmarks.

cs.CL↗

arXiv2025

Fantastic Reasoning Behaviors and Where to Find Them: Unsupervised Discovery of the Reasoning Process

Despite the growing reasoning capabilities of recent large language models (LLMs), their internal mechanisms during the reasoning process remain underexplored. Prior approaches often rely on human-defined concepts (e.g., overthinking, reflection) at the word level to analyze reasoning in a supervised manner. However, such methods are limited, as it is infeasible to capture the full spectrum of potential reasoning behaviors, many of which are difficult to define in token space. In this work, we propose an unsupervised framework (namely, RISE: Reasoning behavior Interpretability via Sparse auto-Encoder) for discovering reasoning vectors, which we define as directions in the activation space that encode distinct reasoning behaviors. By segmenting chain-of-thought traces into sentence-level 'steps' and training sparse auto-encoders (SAEs) on step-level activations, we uncover disentangled features corresponding to interpretable behaviors such as reflection and backtracking. Visualization and clustering analyses show that these behaviors occupy separable regions in the decoder column space. Moreover, targeted interventions on SAE-derived vectors can controllably amplify or suppress specific reasoning behaviors, altering inference trajectories without retraining. Beyond behavior-specific disentanglement, SAEs capture structural properties such as response length, revealing clusters of long versus short reasoning traces. More interestingly, SAEs enable the discovery of novel behaviors beyond human supervision. We demonstrate the ability to control response confidence by identifying confidence-related vectors in the SAE decoder space. These findings underscore the potential of unsupervised latent discovery for both interpreting and controllably steering reasoning in LLMs.

cs.CL↗

arXiv2025

Eliciting Behaviors in Multi-Turn Conversations

Identifying specific and often complex behaviors from large language models (LLMs) in conversational settings is crucial for their evaluation. Recent work proposes novel techniques to find natural language prompts that induce specific behaviors from a target model, yet they are mainly studied in single-turn settings. In this work, we study behavior elicitation in the context of multi-turn conversations. We first offer an analytical framework that categorizes existing methods into three families based on their interactions with the target model: those that use only prior knowledge, those that use offline interactions, and those that learn from online interactions. We then introduce a generalized multi-turn formulation of the online method, unifying single-turn and multi-turn elicitation. We evaluate all three families of methods on automatically generating multi-turn test cases. We investigate the efficiency of these approaches by analyzing the trade-off between the query budget, i.e., the number of interactions with the target model, and the success rate, i.e., the discovery rate of behavior-eliciting inputs. We find that online methods can achieve an average success rate of 45/19/77% with just a few thousand queries over three tasks where static methods from existing multi-turn conversation benchmarks find few or even no failure cases. Our work highlights a novel application of behavior elicitation methods in multi-turn conversation evaluation and the need for the community to move towards dynamic benchmarks.

cs.CL↗

arXiv2025

Attacker's Noise Can Manipulate Your Audio-based LLM in the Real World

This paper investigates the real-world vulnerabilities of audio-based large language models (ALLMs), such as Qwen2-Audio. We first demonstrate that an adversary can craft stealthy audio perturbations to manipulate ALLMs into exhibiting specific targeted behaviors, such as eliciting responses to wake-keywords (e.g., "Hey Qwen"), or triggering harmful behaviors (e.g. "Change my calendar event"). Subsequently, we show that playing adversarial background noise during user interaction with the ALLMs can significantly degrade the response quality. Crucially, our research illustrates the scalability of these attacks to real-world scenarios, impacting other innocent users when these adversarial noises are played through the air. Further, we discuss the transferrability of the attack, and potential defensive measures.

cs.CR↗

arXiv2025

Comparative Analysis of Large Language Models for Context-Aware Code Completion using SAFIM Framework

The advent of Large Language Models (LLMs) has revolutionized code completion, transforming it into a more intelligent and context-aware feature in modern integrated development environments. These advancements have significantly enhanced developers' ability to write efficient and error-free code. This study evaluates the performance of several chat-based LLMs, including Gemini 1.5 Flash, Gemini 1.5 Pro, GPT-4o, GPT-4o-mini, and GPT-4 Turbo, using the Syntax-Aware Fill-in-the-Middle (SAFIM) dataset. This benchmark is specifically designed to assess models' capabilities in syntax-sensitive code generation. Performance metrics, such as cosine similarity with ground-truth completions and latency, were employed to measure both accuracy and efficiency. The findings reveal substantial differences in the models' code completion abilities, offering valuable insights into their respective strengths and weaknesses. This work provides a comparative analysis that underscores the trade-offs between accuracy and speed, establishing a benchmark for future advancements in LLM-based code completion.

cs.SE↗

arXiv2025

Performance Review on LLM for solving leetcode problems

This paper presents a comprehensive performance evaluation of Large Language Models (LLMs) in solving programming challenges from Leetcode, a widely used platform for algorithm practice and technical interviews. We began by crawling the Leetcode website to collect a diverse set of problems encompassing various difficulty levels and topics. Using this dataset, we generated solutions with multiple LLMs, including GPT-4 and GPT-3.5-turbo (ChatGPT-turbo). The generated solutions were systematically evaluated for correctness and efficiency. We employed the pass@k metric to assess the success rates within a given number of attempts and analyzed the runtime performance of the solutions. Our results highlight the strengths and limitations of current LLMs [10] in code generation and problem-solving tasks, providing insights into their potential applications and areas for improvement in automated programming assistance.

cs.SE↗

arXiv2025

Deep Semantic Graph Learning via LLM based Node Enhancement

Graph learning has attracted significant attention due to its widespread real-world applications. Current mainstream approaches rely on text node features and obtain initial node embeddings through shallow embedding learning using GNNs, which shows limitations in capturing deep textual semantics. Recent advances in Large Language Models (LLMs) have demonstrated superior capabilities in understanding text semantics, transforming traditional text feature processing. This paper proposes a novel framework that combines Graph Transformer architecture with LLM-enhanced node features. Specifically, we leverage LLMs to generate rich semantic representations of text nodes, which are then processed by a multi-head self-attention mechanism in the Graph Transformer to capture both local and global graph structural information. Our model utilizes the Transformer's attention mechanism to dynamically aggregate neighborhood information while preserving the semantic richness provided by LLM embeddings. Experimental results demonstrate that the LLM-enhanced node features significantly improve the performance of graph learning models on node classification tasks. This approach shows promising results across multiple graph learning tasks, offering a practical direction for combining graph networks with language models.

cs.AI↗

arXiv2024

AltGen: AI-Driven Alt Text Generation for Enhancing EPUB Accessibility

Digital accessibility is a cornerstone of inclusive content delivery, yet many EPUB files fail to meet fundamental accessibility standards, particularly in providing descriptive alt text for images. Alt text plays a critical role in enabling visually impaired users to understand visual content through assistive technologies. However, generating high-quality alt text at scale is a resource-intensive process, creating significant challenges for organizations aiming to ensure accessibility compliance. This paper introduces AltGen, a novel AI-driven pipeline designed to automate the generation of alt text for images in EPUB files. By integrating state-of-the-art generative models, including advanced transformer-based architectures, AltGen achieves contextually relevant and linguistically coherent alt text descriptions. The pipeline encompasses multiple stages, starting with data preprocessing to extract and prepare relevant content, followed by visual analysis using computer vision models such as CLIP and ViT. The extracted visual features are enriched with contextual information from surrounding text, enabling the fine-tuned language models to generate descriptive and accurate alt text. Validation of the generated output employs both quantitative metrics, such as cosine similarity and BLEU scores, and qualitative feedback from visually impaired users. Experimental results demonstrate the efficacy of AltGen across diverse datasets, achieving a 97.5% reduction in accessibility errors and high scores in similarity and linguistic fidelity metrics. User studies highlight the practical impact of AltGen, with participants reporting significant improvements in document usability and comprehension. Furthermore, comparative analyses reveal that AltGen outperforms existing approaches in terms of accuracy, relevance, and scalability.

cs.AI↗

arXiv2024

Comparative Analysis of Listwise Reranking with Large Language Models in Limited-Resource Language Contexts

Large Language Models (LLMs) have demonstrated significant effectiveness across various NLP tasks, including text ranking. This study assesses the performance of large language models (LLMs) in listwise reranking for limited-resource African languages. We compare proprietary models RankGPT3.5, Rank4o-mini, RankGPTo1-mini and RankClaude-sonnet in cross-lingual contexts. Results indicate that these LLMs significantly outperform traditional baseline methods such as BM25-DT in most evaluation metrics, particularly in nDCG@10 and MRR@100. These findings highlight the potential of LLMs in enhancing reranking tasks for low-resource languages and offer insights into cost-effective solutions.

cs.CL↗

arXiv2024

Robustness of Large Language Models Against Adversarial Attacks

The increasing deployment of Large Language Models (LLMs) in various applications necessitates a rigorous evaluation of their robustness against adversarial attacks. In this paper, we present a comprehensive study on the robustness of GPT LLM family. We employ two distinct evaluation methods to assess their resilience. The first method introduce character-level text attack in input prompts, testing the models on three sentiment classification datasets: StanfordNLP/IMDB, Yelp Reviews, and SST-2. The second method involves using jailbreak prompts to challenge the safety mechanisms of the LLMs. Our experiments reveal significant variations in the robustness of these models, demonstrating their varying degrees of vulnerability to both character-level and semantic-level adversarial attacks. These findings underscore the necessity for improved adversarial training and enhanced safety mechanisms to bolster the robustness of LLMs.

cs.CL↗

arXiv2024

RareAgents: Autonomous Multi-disciplinary Team for Rare Disease Diagnosis and Treatment

Rare diseases, despite their low individual incidence, collectively impact around 300 million people worldwide due to the vast number of diseases. The involvement of multiple organs and systems, and the shortage of specialized doctors with relevant experience, make diagnosing and treating rare diseases more challenging than common diseases. Recently, agents powered by large language models (LLMs) have demonstrated notable applications across various domains. In the medical field, some agent methods have outperformed direct prompts in question-answering tasks from medical examinations. However, current agent frameworks are not well-adapted to real-world clinical scenarios, especially those involving the complex demands of rare diseases. To bridge this gap, we introduce RareAgents, the first LLM-driven multi-disciplinary team decision-support tool designed specifically for the complex clinical context of rare diseases. RareAgents integrates advanced Multidisciplinary Team (MDT) coordination, memory mechanisms, and medical tools utilization, leveraging Llama-3.1-8B/70B as the base model. Experimental results show that RareAgents outperforms state-of-the-art domain-specific models, GPT-4o, and current agent frameworks in diagnosis and treatment for rare diseases. Furthermore, we contribute a novel rare disease dataset, MIMIC-IV-Ext-Rare, to facilitate further research in this field.

cs.CL↗

arXiv2024

Can DeepFake Speech be Reliably Detected?

Recent advances in text-to-speech (TTS) systems, particularly those with voice cloning capabilities, have made voice impersonation readily accessible, raising ethical and legal concerns due to potential misuse for malicious activities like misinformation campaigns and fraud. While synthetic speech detectors (SSDs) exist to combat this, they are vulnerable to ``test domain shift", exhibiting decreased performance when audio is altered through transcoding, playback, or background noise. This vulnerability is further exacerbated by deliberate manipulation of synthetic speech aimed at deceiving detectors. This work presents the first systematic study of such active malicious attacks against state-of-the-art open-source SSDs. White-box attacks, black-box attacks, and their transferability are studied from both attack effectiveness and stealthiness, using both hardcoded metrics and human ratings. The results highlight the urgent need for more robust detection methods in the face of evolving adversarial threats.

cs.SD↗