Large language models (LLMs) often appear to excel on public benchmarks, but these high scores may mask an overreliance on dataset-specific surface cues rather than true language understanding. We introduce the **Chameleon Benchmark Overfit Detector (C-BOD)**, a meta-evaluation framework designed to reveal such overfitting. C-BOD systematically rephrases benchmark inputs via a parameterized transformation that preserves semantic content and labels, enabling the detection of performance degradation indicative of superficial pattern reliance.We conduct extensive experiments across two datasets, three rephrasing models, and multiple distortion levels, evaluating 32 state-of-the-art LLMs. On the MMLU benchmark, C-BOD reveals an average performance drop of 2.75% under modest rephrasings, with over 80% of models exhibiting statistically significant differences. Notably, higher-performing models and larger LLMs tend to show greater sensitivity, suggesting a deeper dependence on benchmark-specific phrasing.Due to its dataset and model-agnostic design, C-BOD can be easily integrated into evaluation pipelines and offers a promising foundation for overfitting mitigation strategies. Our findings challenge the community to look beyond leaderboard scores and prioritize resilience and generalization in LLM evaluation. Our code and benchmark datasets are availableat: https://github.com/nuritci/cbod
论文检索
输入标题、作者或关键词，从 12,319 篇学术成果中精准定位
Large Action models are essential for enabling autonomous agents to perform complex tasks. However, training such models remains challenging due to the diversity of agent environments and the complexity of noisy agentic data. Existing infrastructure offers limited support for scalable, agent-specific fine-tuning and standardized agent data processing. We introduce ActionStudio, a lightweight and extensible data and training framework designed for large action models. ActionStudio unifies diverse agent trajectories using our proposed Unified Format 2.0, supports a range of training workflows with optimized multi-node distributed setup, and integrates robust preprocessing and real-time verification tools. ActionStudio demonstrates up to 9\times higher throughput compared to existing agentic training frameworks, and our trained models yield top performances across public and realistic agent benchmarks. To support the broader research community, we open-source the ActionStudio framework and release actionstudio-98k, a curated dataset of 98k high-quality trajectories.
PLAN-TUNING: Post-Training Language Models to Learn Step-by-Step Planning for Complex Problem Solving
PDF ↗Recently, decomposing complex problems into simple subtasks–a crucial part of human-like natural planning–to solve the given problem has significantly boosted the performance of large language models (LLMs). However, leveraging such planning structures during post-training to boost the performance of smaller open-source LLMs remains underexplored. Motivated by this, we introduce PLAN-TUNING, a unified post-training framework that (i) distills synthetic task decompositions (termed “planning trajectories”) from large-scale LLMs and (ii) fine-tunes smaller models via supervised and reinforcement-learning objectives designed to mimic these planning processes to improve complex reasoning. On GSM8k and the MATH benchmarks, plan-tuned models outperform strong baselines by an average ~7%. Furthermore, plan-tuned models show better generalization capabilities on out-of-domain datasets, with average ~10% and ~12% performance improvements on OlympiadBench and AIME 2024, respectively. Our detailed analysis demonstrates how planning trajectories improves complex reasoning capabilities, showing that PLAN-TUNING is an effective strategy for improving task-specific performance of smaller LLMs.
Traditional information retrieval (IR) methods excel at textual and semantic matching but struggle in reasoning-intensive retrieval tasks that require multi-hop inference or complex semantic understanding between queries and documents. One promising solution is to explicitly rewrite or augment queries using large language models (LLMs) to elicit reasoning-relevant content prior to retrieval. However, the widespread use of large-scale LLMs like GPT-4 or LLaMA3-70B remains impractical due to their high inference cost and limited deployability in real-world systems. In this work, we introduce Reinforced Query Reasoner (RQR), a family of small-scale language models for query reasoning and rewriting in reasoning-intensive retrieval. Our approach frames query reformulation as a reinforcement learning problem and employs a novel semi-rule-based reward function. This enables smaller language models, e.g., Qwen2.5-7B-Instruct and Qwen2.5-1.5B-Instruct, to achieve reasoning performance rivaling large-scale LLMs without their prohibitive inference costs. Experiment results on BRIGHT benchmark show that, with BM25 as retrievers, both RQR-7B and RQR-1.5B models significantly outperform existing baselines, including prompt-based query reasoners and some latest dense retrievers trained for reasoning-intensive retrieval tasks, offering superior adaptability for real-world deployment. All code and dataset will be publicly released.
A Sequential Multi-Stage Approach for Code Vulnerability Detection via Confidence- and Collaboration-based Decision Making
PDF ↗While large language models (LLMs) have shown strong capabilities across diverse domains, their application to code vulnerability detection holds great potential for identifying security flaws and improving software safety. In this paper, we propose a sequential multi-stage approach via confidence- and collaboration-based decision making (ConfColl). The system adopts a three-stage sequential classification framework, proceeding through a single agent, retrieval-augmented generation (RAG) with external examples, and multi-agent reasoning enhanced with RAG. The decision process selects among these strategies to balance performance and cost, with the process terminating at any stage where a high-certainty prediction is achieved. Experiments on a benchmark dataset and a low-resource language demonstrate the effectiveness of our framework in enhancing code vulnerability detection performance.
Enhancing Large Language Model for Knowledge Graph Completion via Structure-Aware Alignment-Tuning
PDF ↗Knowledge graph completion (KGC) aims to infer new knowledge and make predictions from knowledge graphs. Recently, large language models (LLMs) have exhibited remarkable reasoning capabilities. LLM-enhanced KGC methods primarily focus on designing task-specific instructions, achieving promising advancements. However, there are still two critical challenges. First, existing methods often ignore the inconsistent representation spaces between natural language and graph structures. Second, most approaches develop separate instructions for different KGC tasks, leading to duplicate works and time-consuming processes. To address these challenges, we propose SAT, a novel framework that enhances LLMs for KGC via structure-aware alignment-tuning. Specifically, we first introduce hierarchical knowledge alignment to align graph embeddings with the natural language space through multi-task contrastive learning. Then, we propose structural instruction tuning to guide LLMs in performing structure-aware reasoning over KGs, using a unified graph instruction combined with a lightweight knowledge adapter. Experimental results on two KGC tasks across four benchmark datasets demonstrate that SAT significantly outperforms state-of-the-art methods, especially in the link prediction task with improvements ranging from 8.7% to 29.8%
The quality of supervised fine-tuning (SFT) data is crucial for the performance of large multimodal models (LMMs), yet current data enhancement methods often suffer from factual errors and hallucinations due to inadequate visual perception. To address this challenge, we propose VERITAS, a pipeline that systematically integrates vision priors and multiple state-of-the-art LMMs with statistical methods to enhance SFT data quality. VERITAS leverages visual recognition models (RAM++) and OCR systems (PP-OCRv4) to extract structured vision priors, which are combined with images, questions, and answers. Three LMMs (GPT-4o, Gemini-2.5-Pro, Doubao-1.5-pro) evaluate the original answers, providing critique rationales and scores that are statistically fused into a high-confidence consensus score serving as ground truth. Using this consensus, we train a lightweight critic model via Group Relative Policy Optimization (GRPO), enhancing reasoning capabilities efficiently. Each LMM then refines the original answers based on the critiques, generating new candidate answers; we select the highest-scoring one as the final refined answer. Experiments across six multimodal benchmarks demonstrate that models fine-tuned with data processed by VERITAS consistently outperform those using raw data, particularly in text-rich and fine-grained reasoning tasks. Our critic model exhibits enhanced capability comparable to state-of-the-art LMMs while being significantly more efficient. We release our pipeline, datasets, and model checkpoints to advance research in multimodal data optimization.
Foundation multi-modal models are often designed by stitching of multiple existing pretrained uni-modal models: for example, an image classifier with a text model. This stitching process is performed by training a connector module that aims to align the representation spaces of these uni-modal models towards a multi-modal objective. However, given the complexity of training such connectors on large scale web-based datasets coupled with the ever-increasing number of available pretrained uni-modal models, the task of uni-modal models selection and subsequent connector module training becomes computationally demanding. To address this under-studied critical problem, we propose Hypernetwork Model Alignment (Hyma), a novel all-in-one solution for optimal uni-modal model selection and connector training by leveraging hypernetworks. Specifically, our framework utilizes the parameter prediction capability of a hypernetwork to obtain jointly trained connector modules for N \times M combinations of uni-modal models. In our experiments, Hyma reduces the cost of searching for the best performing uni-modal model pair by 10\times, while matching the ranking and trained connector performance obtained via grid search across a suite of diverse multi-modal benchmarks.
Large language models (LLMs) offer an inexpensive yet powerful way to annotate text, but are often inconsistent when compared with experts. These errors can bias downstream estimates of population parameters such as regression coefficients and causal effects. To mitigate this bias, researchers have developed debiasing methods such as Design-based Supervised Learning (DSL) and Prediction-Powered Inference (PPI), which promise valid estimation by combining LLM annotations with a limited number of expensive expert annotations.Although these methods produce consistent estimates under theoretical assumptions, it is unknown how they compare in finite samples of sizes encountered in applied research. We make two contributions: First, we study how each method’s performance scales with the number of expert annotations, highlighting regimes where LLM bias or limited expert labels significantly affect results. Second, we compare DSL and PPI across a range of tasks, finding that although both achieve low bias with large datasets, DSL often outperforms PPI on bias reduction and empirical efficiency, but its performance is less consistent across datasets. Our findings indicate that there is a bias-variance tradeoff at the level of debiasing methods, calling for more research on developing metrics for quantifying their efficiency in finite samples.
Agentic Retrieval-Augmented Generation (RAG) systems enhance Large Language Models (LLMs) by enabling dynamic, multi-step reasoning and information retrieval. However, these systems often exhibit sub-optimal search behaviors like over-search (retrieving redundant information) and under-search (failing to retrieve necessary information), which hinder efficiency and reliability. This work formally defines and quantifies these behaviors, revealing their prevalence across multiple QA datasets and agentic RAG systems (e.g., one model could have avoided searching in 27.7% of its search steps). Furthermore, we demonstrate a crucial link between these inefficiencies and the models’ uncertainty regarding their own knowledge boundaries, where response accuracy correlates with model’s uncertainty in its search decisions. To address this, we propose β-GRPO, a reinforcement learning-based training method that incorporates confidence threshold to reward high-certainty search decisions. Experiments on seven QA benchmarks show that β-GRPO enable a 3B model with better agentic RAG ability, outperforming other strong baselines with a 4% higher average exact match score.
Digital social media platforms frequently contribute to cognitive-behavioral fixation, a phenomenon in which users exhibit sustained and repetitive engagement with narrow content domains. While cognitive-behavioral fixation has been extensively studied in psychology, methods for computationally detecting and evaluating such fixation remain underexplored. To address this gap, we propose a novel framework for assessing cognitive-behavioral fixation by analyzing users’ multimodal social media engagement patterns. Specifically, we introduce a multimodal topic extraction module and a cognitive-behavioral fixation quantification module that collaboratively enable adaptive, hierarchical, and interpretable assessment of user behavior. Experiments on existing benchmarks and a newly curated multimodal dataset demonstrate the effectiveness of our approach, laying the groundwork for scalable computational analysis of cognitive fixation. All code in this project is publicly available for research purposes at https://github.com/Liskie/cognitive-fixation-evaluation.
Mitigating Hallucinations in Large Vision-Language Models via Entity-Centric Multimodal Preference Optimization
PDF ↗Large Visual Language Models (LVLMs) have demonstrated impressive capabilities across multiple tasks. However, their trustworthiness is often challenged by hallucinations, which can be attributed to the modality misalignment and the inherent hallucinations of their underlying Large Language Models (LLMs) backbone. Existing preference alignment methods focus on aligning model responses with human preferences while neglecting image-text modality alignment, resulting in over-reliance on LLMs and hallucinations. In this paper, we propose Entity-centric Multimodal Preference Optimization (EMPO), which achieves enhanced modality alignment than existing human preference alignment methods. Besides, to overcome the scarcity of high-quality multimodal preference data, we utilize open-source instruction datasets to automatically construct high-quality preference data across three aspects: image, instruction, and response. Experiments on two human preference datasets and five multimodal hallucination benchmarks demonstrate the effectiveness of EMPO, e.g., reducing hallucination rates by 80.4% on Object HalBench and 52.6% on MM HalBench, thereby enhancing the trustworthiness of LVLMs. The code and dataset will be made publicly available.
LLM-Driven Completeness and Consistency Evaluation for Cultural Heritage Data Augmentation in Cross-Modal Retrieval
PDF ↗Cross-modal retrieval is essential for interpreting cultural heritage data, but its effectiveness is often limited by incomplete or inconsistent textual descriptions, caused by historical data loss and the high cost of expert annotation. While large language models (LLMs) offer a promising solution by enriching textual descriptions, their outputs frequently suffer from hallucinations or miss visually grounded details. To address these challenges, we propose C^3, a data augmentation framework that enhances cross-modal retrieval performance by improving the completeness and consistency of LLM-generated descriptions. C^3 introduces a bidirectional validation mechanism to assess semantic coverage using both visual cues and language-model outputs. Furthermore, to mitigate factual inconsistencies, we formulate a Markov Decision Process to supervise Chain-of-Thought reasoning, guiding consistency verification through adaptive query control. Experiments on the cultural heritage dataset CulTi and general benchmarks MSCOCO and Flickr30K demonstrate that C^3 achieves state-of-the-art performance in both fine-tuned and zero-shot settings.
Graph-theoretic problems arise in real-world applications like logistics, communication networks, and traffic optimization. These problems are often complex, noisy, and irregular, posing challenges for traditional algorithms. Large language models offer potential solutions but face several challenges, including limited accuracy, input length constraints, and suboptimal algorithm selection. To address these challenges, we propose MA-GTS(Multi-Agent Graph Theory Solver), a multi-agent framework that decomposes these complex problems through agent collaboration. MA-GTS maps the implicitly expressed text-based graph data into clear, structured graph representations and dynamically selects the most suitable algorithm based on problem constraints and graph structure scale. We validate MA-GTS using the G-REAL dataset, a real-world-inspired graph theory dataset we created. Experimental results show that MA-GTS outperforms state-of-the-art methods in cost-effectiveness, accuracy, and scalability, achieving strong results on multiple benchmarks (G-REAL 93.6%, GraCoRe 96.9% ,NLGraph 98.4%) with robust performance on both closed- and open-source models.
Fine-tuning multi-turn dialogue systems requires high-quality supervision but often suffers from degraded performance when exposed to low-quality data. Supervision errors in early turns can propagate across subsequent turns, undermining coherence and response quality. Existing methods typically address data quality via static prefiltering, which decouples quality control from training and fails to mitigate turn-level error propagation. In this context, we propose **ReSURE** (REgularizing Supervision UnREliability), an adaptive learning method that dynamically down-weights unreliable supervision without explicit filtering. ReSURE estimates per-turn loss distributions using Welford’s online statistics and reweights sample losses on the fly accordingly. Experiments on both single-source and mixed-quality datasets show improved stability and response quality. Notably, ReSURE enjoys positive Spearman correlations (0.21 ~ 1.0 across multiple benchmarks) between response scores and number of samples regardless of data quality, which potentially paves the way for utilizing large-scale data effectively.
We introduce Debate Speech Evaluation as a novel and challenging benchmark for assessing LLM judges. Evaluating debate speeches requires a deep understanding of the speech at multiple levels, including argument strength and relevance, the coherence and organization of the speech, the appropriateness of its style and tone, and so on. This task involves a unique set of cognitive abilities that previously received limited attention in systematic LLM benchmarking. To explore such skills, we leverage a dataset of over 600 meticulously annotated debate speeches and present the first in-depth analysis of how state-of-the-art LLMs compare to human judges on this task. Our findings reveal a nuanced picture: while larger models can approximate individual human judgments in some respects, they differ substantially in their overall judgment behavior. We also investigate the ability of frontier LLMs to generate persuasive, opinionated speeches, showing that models may perform at a human level on this task.
The performance of large language models (LLMs) is closely tied to their training data, which can include copyrighted material or private information, raising legal and ethical concerns. Additionally, LLMs face criticism for dataset contamination and internalizing biases. To address these issues, the Pre-Training Data Detection (PDD) task was proposed to identify if specific data was included in an LLM’s pre-training corpus. However, existing PDD methods often rely on superficial features like prediction confidence and loss, resulting in mediocre performance. To improve this, we introduce NA-PDD, a novel algorithm analyzing differential neuron activation patterns between training and non-training data in LLMs. This is based on the observation that these data types activate different neurons during LLM inference. We also introduce CCNewsPDD, a temporally unbiased benchmark employing rigorous data transformations to ensure consistent time distributions between training and non-training data. Our experiments demonstrate that NA-PDD significantly outperforms existing methods across three benchmarks and multiple LLMs.
Retrieval over Classification: Integrating Relation Semantics for Multimodal Relation Extraction
PDF ↗Relation extraction (RE) aims to identify semantic relations between entities in unstructured text. Although recent work extends traditional RE to multimodal scenarios, most approaches still adopt classification-based paradigms with fused multimodal features, representing relations as discrete labels. This paradigm has two significant limitations: (1) it overlooks structural constraints like entity types and positional cues, and (2) it lacks semantic expressiveness for fine-grained relation understanding. We propose **R**etrieval **O**ver **C**lassification (ROC), a novel framework that reformulates multimodal RE as a retrieval task driven by relation semantics. ROC integrates entity type and positional information through a multimodal encoder, expands relation labels into natural language descriptions using a large language model, and aligns entity-relation pairs via semantic similarity-based contrastive learning. Experiments show that our method achieves state-of-the-art performance on the benchmark datasets MNRE and MORE and exhibits stronger robustness and interpretability.
TrojanWave: Exploiting Prompt Learning for Stealthy Backdoor Attacks on Large Audio-Language Models
PDF ↗Prompt learning has emerged as an efficient alternative to full fine-tuning for adapting large audio-language models (ALMs) to downstream tasks. While this paradigm enables scalable deployment via Prompt-as-a-Service frameworks, it also introduces a critical yet underexplored security risk of backdoor attacks. In this work, we present TrojanWave, the first backdoor attack tailored to the prompt-learning setting in frozen ALMs. Unlike prior audio backdoor methods that require training from scratch on full datasets, TrojanWave injects backdoors solely through learnable prompts, making it highly scalable and effective in few-shot settings. TrojanWave injects imperceptible audio triggers in both time and spectral domains to effectively induce targeted misclassification during inference. To mitigate this threat, we further propose TrojanWave-Defense, a lightweight prompt purification method that neutralizes malicious prompts without hampering the clean performance. Extensive experiments across 11 diverse audio classification benchmarks demonstrate the robustness and practicality of both the attack and defense. Our code is publicly available at https://asif-hanif.github.io/trojanwave/.
Although the power of LLM tool-use agents has ignited a flurry of recent research in this area, the curation of tool-use training data remains an open problem\textemdashespecially for online RL training. Existing approaches to synthetic tool-use data generation tend to be non-interactive and/or non-compositional. We introduce RandomWorld, a pipeline for the procedural generation of interactive tools and compositional tool-use data. We show that models tuned via SFT and RL on synthetic RandomWorld data improve on a range of tool-use benchmarks, and set the new SoTA for two metrics on the NESTFUL dataset. Further experiments show that downstream performance scales with the amount of RandomWorld-generated training data, opening up the possibility of further improvement through the use of entirely synthetic data.