paperlist.ai

AI conference papers

AI conference paper library

Search papers from 30+ top conferences, including NeurIPS, ICLR, ICML, CVPR, ACL and AAAI. Filter by keyword, then open any paper for its abstract, authors and links.

arXiv

Poisoning Prompt-Guided Sampling in Video Large Language Models

Yuxin Cao, Wei Song, Jingling Xue, Jin Song Dong

Video Large Language Models (VideoLLMs) are increasingly deployed as automated moderators on user-generated video platforms, where a few unwatched seconds of harmful footage are enough to suppress a safety alert. Because encoding every frame is prohibitive, modern VideoLLMs rely on prompt-guided sampling (PGS), which scores frames against the user prompt and forwards only the top-ranked ones to the visual encoder. Uniform and semantic samplers are known to be defeated by simple frame replacement, whereas PGS, the most prompt-aware family, has escaped scrutiny, and its prompt awareness in fact repairs the omission failures that defeat the other two. We show that this repair is superficial, since PoisonVID, a transfer attack, poisons the sampler's ranking so that harmful clips are never surfaced, without access to target weights, gradients, or sampling internals. It optimizes one video-level perturbation under a relevance-suppression loss defined over a depiction set of paraphrased harmful descriptions written by a shadow VideoLLM and a general-purpose language model, which drives perturbed harmful frames out of the prompt-conditioned subspace that PGS reads. Samplers that never consult that score keep the frames they always kept, which locates the failure at selection rather than at the encoder. Across three PGS methods, six VideoLLMs, and six harmful categories, PoisonVID attains 84% to 97% average attack success over the 18 sampler and model pairs and survives seven defenses. Re-encoding at lower resolution on ingest gives back part of what was evicted and costs the attack 48 points, which bounds the threat without closing it. PGS therefore buys accuracy with a structural safety debt, and sampler design will now have to repay that.

arXiv

MeSS: City Mesh-Guided Outdoor Scene Generation with Cross-View Consistent Diffusion

Xuyang Chen, Zhijun Zhai, Kaixuan Zhou, Zengmao Wang, Jianan He, Dong Wang, Yanfeng Zhang, Mingwei Sun, rüdiger westermann, Konrad Schindler, Liqiu Meng

Mesh models have become increasingly accessible for numerous cities; however, the lack of realistic textures restricts their application in virtual urban navigation and autonomous driving. To address this, this paper proposes MeSS (Meshbased Scene Synthesis) for generating high-quality, styleconsistent outdoor scenes with city mesh models serving as the geometric prior. While image and video diffusion models can leverage spatial layouts (such as depth maps or HD maps) as control conditions to generate street-level perspective views, they are not directly applicable to 3D scene generation. Video diffusion models excel at synthesizing consistent view sequences that depict scenes but often struggle to adhere to predefined camera paths or align accurately with rendered control videos. In contrast, image diffusion models, though unable to guarantee cross-view visual consistency, can produce more geometry-aligned results when combined with ControlNet. Building on this insight, our approach enhances image diffusion models by improving cross-view consistency. The pipeline comprises three key stages: first, we generate geometrically consistent sparse views using Cascaded Outpainting ControlNets; second, we propagate denser intermediate views via a component dubbed AGInpaint; and third, we globally eliminate visual inconsistencies (e.g., varying exposure) using the GCAlign module. Concurrently with generation, a 3D Gaussian Splatting (3DGS) scene is reconstructed by initializing Gaussian balls on the mesh surface. Our method outperforms existing approaches in both geometric alignment and generation quality. Once synthesized, the scene can be rendered in diverse styles through relighting and style transfer techniques. project page: https://albertchen98.github.io/mess/

arXiv

Two-Way Garment Transfer: Unified Diffusion Framework for Dressing and Undressing Synthesis

Angang Zhang, Fang Deng, Hao Chen, Zhongjian Chen, Junyan Li

While recent advances in virtual try-on (VTON) have achieved realistic garment transfer to human subjects, its inverse task, virtual try-off (VTOFF), which aims to reconstruct canonical garment templates from dressed humans, remains critically underexplored and lacks systematic investigation. Existing works predominantly treat them as isolated tasks: VTON focuses on garment dressing while VTOFF addresses garment extraction, thereby neglecting their complementary symmetry. To bridge this fundamental gap, we propose the Two-Way Garment Transfer Model (TWGTM), to the best of our knowledge, the first unified framework for joint clothing-centric image synthesis that simultaneously resolves both mask-guided VTON and mask-free VTOFF through bidirectional feature disentanglement. Specifically, our framework employs dual-conditioned guidance from both latent and pixel spaces of reference images to seamlessly bridge the dual tasks. On the other hand, to resolve the inherent mask dependency asymmetry between mask-guided VTON and mask-free VTOFF, we devise a phased training paradigm that progressively bridges this modality gap. Extensive qualitative and quantitative experiments conducted across the DressCode and VITON-HD datasets validate the efficacy and competitive edge of our proposed approach.

arXiv

What is the Right Embedding Space for Contrastive Learning in Referring Expression Counting?

Kostas Triaridis, Panagiotis Kaliosis, E-Ro Nguyen, Jingyi Xu, Dimitris Samaras, Hieu Le

Referring Expression Counting (REC) requires distinguishing visually similar objects described by fine-grained text cues. Existing methods tackle this via image-text contrastive learning of visual features which aims to distinguish visual features corresponding to positive and negative referring expressions. However, this strategy relies heavily on accurate image-text alignment and is constrained by the limited number of available negative expressions. We argue that these limitations can be avoided by performing contrastive learning entirely in the visual embedding space. To this end, we introduce C-REX, a simple but effective supervised contrastive learning framework that learns fine-grained visual representations by contrasting visual tokens within the same image. By shifting the negative samples from incorrect Referring Expressions to visual tokens from the image, C-REX vastly increases the number of potential negatives, providing richer and more stable supervision. This design leads to stronger fine-grained visual discrimination and better generalization across complex open-world counting settings. C-REX serves as a plug-in module that can be seamlessly applied to any existing REC model without architectural changes. We apply it in three different REC architectures and achieve state-of-the-art results improving by up to 28% in MAE and 24.5% in RMSE. Moreover, we show that our framework is versatile and general, and can be applied to other counting tasks like class-agnostic counting, improving the performance of existing models.

arXiv

VLC Fusion: Vision-Language Conditioned Sensor Fusion for Robust Object Detection

Aditya Taparia, Noel Ngu, Mario Leiva, Joshua Shay Kricheli, John Corcoran, Nathaniel D. Bastian, Gerardo Simari, Paulo Shakarian, Ransalu Senanayake

Although fusing multiple sensor modalities can enhance object detection performance, existing fusion approaches often overlook subtle variations in environmental conditions and sensor inputs. As a result, they struggle to adaptively weight each modality under such variations. To address this challenge, we introduce Vision-Language Conditioned Fusion (VLC Fusion), a novel fusion framework that leverages a Vision-Language Model (VLM) to condition the fusion process on nuanced environmental cues. By capturing high-level environmental context such as darkness, rain, and camera blurring, the VLM guides the model to dynamically adjust modality weights based on the current scene, ensuring robustness against environmental shifts. We evaluate VLC Fusion on real-world autonomous driving and military target detection datasets that include image, LIDAR, and mid-wave infrared modalities. Our experiments show that VLC Fusion consistently outperforms conventional fusion baselines, achieving improved detection accuracy in both seen and unseen scenarios.

arXiv

Learning Attribute-aware Representations for Few-shot Scene Text Segmentation

Yifan Tang, Chenming Li, Chengxu Liu, Yuanting Fan, Dangfeng Yang, Yong Huang, Cun Xin, Yu Li, Xingsong Hou, Xueming Qian

Supervised scene text segmentation has achieved notable progress in recent years. However, its development is largely constrained by the scarcity of high-quality datasets and the high cost of pixel-level annotations. To address this limitation, we explore few-shot learning for text segmentation and propose TSAL, an attribute-aware few-shot framework that leverages a pre-trained CLIP model to learn transferable text attributes for segmentation. Our framework comprises two complementary branches: I) a Visual-Guided Branch that extracts semantic and textural features for foreground text and background regions, respectively, and II) an Adaptive Prompt-Guided Branch that employs learnable prompt templates to capture diverse text attributes with minimal data dependence. To effectively align textual attributes with visual representations, we further introduce an Adaptive Feature Alignment~(AFA) module, which aligns learnable attribute tokens with visual features and prompt prototypes, enabling the model to capture both general and distinctive textual characteristics. As a result, TSAL can accurately segment text regions using only a few annotated samples. Extensive experiments demonstrate that our method achieves state-of-the-art performance across several public text segmentation benchmarks under few-shot settings and exhibits strong generalization to text-related tasks.

arXiv

Compound and Parallel Modes of Tropical Convolutional Neural Networks

Mingbo Li, Liying Liu, Charles Wiranto, Ye Luo

Convolutional neural networks (CNNs) are foundational to many state-of-the-art computer vision systems, yet their reliance on multiplication-intensive computations poses challenges for deployment on resource-constrained devices. While tropical convolutional neural networks (TCNNs) reduce this computational burden by replacing multiplications with cheaper min/maxplus operations, they often do so at the cost of reduced model accuracy. To address this tradeoff, we introduce two novel extensions of tropical convolution: compound tropical convolution (cTCNN) and parallel tropical convolution (pTCNN). These operators combine minplus and maxplus algebraic operations within a single layer to enhance representational capacity while maintaining low computational cost. We provide an open-source implementation of these operators in a PyTorch-compatible framework, featuring optimized GPU kernels developed with TileLang. Through extensive experiments on image classification and semantic segmentation benchmarks, we demonstrate that our proposed cTCNN and pTCNN layers achieve competitive performance against standard CNNs while significantly reducing the number of multiplications. Moreover, we show that hybrid models, which integrate both tropical and conventional convolutions, can further improve the accuracy-efficiency balance. Our findings suggest that these tropical convolution variants are viable and effective components for building efficient deep learning models

arXiv

MaterialFusion: High-Quality, Zero-Shot, and Controllable Material Transfer with Diffusion Models

Kamil Garifullin, Maxim Nikolaev, Andrey Kuznetsov, Aibek Alanov

Manipulating the material appearance of objects in images is critical for applications like augmented reality, virtual prototyping, and digital content creation. We present MaterialFusion, a novel framework for high-quality material transfer that allows users to adjust the degree of material application, achieving an optimal balance between new material properties and the object's original features. MaterialFusion seamlessly integrates the modified object into the scene by maintaining background consistency and mitigating boundary artifacts. To thoroughly evaluate our approach, we have compiled a dataset of real-world material transfer examples and conducted complex comparative analyses. Through comprehensive quantitative evaluations and user studies, we demonstrate that MaterialFusion significantly outperforms existing methods in terms of quality, user control, and background preservation. Code is available at https://github.com/ControlGenAI/MaterialFusion.

arXiv

Toward Visual Grounding: A Survey

Linhui Xiao, Xiaoshan Yang, Xiangyuan Lan, Yaowei Wang, Changsheng Xu

Visual Grounding, also known as Referring Expression Comprehension and Phrase Grounding, aims to ground the specific region(s) within the image(s) based on the given expression text. This task simulates the common referential relationships between visual and linguistic modalities, enabling machines to develop human-like multimodal comprehension capabilities. Consequently, it has extensive applications in various domains. However, since 2021, visual grounding has witnessed significant advancements, with emerging new concepts such as grounded pre-training, grounding multimodal LLMs, generalized visual grounding, and giga-pixel grounding, which have brought numerous new challenges. In this survey, we first examine the developmental history of visual grounding and provide an overview of essential background knowledge. We systematically track and summarize the advancements, and then meticulously define and organize the various settings to standardize future research and ensure a fair comparison. Additionally, we delve into numerous related datasets and applications, and highlight several advanced topics. Finally, we outline the challenges confronting visual grounding and propose valuable directions for future research, which may serve as inspiration for subsequent researchers. By extracting common technical details, this survey encompasses the representative work in each subtopic over the past decade. To the best of our knowledge, this paper represents the most comprehensive overview currently available in the field of visual grounding. This survey is designed to be suitable for both beginners and experienced researchers, serving as an invaluable resource for understanding key concepts and tracking the latest research developments. We keep tracing related work at https://github.com/linhuixiao/Awesome-Visual-Grounding.

arXiv

Efficient unsupervised domain adaptation via self-supervised vision transformer and synergistic cross-domain alignment

Ali Abedi, Q. M. Jonathan Wu, Ning Zhang, Farhad Pourpanah

Unsupervised domain adaptation (UDA) aims to mitigate domain shift, where the distribution of labeled source data differs from that of unlabeled target data. Despite recent advances, existing methods often rely on fine-tuning large backbone models, which leads to high computational cost and limits scalability in resource-constrained environments. This limitation highlights the need for parameter-efficient approaches that maintain strong performance with reduced training complexity. Self-supervised foundation models such as DINOv2 provide highly transferable representations and raise the question of whether effective domain adaptation can be achieved without full fine-tuning. To address this question, we propose Efficient Unsupervised Domain Adaptation (EUDA), a parameter-efficient framework that leverages a frozen DINOv2 backbone as a feature extractor and updates only a lightweight bottleneck and classification head. We also adopt a synergistic domain alignment loss (SDAL), which combines cross-entropy (CE) and maximum mean discrepancy (MMD) to promote both discriminative learning and cross-domain alignment. Experimental results on Office-Home, Office-31, VisDA-2017, and DomainNet demonstrate that EUDA achieves competitive performance across diverse domain complexities, while reducing the number of trainable parameters by 42 to 99.7%. These results show the suitability of the proposed method for resource-constrained and distributed environments.

arXiv

CollaFuse: Collaborative Diffusion Models

Simeon Allmendinger, Domenique Zipperling, Lukas Struppek, Niklas Kühl

In the landscape of generative artificial intelligence, diffusion-based models have emerged as a promising method for generating synthetic images. However, the application of diffusion models poses numerous challenges, particularly concerning data availability, computational requirements, and privacy. Traditional approaches to address these shortcomings, like federated learning, often impose significant computational burdens on individual clients, especially those with constrained resources. In response to these challenges, we introduce the novel approach CollaFuse for distributed collaborative diffusion models inspired by split learning. Our approach facilitates collaborative training of diffusion models while alleviating client computational burdens during image synthesis. This reduced computational burden is achieved by retaining data and computationally inexpensive processes locally at each client while outsourcing the computationally expensive processes to shared, more efficient server resources. Through experiments on the common datasets CelebA, CIFAR-10, and Animals-with-Attributes2, our approach demonstrates enhanced performance while decreasing information disclosure as it reduces the necessity for sharing raw data. These capabilities hold significant potential across various application areas, including the design of edge computing solutions. Thus, our work advances distributed machine learning by contributing to the evolution of collaborative diffusion models.

arXiv

Neural Surface Reconstruction from Sparse Views Using Epipolar Geometry

Xinhai Chang, Kaichen Zhou

Reconstructing accurate surfaces from sparse multi-view images remains challenging due to severe geometric ambiguity and occlusions. Existing generalizable neural surface reconstruction methods primarily rely on cost volumes that summarize multi-view features using simple statistics (e.g., mean and variance), which discard critical view-dependent geometric structure and often lead to over-smoothed reconstructions. We propose EpiS, a generalizable neural surface reconstruction framework that explicitly leverages epipolar geometry for sparse-view inputs. Instead of directly regressing geometry from cost-volume statistics, EpiS uses coarse cost-volume features to guide the aggregation of fine-grained epipolar features sampled along corresponding epipolar lines across source views. An epipolar transformer fuses multi-view information, followed by ray-wise aggregation to produce SDF-aware features for surface estimation. To further mitigate information loss under sparse views, we introduce a geometry regularization strategy that leverages a pretrained monocular depth model through scale-invariant global and local constraints. Extensive experiments on DTU and BlendedMVS demonstrate that EpiS significantly outperforms state-of-the-art generalizable surface reconstruction methods under sparse-view settings, while maintaining strong generalization without per-scene optimization.

arXiv

MSTAR: Multi-Scale Backbone Architecture Search for Timeseries Classification

Tue M. Cao, Nhat H. Tran, Hieu H. Pham, Hung T. Nguyen, Le P. Nguyen

Most of the previous approaches to Time Series Classification (TSC) highlight the significance of receptive fields and frequencies while overlooking the time resolution. Hence, unavoidably suffered from scalability issues as they integrated an extensive range of receptive fields into classification models. Other methods, while having a better adaptation for large datasets, require manual design and yet not being able to reach the optimal architecture due to the uniqueness of each dataset. We overcome these challenges by proposing a novel multi-scale search space and a framework for Neural architecture search (NAS), which addresses both the problem of frequency and time resolution, discovering the suitable scale for a specific dataset. We further show that our model can serve as a backbone to employ a powerful Transformer module with both untrained and pre-trained weights. Our search space reaches the state-of-the-art performance on four datasets on four different domains while introducing more than ten highly fine-tuned models for each data.

arXiv

Deep Frequency-Aware Functional Maps for Robust Shape Matching

Feifan Luo, Qinsong Li, Ling Hu, Haibo Wang, Xinru Liu, Shengjun Liu, Hongyang Chen

Deep functional map frameworks are widely employed for 3D shape matching. However, most existing deep functional map methods cannot adaptively capture important frequency information for functional map estimation in specific matching scenarios, i.e., lacking \textit{frequency awareness}, resulting in poor performance when dealing with large deformable shape matching. To this end, we propose a novel unsupervised learning-based framework called Deep Frequency-Aware Functional Maps, which can gracefully cope with various shape matching scenarios. We first introduce a general constraint called Spectral Filter Operator Preservation to compute desirable functional maps, where the spectral filter operator encodes informative frequency information and can promote frequency awareness for deep functional map frameworks by learning a set of filter functions. Then, we directly utilize the proposed constraint as a loss function to supervise functional maps, pointwise maps, and filter functions simultaneously, where the filter functions are derived from the orthonormal Jacobi basis, and the coefficients of the basis are learnable parameters. Finally, we develop an effective refinement strategy to improve the final pointwise map, which incorporates our constraint and learned filter functions, leading to more robust and accurate correspondences during the inference process. Extensive experimental results on various datasets demonstrate that our approach outperforms the existing state-of-the-art methods, especially in challenging settings like datasets with non-isometric deformation and inconsistent topology.

arXiv

Modeling Long-Term Memory and Temporal Attention Shifts for Video Salient Object Ranking with a New Benchmark

Jiaying Lin, Rui Song, Huankang Guan, Shuanglin Li, Shuquan Ye, Rynson W. H. Lau

Salient Object Ranking (SOR) aims to estimate the relative saliency order among multiple salient objects. While SOR has been extensively studied in static images, Video Salient Object Ranking (VSOR) remains largely underexplored due to the lack of effective temporal saliency modeling. In particular, existing VSOR methods rely on short input frame clips, which limits their ability to capture long-term saliency evolution and identify dynamic attention shifts. To address these challenges, we propose LoTAS, a long-term memory framework for VSOR that jointly models historical saliency states and temporal attention transitions. To model historical saliency, we propose a Temporal Context Decoder (TCD) and a Rank-aware Saliency State Encoder (RSSE). The TCD retrieves historical saliency states from memory queries to provide references to previously salient instances and long-range temporal context, while the RSSE encodes current predictions into rank-aware state embeddings and updates the memory for future frames, allowing reliable ranking cues to accumulate across long video sequences. To capture temporal attention transitions, we introduce explicit inter-frame rank-transition supervision and jointly learn a binary transition predictor as an auxiliary task alongside ordinal ranking. In addition, to address the limited video types and scene diversity in the existing VSOR dataset, we propose a challenging dataset that covers diverse video types and scenes with 124 videos and 16,610 frames. Experimental results demonstrate that our method outperforms state-of-the-art VSOR methods. We will make the code and our proposed dataset available.

arXiv

OneDayAgent: Towards a Long-Horizon Harness for Autonomous Agents

Jingsheng Zheng, Xinyuan Fang, Jintian Zhang, Zhengke Gui, Huajun Chen, Ningyu Zhang

LLM agents are increasingly applied to open-ended everyday requests that span work, study, and life. These tasks are long-horizon, cross-environment, and multimodal, forcing the agent to preserve goals and constraints across many steps while navigating heterogeneous tools and attachments. While prior work has addressed individual failure modes such as goals drift, states loss, and context overflow, whether a single harness can manage them jointly and remain effective across backends has received less study. We present OneDayAgent, a long-horizon harness for autonomous agents. OneDayAgent turns an open-ended request into a managed execution process that decomposes tasks into bounded subtasks, maintains execution memory under context pressure, and verifies and repairs the final deliverable. We evaluate OneDayAgent on AgentIF-OneDay across 104 tasks. With the GLM-5.2 backend, OneDayAgent sets a new state of the art with an overall score of 0.821. The same harness runs across five backend LLMs from three model families, indicating the harness generalizes across backends without tuning, even as different models induce distinct execution styles under the same workflow.

arXiv

SafeCommit: Certifying When Memory-Grounded Agents May Safely Act

Mayur Akewar, Ravi Ranjan

Long-horizon agents increasingly use persistent memory and tools to take actions with external side effects. A central failure mode is premature commitment: an agent acts before resolving whether its memory grounding is stale, conflicting, incomplete, or corrupted. We formalize this problem as safe commitment under memory uncertainty and introduce SafeCommit, a risk controlled layer between agent reasoning and external execution. The layer constructs a calibrated set of plausible latent worlds from memory, observations, tool outputs, provenance, and policy constraints. It permits a side effectful action only when a conformal action certificate shows that the action is safe in every retained world. Otherwise, it selects a low-side-effect probe that targets the worlds blocking certification, or returns a conservative fallback. Under calibrated world coverage, the probability of an unsafe certified commit is at most the target level α; with imperfect world proposal, the bound separates calibration and representation error. A dependency-free controlled simulator illustrates the safety-utility tradeoff and reproduces all reported results with one command. The goal is to offer a concrete approach for deciding not only what an agent should do, but when the available evidence is sufficient to safely do it.

arXiv

Eliciting Intrinsic Hallucinations in LLMs via Semantically Equivalent Adversarial Attacks

Atri Vivek Sharma, Brian Formento, Alessio Lomuscio

Large language models (LLMs) are often used in conjunction with external knowledge sources to improve their factual accuracy and decrease hallucinations, through methods such as Retrieval-Augmented Generation (RAG). However, these systems remain susceptible to intrinsic hallucinations, where the model generates unfaithful or fabricated information that is not supported by the retrieved evidence. We propose a novel framework to assess model robustness against this phenomenon by stress-testing using natural, semantically equivalent variations of a user query found via adversarial optimization methods. We apply our framework, which enforces strict semantic equivalence constraints and an intrinsic hallucination objective, to a range of adversarial attack techniques across white-box, gray-box, and black-box adversarial settings. Evaluating these attacks on 5 open-source and 5 closed-source generator models across 3 datasets, we demonstrate that even state-of-the-art models are highly susceptible to meaning-preserving perturbations, which significantly degrade contextual faithfulness (by up to 50% for GPT-5-mini). Our findings indicate that faithful use of in-context evidence remains fragile even in state-of-the-art LLMs, motivating architectures and training objectives that enforce robust grounding independent of surface query form. Code is available at: https://github.com/atriviveksharma/intrinsic_hall

arXiv

LLM-based Vulnerability Discovery in Business Process Documentation

Ben Falchuk, Himanshu Garg, Euthimios Panagos, Sioan Zohar

Just like software and hardware, business processes are susceptible to vulnerabilities that can lead to product quality issues, delays, and increased costs. Business process vulnerabilities can arise from a variety of sources, including conflicting requirements, ambiguous documentation, invalid measurement spec-ifications, omission of quality checks, or implementations that differ from speci-fications. MIRABELLE is a system that identifies and characterizes business logic (BL) vulnerabilities from available business process representations, in-cluding ISO 9000/9001 documentation, user guides, work instructions, and pro-cess execution logs. MIRABELLE leverages recent advances in AI/ML to pro-cess available business process documentation and generate attributed graph rep-resentations of the business logic that can be processed using both graph and for-mal logic approaches for identifying potential vulnerabilities. However, extract-ing the business logic (e.g., operation execution sequences, decisions, input/out-put resources) from mostly natural language artifacts is challenging due to the required domain expertise, inherent process complexity, and the sometimes very large volumes of information. This paper focuses on our experimentation with Large Language Models (LLMs) and their role within MIRABELLE. We report on the performance of several LLMs across vital stages of vulnerability detection, from grammatical and technical error-flagging in short phrasings, to complete process structure recovery and extraction.

arXiv

The Fairness Collapse Phenomenon: Bias Amplification in Language Models Trained on Synthetic Data

Irina Proskurina, Antoine Gourru, Julien Velcin

Generative models trained on artificially generated data have been shown to exhibit model collapse, resulting in significant performance degradation. As synthetic content increasingly contaminates the training corpora of language models, this raises critical concerns about the use of open data in continued pretraining. Although previous work has demonstrated model collapse in language models, it remains unclear whether exposure to synthetic data amplifies or attenuates the social biases already present in pretrained models. Because language models are known to reproduce and amplify demographic stereotypes, recursive training on self-generated data may create a self-reinforcing feedback loop in which biased associations become progressively stronger across generations. We call this hypothesized phenomenon fairness collapse. In this work, we construct controlled training regimes in which models are repeatedly trained on synthetic data using the Bias in Bios dataset. Across experiments, we observe a consistent and concerning pattern: fairness degradation emerges before substantial degradation is reflected by standard language-modeling metrics. This result highlights a critical risk associated with synthetic data contamination in language model training: bias can increase silently before strong indicators of model collapse become apparent.

Keep browsing

More papers from top AI conferences

30+ top conferences, by conference, year, and topic.