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

CIRCLED: A Multi-turn CIR Dataset with Consistent Dialogues across Domains

Tomohisa Takeda, Yu-Chieh Lin, Yuji Nozawa, Youyang Ng, Osamu Torii, Yusuke Matsui

Existing Multi-Turn Composed Image Retrieval (MTCIR) datasets lack dialogue-historyconsistency and are restricted to the fashion domain. To address these limitations, we construct CIRCLED by extending FashionIQ, CIRR, and CIRCO. In CIRCLED, the query ateach turn progressively approaches the target image. Data are generated via a CIReVLbased retrieval pipeline and curated with multiple filters on retrieval success, turn length, consistency, and information redundancy to ensure quality. In total, we collect 22,608 multiturn sessions across nine subsets, substantially exceeding Multi-turn FashionIQ (11,505 sessions) in both scale and generality. We further apply multiple baseline methods and quantitatively assess retrieval accuracy on CIRCLED. Our work provides a practical, highquality benchmark to facilitate future research on multi-turn CIR. The dataset is publicly available at https://huggingface.co/datasets/tk1441/CIRCLED, and the code at https://github.com/mti-lab/circled.

arXiv

Component-Aware Structure-Preserving Style Transfer for Satellite Visual Sim2Real Data Construction

Zongwu Xie, Yonglong zhang, Yifan Yang, Yang Liu, Baoshi Cao, Guanghu Xie

For camera-based satellite visual sensing, Sim2Real data construction requires images that approach real-domain sensor appearance while retaining the annotations inherited from simulation. Real sensor images of satellite targets with reliable pose labels and component-level masks are difficult to acquire at scale, whereas synthetic rendering provides exact geometric annotations but suffers from a visible appearance gap. This paper presents a component-aware structure-preserving style transfer framework for satellite visual synthetic-to-real data construction. The method builds weakly paired real--synthetic samples from calibrated real acquisition, ArUco-based camera-pose measurement, CAD rendering, and component masks. It then extracts part-wise real-domain style codes from unlabeled real images and injects them into corresponding synthetic satellite regions through mask-aligned modulation. To keep the generated images usable for downstream sensor-data supervision, adversarial training is combined with local contrastive consistency, self-regularization, and edge-preserving constraints. Experiments are conducted on 5,000 rendered satellite images and 100 real images captured in a calibrated setup. The real images provide target-domain appearance references and final evaluation images, while the downstream GDRNet pose estimator is trained only on synthetic or translated synthetic images. Compared with representative image-translation baselines, the proposed method achieves the lowest image distribution discrepancy, with an FID of 54.32 and a KID of 0.048. When the translated data are used to train GDRNet in this target-domain adaptation setting, the ADD pass rate improves to 0.260 and the AUC improves to 0.611. These results indicate that component-level appearance transfer can improve annotation-preserving satellite visual Sim2Real data generation in the considered calibrated setup.

arXiv

Counterfactual Stress Testing for Image Classification Models

Moritz Stammel, Fabio De Sousa Ribeiro, Raghav Mehta, Mélanie Roschewitz, Ben Glocker

Deep learning models in medical imaging often fail when deployed in new clinical environments due to distribution shifts in demographics, scanner hardware, or acquisition protocols. A central challenge is underspecification, where models with similar validation performance exhibit divergent real-world failure modes. Although stress testing has emerged as a tool to assess this, current methods typically rely on simple, uninformed perturbations (e.g., brightness or contrast changes), which fail to capture clinically realistic variation and can overestimate robustness. In this work, we introduce a counterfactual stress testing framework based on causal generative models that create realistic "what if" images by intervening on attributes such as scanner type and recorded sex while largely preserving anatomical identity, enabling controlled and semantically meaningful evaluation under targeted distribution shifts. Across two imaging modalities (chest X-ray and mammography), three model architectures, and multiple shift scenarios, we show that counterfactual stress tests provide a substantially more accurate proxy for real out-of-distribution performance than classical perturbations, capturing the direction and relative magnitude of performance changes and showing stronger overall rank agreement. These results suggest that causal generative models can provide informative synthetic stress tests for assessing robustness under targeted distribution shifts prior to deployment.

arXiv

DF3DV-1K: A Large-Scale Dataset and Benchmark for Distractor-Free Novel View Synthesis

Cheng-You Lu, Yi-Shan Hung, Wei-Ling Chi, Hao-Ping Wang, Charlie Li-Ting Tsai, Yu-Cheng Chang, Yu-Lun Liu, Thomas Do, Chin-Teng Lin

Advances in radiance fields have enabled photorealistic novel view synthesis. In several domains, large-scale real-world datasets have been developed to support comprehensive benchmarking and to facilitate progress beyond scene-specific reconstruction. However, for distractor-free radiance fields, a large-scale dataset with clean and cluttered images per scene remains lacking, limiting the development. To address this gap, we introduce DF3DV-1K, a large-scale real-world dataset comprising 1,048 scenes, each providing clean and cluttered image sets for benchmarking. In total, the dataset contains 89,924 images captured using consumer cameras to mimic casual capture, spanning 128 distractor types and 161 scene themes across indoor and outdoor environments. A curated subset of 41 scenes, DF3DV-41, is systematically designed to evaluate the robustness of distractor-free radiance field methods under challenging scenarios. Using DF3DV-1K, we benchmark nine recent distractor-free radiance field methods and 3D Gaussian Splatting, identifying the most robust methods and the most challenging scenarios. Beyond benchmarking, we demonstrate an application of DF3DV-1K by fine-tuning a diffusion-based 2D enhancer to improve radiance field methods, achieving average improvements of 0.96 dB PSNR and 0.057 LPIPS on the held-out set (e.g., DF3DV-41) and the On-the-go dataset. We hope DF3DV-1K facilitates the development of distractor-free vision and promotes progress beyond scene-specific approaches. The dataset and leaderboard are available at https://johnnylu305.github.io/df3dv1k_web/.

arXiv

ProtoFlow: Mitigating Forgetting in Class-Incremental Remote Sensing Segmentation via Low-Curvature Prototype Flow

Jiekai Wu, Rong Fu, Chuangqi Li, Zijian Zhang, Guangxin Wu, Hao Zhang, Shiyin Lin, Yang Li, Dongxu Zhang, Amir H. Gandomi, Simon Fong, Pengbin Feng

Remote sensing segmentation in real deployment is inherently continual: new semantic categories emerge, and acquisition conditions shift across seasons, cities, and sensors. Despite recent progress, many incremental approaches still treat training steps as isolated updates, which leaves representation drift and forgetting insufficiently controlled. We present ProtoFlow, a time-aware prototype dynamics framework that models class prototypes as trajectories and learns their evolution with an explicit temporal vector field. By jointly enforcing low-curvature motion and inter-class separation, ProtoFlow stabilizes prototype geometry throughout incremental learning. Experiments on standard class- and domain-incremental remote sensing benchmarks show consistent gains over strong baselines, including up to 1.5-2.0 points improvement in mIoUall, together with reduced forgetting. These results suggest that explicitly modeling temporal prototype evolution is a practical and interpretable strategy for robust continual remote sensing segmentation. Open-source code:https://github.com/dudududke/protoflow.

arXiv

RecGen3D: Reconstruction-Guided 3D Generation in a Shared Canonical Space

Zhisheng Huang, Jiahao Chen, Cheng Lin, Chenyu Hu, Hanzhuo Huang, Zhengming Yu, Mengfei Li, Yuheng Liu, Zekai Gu, Zibo Zhao, Yuan Liu, Xin Li, Wenping Wang

Sparse-view 3D modeling represents a fundamental tension between reconstruction fidelity and generative plausibility. While feed-forward reconstruction excels in efficiency and input alignment, it often lacks the global priors needed for structural completeness. Conversely, diffusion-based generation provides rich geometric details but struggles with multi-view consistency. We present RecGen3D, a framework that combines these two paradigms into a cooperative system. To overcome inherent conflicts in coordinate spaces, 3D representations, and training objectives, we align both models within a shared canonical space. We employ decoupled cooperative learning, which maintains stable training while enabling seamless collaboration during inference. Specifically, the reconstruction module is adapted to provide canonical geometric anchors, while the diffusion generator leverages latent-augmented conditioning to refine and complete the geometric structure. Experimental results demonstrate that RecGen3D achieves superior fidelity and robustness, outperforming existing methods in creating complete and consistent 3D models from sparse observations.

arXiv

InverFill: One-Step Inversion for Enhanced Few-Step Diffusion Inpainting

Duc Vu, Kien Nguyen, Trong-Tung Nguyen, Ngan Nguyen, Phong Nguyen, Khoi Nguyen, Cuong Pham, Anh Tran

Recent diffusion-based models achieve photorealism in image inpainting but require many sampling steps, limiting practical use. Few-step text-to-image models offer faster generation, but naively applying them to inpainting yields poor harmonization and artifacts between the background and inpainted region. We trace this cause to random Gaussian noise initialization, which under low function evaluations causes semantic misalignment and reduced fidelity. To overcome this, we propose InverFill, a one-step inversion method tailored for inpainting that injects semantic information from the input masked image into the initial noise, enabling high-fidelity few-step inpainting. Instead of training inpainting models, InverFill leverages few-step text-to-image models in a blended sampling pipeline with semantically aligned noise as input, significantly improving vanilla blended sampling and even matching specialized inpainting models at low NFEs. Moreover, InverFill does not require real-image supervision and only adds minimal inference overhead. Extensive experiments show that InverFill consistently boosts baseline few-step models, improving image quality and text coherence without costly retraining or heavy iterative optimization.

arXiv

Radiometric fingerprinting of object surfaces using mobile laser scanning and semantic 3D road space models

Benedikt Schwab, Thomas H. Kolbe

Although semantic 3D city models are internationally available and becoming increasingly detailed, the incorporation of material information remains largely untapped. However, a structured representation of materials and their physical properties could substantially broaden the application spectrum and analytical capabilities for urban digital twins. At the same time, the growing number of repeated mobile laser scans of cities and their street spaces yields a wealth of observations influenced by the material characteristics of the corresponding surfaces. To leverage this information, we propose radiometric fingerprints of object surfaces by grouping LiDAR observations reflected from the same semantic object under varying distances, incident angles, environmental conditions, sensors, and scanning campaigns. Our study demonstrates how 312.4 million individual beams acquired across four campaigns using five LiDAR sensors on the Audi Autonomous Driving Dataset (A2D2) vehicle can be automatically associated with 6368 individual objects of the semantic 3D city model. The model comprises a comprehensive and semantic representation of four inner-city streets at Level of Detail (LOD) 3 with centimeter-level accuracy. It is based on the CityGML 3.0 standard and enables fine-grained sub-differentiation of objects. The extracted radiometric fingerprints for object surfaces reveal recurring intra-class patterns that indicate class-dominant materials. The semantic model, the method implementations, and the developed geodatabase solution 3DSensorDB are released under: https://github.com/tum-gis/sensordb

arXiv

PhotoBench: Beyond Visual Matching Towards Personalized Intent-Driven Photo Retrieval

Tianyi Xu, Rong Shan, Junjie Wu, Jiadeng Huang, Teng Wang, Jiachen Zhu, Wenteng Chen, Minxin Tu, Quantao Dou, Zhaoxiang Wang, Changwang Zhang, Weinan Zhang, Jun Wang, Jianghao Lin

Personal photo albums are not merely collections of static images but living, ecological archives defined by temporal continuity, social entanglement, and rich metadata, which makes the personalized photo retrieval non-trivial. However, existing retrieval benchmarks rely heavily on context-isolated web snapshots, failing to capture the multi-source reasoning required to resolve authentic, intent-driven user queries. To bridge this gap, we introduce PhotoBench, the first benchmark constructed from authentic, personal albums. It is designed to shift the paradigm from visual matching to personalized multi-source intent-driven reasoning. Based on a rigorous multi-source profiling framework, which integrates visual semantics, spatial-temporal metadata, social identity, and temporal events for each image, we synthesize complex intent-driven queries rooted in users' life trajectories. Extensive evaluation on PhotoBench exposes two critical limitations: the modality gap, where unified embedding models collapse on non-visual constraints, and the source fusion paradox, where agentic systems perform poor tool orchestration. These findings indicate that the next frontier in personal multimodal retrieval lies beyond unified embeddings, necessitating robust agentic reasoning systems capable of precise constraint satisfaction and multi-source fusion. Our PhotoBench is available.

arXiv

Driving with DINO: Vision Foundation Features as a Unified Bridge for Sim-to-Real Generation in Autonomous Driving

Xuyang Chen, Conglang Zhang, Chuanheng Fu, Zihao Yang, Kaixuan Zhou, Yizhi Zhang, Yanfeng Zhang, Mingwei Sun, Zhen Dong, Xiaoxiao Long, Zengmao Wang, Liqiu Meng

Driven by the emergence of Controllable Video Diffusion, existing Sim2Real methods for autonomous driving video generation typically rely on explicit intermediate representations to bridge the domain gap. However, these modalities face a fundamental Consistency-Realism Dilemma. Low-level signals (e.g., edges, blurred images) ensure precise control but compromise realism by "baking in" synthetic artifacts, whereas high-level priors (e.g., depth, semantics, HDMaps) facilitate photorealism but lack the structural detail required for consistent guidance. In this work, we present Driving with DINO (DwD), a novel framework that leverages Vision Foundation Module (VFM) features as a unified bridge between the simulation and real-world domains. We first identify that these features encode a spectrum of information, from high-level semantics to fine-grained structure. To effectively utilize this, we employ Principal Subspace Projection to discard the high-frequency elements responsible for "texture baking," while concurrently introducing Random Channel Tail Drop to mitigate the structural loss inherent in rigid dimensionality reduction, thereby reconciling realism with control consistency. Furthermore, to fully leverage DINOv3's high-resolution capabilities for enhancing control precision, we introduce a learnable Spatial Alignment Module that adapts these high-resolution features to the diffusion backbone. Finally, we propose a Causal Temporal Aggregator employing causal convolutions to explicitly preserve historical motion context when integrating frame-wise DINO features, which effectively mitigates motion blur and guarantees temporal stability. Project page: https://albertchen98.github.io/DwD-project/

arXiv

Deep Models, Shallow Alignment: Uncovering the Granularity Mismatch in Neural Decoding

Yang Du, Siyuan Dai, Yonghao Song, Paul M. Thompson, Haoteng Tang, Liang Zhan

Neural visual decoding is a central problem in brain-computer interface research, aiming to reconstruct human visual perception and to elucidate the structure of neural representations. Recent contrastive neural visual decoding methods commonly align neural signals with the final embeddings of pretrained vision encoders. However, such representations are optimized for high-level semantic invariance, whereas EEG/MEG signals contain information spanning multiple levels of visual abstraction, potentially creating a representational granularity mismatch. Motivated by prior evidence that brain representations correspond to multiple levels of the DNN hierarchy, we propose Shallow Alignment, a granularity-calibration framework that systematically explores intermediate visual representations as alignment targets for neural decoding. Extensive experiments across multiple benchmarks demonstrate that Shallow Alignment significantly outperforms standard final-layer alignment, with performance gains ranging from 22% to 58% across diverse vision backbones. Notably, our approach reveals a positive scaling trend in neural visual decoding, enabling decoding performance to improve consistently with the capacity of pre-trained vision backbones. We further conduct systematic empirical analyses to shed light on the mechanisms underlying the observed performance gains. Code is available at https://github.com/yangdu-neuroai/shallow-alignment.

arXiv

CFM: Language-aligned Concept Foundation Model for Vision

Kai Wittenmayer, Sukrut Rao, Amin Parchami-Araghi, Bernt Schiele, Jonas Fischer

Language-aligned vision foundation models perform strongly across diverse downstream tasks. Yet, their learned representations remain opaque, making interpreting their decision-making difficult. Recent work decompose these representations into human-interpretable concepts, but provide poor spatial grounding and are limited to image classification tasks. In this work, we propose CFM, a language-aligned concept foundation model for vision that provides fine-grained concepts, which are human-interpretable and spatially grounded in the input image. When paired with a foundation model with strong semantic representations, we get explanations for any of its downstream tasks. Examining local co-occurrence dependencies of concepts allows us to define concept relationships through which we improve concept naming and obtain richer explanations. On benchmark data, we show that CFM provides performance on classification, segmentation, and captioning that is competitive with opaque foundation models while providing fine-grained, high quality concept-based explanations. Code at https://github.com/kawi19/CFM. Interactive visualizations at https://concept-foundation-model.mpi-inf.mpg.de.

arXiv

Fine-tuning an ECG Foundation Model to Predict Coronary CT Angiography Outcomes

Yujie Xiao, Qinghao Zhao, Gongzheng Tang, Hao Zhang, Zhuoran Kan, Deyun Zhang, Jun Li, Guangkun Nie, Xiaocheng Fang, Haoyu Wang, Shun Huang, Tong Liu, Jian Liu, Kangyin Chen, Shenda Hong

Coronary artery disease (CAD) remains a major global public health burden, yet scalable pre-imaging risk stratification tools are limited. In this multicenter study, we developed and validated an artificial intelligence-enabled electrocardiography (AI-ECG) model using coronary computed tomographic angiography (CCTA) as the anatomical reference to predict vessel-specific hemodynamically significant stenosis ($\geq 70\%$ for RCA, LAD, LCX; $\geq 50\%$ for LM). The model was evaluated in internal and external cohorts, clinically normal ECGs, and prespecified demographic and clinical subgroups. It showed discrimination across vessels in internal validation and consistent external and normal ECG performance. Predicted probabilities increased with CCTA-defined stenosis severity and were converted into vessel-specific low-, intermediate-, and high-risk strata. Calibration and decision curve analyses supported its clinical utility. Integration with guideline-based pre-test probability improved risk reclassification, enhanced rule-out performance, and reduced the gray-zone proportion. In longitudinal follow-up, model-defined risk groups showed clear separation in major adverse cardiovascular events. Waveform- and attribution-based analyses identified structured ECG differences and physiologically meaningful signal regions linked to high-risk predictions. These results support AI-ECG as a feasible tool for pre-imaging risk stratification and clinical triage, warranting prospective validation in broader clinical settings.

arXiv

What Color Is the Text? A Benchmark for Hallucination Induced by Image-Embedded Prompt

Jinkun Zhao, Lei Huang, Haixin Ge, Wenjun Wu

We introduce Embedded Stroop, a controlled diagnostic paradigm for measuring image-embedded prompt interference in Multimodal Large Language Models (MLLMs), where the query is rendered directly inside the visual input. Using the What-Color-Is-the-Text (WCIT) benchmark, which covers 59 fine-grained colors under Standard, Flipped, and Masked variants, we evaluate 16 proprietary and open-source models. To distinguish semantic capture from general color-naming failure, we decompose model responses into Accuracy, Stroop Hallucination Rate (SHR; answering the embedded word rather than the true text color), and Other Error Rate, and validate the effect with permutation tests, with 56 of 64 conditions remaining significant after FDR correction. Although exact color accuracy is low (6.3%) under the 59-color vocabulary, mapping predictions to 11 basic color families shows that models retain coarse color perception (38.4%) while still exhibiting a substantial SHR (21.6%). A conditional analysis restricted to colors correctly named in the Standard setting further confirms the effect, with pooled conditional SHR exceeding 50\%. Masking or flipping the embedded text reduces Stroop hallucinations, suggesting that semantic legibility can dominate visual color perception in MLLMs.

arXiv

Optimizing Multi-Modality Trackers via Significance-Regularized Tuning

Zhiwen Chen, Jinjian Wu, Zhiyu Zhu, Yifan Zhang, Guangming Shi, Junhui Hou

This paper tackles the critical challenge of optimizing multi-modality trackers by effectively adapting pre-trained models for RGB data. Existing fine-tuning paradigms oscillate between excessive flexibility and over-restriction, both leading to suboptimal plasticity-stability trade-offs. To mitigate this dilemma, we propose a novel significance-regularized fine-tuning framework, which delicately refines the learning process by incorporating intrinsic parameter significance. Through a comprehensive investigation of the transition from pre-trained to multi-modality contexts, we identify that parameters crucial to preserving foundational patterns and managing cross-domain shifts are the primary drivers of this issue. Specifically, we first probe the tangent space of pre-trained weights to measure and orient prior significance, dedicated to preserving generalization. Subsequently, we characterize transfer significance during the fine-tuning phase, emphasizing adaptability and stability. By incorporating these parameter significance terms as unified regularization, our method markedly enhances transferability across modalities. Extensive experiments showcase the superior performance of our method, surpassing current state-of-the-art techniques across various multi-modal tracking benchmarks. The source code and models are publicly available at https://github.com/zhiwen-xdu/SRTrack.

arXiv

An Automated Pipeline for Few-Shot Bird Call Classification: A Case Study with the Tooth-Billed Pigeon

Abhishek Jana, Moeumu Uili, James Atherton, Mark O'Brien, Joe Wood, Leandra Brickson

This paper presents a largely automated one-shot bird call classification pipeline, incorporating targeted manual quality control steps, designed for rare species absent from large publicly available classifiers like BirdNET and Perch. While these models excel at detecting common birds with abundant training data, they lack options for species with only 1-3 known recordings, a critical limitation for conservationists monitoring the last remaining individuals of endangered birds. To address this, we leverage the embedding space of large bird classification networks and develop a classifier using cosine similarity, combined with filtering and denoising preprocessing techniques, to optimize detection with minimal training data. We evaluate various embedding spaces using clustering metrics and validate our approach in both a simulated scenario with Xeno-Canto recordings and a real-world test on the critically endangered tooth-billed pigeon (Didunculus strigirostris), which has no existing classifiers and only three confirmed recordings. The final model achieved 1.0 recall and 0.95 accuracy in detecting tooth-billed pigeon calls, making it practical for use in the field. This open-source system provides a practical tool for conservationists seeking to detect and monitor rare species on the brink of extinction.

arXiv

Beyond Isolated Heads: Multi-Overlapped-Head Self-Attention for Vision Transformers

Tianxiao Zhang, Bo Luo, Guanghui Wang

Multi-Head Self-Attention (MHSA) is the cornerstone of Vision Transformers, allowing models to capture diverse feature representations by projecting tokens into independent subspaces. However, the standard MHSA strictly isolates these heads, preventing any information exchange during the attention computation itself. In this paper, we propose Multi-Overlapped-Head Self-Attention (MOHSA), a novel mechanism that replaces the hard division of attention heads with a soft, overlapping division. By allowing queries, keys, and values to partially overlap with those of adjacent heads, MOHSA fosters rich inter-head communication directly within the attention mechanism. We extensively investigate various overlap dimension scheduling strategies to identify optimal configurations. Comprehensive experiments in several Transformer architectures demonstrate that MOHSA outperforms standard MHSA on the CIFAR-10, CIFAR-100, Tiny-ImageNet, and ImageNet-1k datasets, offering a significant performance boost with negligible computational overhead. Code link: https://github.com/ZTX-100/MOHSA.

arXiv

Vision Foundation Model Driven Foreground-Aware Pseudo-LiDAR Generation for Monocular 3D Object Detection

Bonan Ding, Jin Xie, Jing Nie, Jiale Cao, Yanwei Pang

Pseudo-LiDAR has become a promising paradigm for monocular 3D object detection by transforming monocular images into point cloud representations that can be processed by LiDAR-based 3D object detectors. Recent vision foundation models provide powerful geometric and semantic priors, creating new opportunities for improving the quality of pseudo-LiDAR generation. However, effectively exploiting these priors to produce reliable pseudo-LiDAR remains challenging due to inaccurate depth estimation, insufficient foreground awareness, and redundant background points.In this paper, we propose VFMM3D, a vision foundation model-driven framework for foreground-aware pseudo-LiDAR generation. VFMM3D leverages the depth prior provided by the Depth Anything Model (DAM) and the foreground prior provided by the Segment Anything Model (SAM). A foreground-aware pseudo-LiDAR painting operation is introduced to associate projected 3D points with object-level foreground information, thereby strengthening object structures and suppressing irrelevant background regions. Furthermore, a sparsification strategy is developed to remove redundant points, reduce computational overhead, and improve the compatibility of the generated point clouds with LiDAR-based 3D object detectors. Comprehensive experiments are conducted on two challenging 3D object detection datasets, KITTI and Waymo. Our VFMM3D establishes a new state-of-the-art performance on both datasets. Additionally, experimental results demonstrate the generality of VFMM3D, showcasing its seamless integration into various LiDAR-based 3D object detectors.

arXiv

Measuring Activation Control in Large Language Models

Marek Mateusz Kowalski, Joshua Fonseca Rivera, Uzay Macar, David Demitri Africa

Safe deployment of increasingly capable models will likely come to rely on latent-space monitoring as a complement to behavioral evaluations, especially when evaluation-aware models exhibit scheming or deception. However, if models can also control their own activations, deception could extend into the latent space itself. With this in mind, we introduce the Activation Controllability Benchmark to quantify the extent to which models can modulate their residual stream via natural-language instruction. Across model families and capability levels, we find that most LLMs can control the direction and magnitude of their residual stream activations with some degree of temporal resolution, though performance varies considerably across models. In simple tasks, this level of control can evade activation-based monitoring methods (including linear probes, natural language autoencoders, activation oracles, and the Jacobian lens), albeit imperfectly. These results suggest that control over the activation space itself could become a confound for monitoring as introspective capabilities increase; therefore, we recommend that frontier labs and evaluators track activation controllability in future models.

arXiv

Mitigating Database Leakage in RAG Systems with Keyword-Grounded Fact Substitution

Ziliang Zhang, Yubo Zhu, Wei Tong, Jingyu Hua, Zijian Wang, Yuan Zhang, Sheng Zhong

Retrieval-Augmented Generation (RAG) has emerged as a powerful paradigm for combining large language models (LLMs) with external knowledge sources. However, RAG systems remain vulnerable to prompt injection attacks, which may mislead the retriever or generator to expose sensitive database contents. To address this issue, we propose KFS-RAG, a defense that mitigates information leakage by reformulating the retrieved context. Specifically, our method first identifies a small set of influential keywords from the retrieved context via an attention rollout plus a causal perturbation mechanism. These keywords are then used to guide an auxiliary LLM to generate a compact set of keyword-grounded facts from the retrieved passages. Finally, the original context is substituted with these curated facts, ensuring that the generator operates on sanitized evidence rather than the raw retrieved text. Experimental evaluations demonstrate that KFS-RAG significantly reduces the risk of database leakage under injection attacks while maintaining response accuracy and relevance. This work highlights a practical pathway toward building secure and trustworthy RAG systems.

Keep browsing

More papers from top AI conferences

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