Poster Session 2
(11/6, Fri, 16:45~18:00)
(11/6, Fri, 16:45~18:00)
[Part 2] 11/6(Fri.) 16:45~13:30
[2-1] Disambiguating 2D-3D Correspondences in Gaussian Splatting-based Feature Fields for Visual Localization (NeurIPS 2026)
이미소(석박통합과정)
While Gaussian Splatting-based Feature Fields (GSFFs) have shown promise for visual localization, this paper highlights that photometrically optimized GSFFs are inherently not suitable for 2D-3D matching. The volumetric extent of each Gaussian induces many-to-one pixel-to-point mappings that destabilize PnP-based pose estimation, while photometric optimization gives rise to superfluous Gaussians devoid of multi-view consistency. To address these issues, we propose SplitGS-Loc, a GSFFs construction framework for localization, which disambiguates 2D-3D correspondences by exploiting Gaussian attributes. Our key design, Mixture-of-Gaussians-based splitting, decomposes each Gaussian into smaller Gaussians, replacing ambiguous many-to-one with precise one-to-one correspondences. In parallel, we exploit composition weights from GS rasterization to select Gaussians that significantly and consistently contribute across multiple views and aggregate discriminative features through strong pixel-Gaussian associations. The resulting compact yet discriminative feature fields enable stable PnP convergence, achieving competitive performance on localization benchmarks. Extensive experiments validate that SplitGS-Loc extends the utility of photometric GSFFs to accurate and efficient localization by exploiting Gaussian attributes, without per-scene training for feature fields or iterative pose refinement.
[2-2] Temporally Consistent Long-Term Memory for 3D Single Object Tracking (CVPR 2026 Findings)
유재준(박사과정)
3D Single Object Tracking (3D-SOT) aims to localize a target object across a sequence of LiDAR point clouds, given its 3D bounding box in the first frame. Recent methods have adopted a memory-based approach to utilize previously observed features of the target object, but remain limited to only a few recent frames. This work reveals that their temporal capacity is fundamentally constrained to short-term context due to severe temporal feature inconsistency and excessive memory overhead. To this end, we propose a robust long-term 3D-SOT framework, ChronoTrack, which preserves the temporal feature consistency while efficiently aggregating the diverse target features via long-term memory. Based on a compact set of learnable memory tokens, ChronoTrack leverages long-term information through two complementary objectives: a temporal consistency loss and a memory cycle consistency loss. The former enforces feature alignment across frames, alleviating temporal drift and improving the reliability of proposed long-term memory. In parallel, the latter encourages each token to encode diverse and discriminative target representations observed throughout the sequence via memory-point-memory cyclic walks. As a result, ChronoTrack achieves new state-of-the-art performance on multiple 3D-SOT benchmarks, demonstrating its effectiveness in long-term target modeling with compact memory while running at real-time speed of 42 FPS on a single RTX 4090 GPU. The code is available at https://github.com/ujaejoon/ChronoTrack
[2-3] Analyzing the Training Dynamics of Image Restoration Transformer (ICLR 2026)
이민규(박사과정)
Traditional Image Restoration Transformers coupled with standard per-token LayerNorm suffers in preserving spatial information. Thus, we propose a image restoration tailored normalization method i-LN.
[2-4] PDF-GS: Progressive Distractor Filtering for Robust 3D Gaussian Splatting (CVPR 2026 Findings)
서강민(석사과정)
Recent advances in 3D Gaussian Splatting (3DGS) have enabled impressive real-time photorealistic rendering. However, conventional training pipelines inherently assume full multi-view consistency among input images, which makes them sensitive to distractors that violate this assumption and cause visual artifacts. In this work, we revisit an underexplored aspect of 3DGS: its inherent ability to suppress inconsistent signals. Building on this insight, we propose PDF-GS (Progressive Distractor Filtering for Robust 3D Gaussian Splatting), a framework that amplifies this self-filtering property through a progressive multi-phase optimization. The progressive filtering phases gradually remove distractors by exploiting discrepancy cues, while the following reconstruction phase restores fine-grained, view-consistent details from the purified Gaussian representation. Through this iterative refinement, PDF-GS achieves robust, high-fidelity, and distractor-free reconstructions, consistently outperforming baselines across diverse datasets and challenging real-world conditions. Moreover, our approach is lightweight and easily adaptable to existing 3DGS frameworks, requiring no architectural changes or additional inference overhead, leading to a new state-of-the-art performance.
[2-5] Towards Generalizable Partially Relevant Video Retrieval (NeurIPS 2026)
전우진(박사과정)
Partially Relevant Video Retrieval (PRVR) aims to retrieve an untrimmed video that contains at least one segment relevant to a text query. Despite its practical motivation, existing PRVR methods are typically trained and evaluated on a single curated benchmark, limiting insight into whether they learn generalizable fine-grained text-video alignment. To examine this issue, we introduce a multi-source-to-unseen-target protocol, in which multiple source benchmarks are merged into a single training corpus without source labels and the trained model is evaluated on a held-out target benchmark. Under this protocol, existing PRVR methods generalize poorly to unseen targets despite leveraging the general image-text alignment of CLIP. Surprisingly, they even underperform zero-shot CLIP, suggesting that PRVR training can actively erode the transferable alignment inherited from CLIP. To address this limitation, we propose Generalization-Aware Learning GAL for PRVR, a two-branch framework that preserves CLIP-derived transferability while learning PRVR-specific fine-grained text-segment matching. GAL consists of an anchored branch that maintains CLIP's transferability and an adaptive branch that learns task-oriented segment-level alignment. We train these branches through conservative residual updates, per-sample branch specialization, and CLIP-text relation distillation, enabling complementary learning without overwriting the alignment structure provided by CLIP. Under the proposed protocol, GAL achieves state-of-the-art unseen-target retrieval with competitive source performance, showing that fine-grained matching can be learned without collapsing transferable alignment.
[2-6] SynSym: A Synthetic Data Generation Framework for Psychiatric Symptom Identification (KDD 2026)
강미경(석박통합과정)
Psychiatric symptom identification on social media aims to infer fine-grained mental health symptoms from user-generated posts, allowing a detailed understanding of users’ mental states. However, the construction of large-scale symptom-level datasets remains challenging due to the resource-intensive nature of expert labeling and the lack of standardized annotation guidelines, which in turn limits the generalizability of models to identify diverse symptom expressions from user-generated text. To address these issues, we propose SynSym, a synthetic data generation framework for constructing generalizable datasets for symptom identification. Leveraging large language models (LLMs), SynSym constructs high-quality training samples by (1) expanding each symptom into sub-concepts to enhance the diversity of generated expressions, (2) producing synthetic expressions that reflect psychiatric symptoms in diverse linguistic styles, and (3) composing realistic multi-symptom expressions, informed by clinical co-occurrence patterns. We validate SynSym on three benchmark datasets covering different styles of depressive symptom expression. Experimental results demonstrate that models trained solely on the synthetic data generated by SynSym perform comparably to those trained on real data, and benefit further from additional fine-tuning with real data. These findings underscore the potential of synthetic data as an alternative resource to real-world annotations in psychiatric symptom modeling, and SynSym serves as a practical framework for generating clinically relevant and realistic symptom expressions.
[2-7] learning to refuse: refusal-aware reinforcement fine-tuning for hard-irrelevant queries in video temporal grounding (CVPR 2026)
이성준(석박통합과정)
Video Temporal Grounding (VTG) aims to localize a temporal segment in a video corresponding to a natural language query. However, existing VTG models assume that a relevant segment always exists, causing them to always predict a target segment even when the query is irrelevant to the video. While recent approaches attempt to handle irrelevant queries, they can only reject those that are entirely unrelated to the video and still fail to handle hard-irrelevant queries that are semantically similar but not actually relevant. To address this, we propose Refusal-Aware Reinforcement Fine-Tuning (RA-RFT) to effectively refuse hard-irrelevant queries in VTG. Our method is based on the Group Relative Policy Optimization (GRPO) framework and integrates four reward objectives-format, refuse-IoU, explain, and query correction-to improve both relevance discrimination and fine-grained semantic reasoning. In addition, to effectively support RA-RFT, we construct a Hard-Irrelevant VTG (HI-VTG) dataset, which includes hard-irrelevant queries and their refusal answers. We demonstrate the effectiveness of our method across various relevance-aware VTG scenarios, including hard-irrelevant VTG, simply-shuffled RA-VTG, and human-annotated RA-VTG settings. We also show that the proposed method is scalable by applying it to various LVLM-based VTG models. Our code is available athttps://github.com/JINSUBY/RA-RFT..
[2-8] InterventionBench: Can LLM Agents Adapt to Mid-task User Interventions? (EMNLP 2026 Findings)
이재훈(박사과정)
LLM agents are increasingly deployed on long- horizon tasks, where a user often revises the goal partway through rather than stating it once at the start. Adapting an ongoing trajectory to such a mid-task intervention is a basic requirement for agents that work with a user, yet exist- ing agentic benchmarks evaluate agents under a goal that stays fixed for the whole task. To address this gap, we introduce InterventionBench, a benchmark that measures how well an agent adapts to a mid-task intervention. Built on AppWorld and WebArena, InterventionBench inserts a user message that revises the goal partway through each task, and defines the revision along three operations and three components of the goal, forming nine intervention types that locate which kind of revision an agent fails on. The released benchmark contains 2,250 intervention instances. We evaluate twelve LLM agents and find that every intervention type degrades performance. Add interventions, which extend the goal rather than replace it, are consis- tently the hardest. These results establish adapting to a mid-task intervention as a substantial and open challenge for current LLM agents. Our code and data are available at https://github.com/jhn25/InterventionBench.
[2-9] Group-Aware Adaptive Retrieval for Evidence Navigation (EMNLP 2026)
박준(석박통합과정)
Reasoning-intensive retrieval addresses queries whose relevance cannot be identified by surface-level matching, thereby requiring multi-step reasoning. Because relevant documents rarely appear in the initial candidate set, retrieval systems suffer from the bounded recall problem. Existing methods iteratively expand a candidate pool at the document level over a corpus graph, examining each neighbor in isolation and drifting toward a narrow region of the corpus. To address this problem, we propose Group-Aware Adaptive Retrieval for Evidence Navigation (GAREN), which explores the corpus graph through group-level expansion. GAREN organizes documents into semantically coherent and distinguishable groups based on their connections in the corpus graph. The information in each group indicates what can be accessed by expanding through it, providing guidance beyond individual document-level signals. At each iteration, GAREN uses a group-level navigator to identify promising expansion directions, retrieves documents from the selected groups, and applies a document-level reranker to evaluate the updated candidate set. Extensive experiments show that GAREN achieves up to 8.0% improvement over the strongest baseline on BRIGHT.
[2-10] ICGS: Illumination Cue-Driven Densification for 3D Gaussian Splatting (ACCV 2026)
김동빈(석사과정)
We present Illumination Cue-driven densification for 3D Gaussian Splatting (ICGS). Effective densification should allocate more primitives to regions with high geometric complexity, but the underlying scene geometry is unavailable under image-space supervision alone. Hence, existing methods primarily rely on photometric signals such as view-space positional gradients, which can become weak once a primitive is locally stationary and may therefore under-densify regions containing fine geometric structures. To better drive densification, we leverage implicit structural cues reflected in illumination patterns. Our key observation is that the underlying 3D geometry is unknown, but its surface relief, orientation, and visibility relationships leave indirect traces in illumination patterns such as shading and shadows. The statistics of such patterns can reveal structural discrepancies in the representation of underlying scene structures. Our experiments show that our approach greatly improves rendering quality with fewer primitives.
[2-11] Reconfiguring Procedural Knowledge for Compositional Robot Skill Adaptation (NeurIPS 2026)
이대희(석박통합과정)
General-purpose behavior models increasingly internalize broad procedural knowledge during training, making their adaptation to new goals and skill interfaces less a problem of learning new behaviors from scratch than of reconfiguring existing competence. We study this capability as reconfigurability, the ability to realize new behaviors by cost-efficiently restructuring existing components and relations under sparse governing information, where task and procedure specifications provide high-level goals and reusable procedural factors but leave concrete bindings and compositions underdetermined. We propose RecPo, which stores skill-induced transition records task-agnostically in a procedure knowledge base and composes them only for a new task. The knowledge base indexes each record under multiple procedure views, each exposing one factor while preserving local transition relations. At test time, RecPo reads a sparse task demonstration through these views and turns view-level changes into a query whose resolution over stored relations yields an executable skill sequence. On an extended Franka Kitchen benchmark with held-out long-horizon tasks outside stored realizations, RecPo achieves 91.3% ordered subtask completion at 7.2 ms per episode and outperforms skill-retrieval and skill-chaining baselines given the same memory. Real-robot case studies further show that local updates to views and records enable deployment-time adaptation without task-specific retraining.
[2-12] Skill-Adaptive Noise Scheduling for Diffusion Policies (NeurIPS 2026)
김우경(박사과정)
Recent advances in diffusion policies have demonstrated strong effectiveness when integrated into hierarchical skill-based learning frameworks, leveraging offline data to capture multimodal and temporally abstracted behaviors. However, prior approaches have rarely considered the diffusion process as a learnable component, adopting fixed noise schedules applied uniformly across skills, regardless of their distinct behavioral contexts. Such a design fundamentally limits the achievable performance, particularly for contact-rich and dexterous manipulation tasks where different skills exhibit distinct couplings between action dimensions. In this paper, we present a skill-adaptive noise diffusion framework (SaND) which introduces a learnable, skill-conditioned noise schedule that modulates the diffusion process per action dimension, adapting the noising dynamics to each skill. To prevent the learned noise schedule from collapsing into trivial solutions, we employ an auxiliary inverse objective that reconstructs the skill embedding from the Legendre coefficients of the derivative of the noise schedule, encouraging the schedules to remain skill-discriminative. This formulation naturally enables adaptive control over the denoising steps for each skill, where noise schedules are clustered in the Legendre coefficient space, and the appropriate number of denoising steps is determined at the point where reconstruction error abruptly increases. At inference, we identify each skill embedding's cluster from its noise schedule and apply the corresponding number of denoising steps, dedicating more compute to skills requiring fine-grained control and less to simpler ones. Through experiments, we demonstrate that SaND outperforms existing diffusion-based skill learning and planning baselines on contact-rich and dexterous manipulation tasks, achieving higher success rates while requiring fewer denoising steps.
[2-13] Distributional Open-Ended Evaluation of LLM Cultural Value Alignment Based on Value Codebook (ICML 2026)
이재혁(박사과정)
As LLMs are globally deployed, aligning their cultural value orientations is critical for safety and user engagement. However, existing benchmarks face the Construct-Composition-Context (C^3) challenge: relying on discriminative, multiple-choice formats that probe value knowledge rather than true orientations, overlook subcultural heterogeneity, and mismatch with real-world open-ended generation. We introduce DOVE, a distributional evaluation framework that directly compares human-written text distributions with LLM-generated outputs. DOVE utilizes a rate-distortion variational optimization objective to construct a compact value codebook from 10K documents, mapping text into a structured value space to filter semantic noise. Alignment is measured using unbalanced optimal transport, capturing intra-cultural distributional structures and subgroup diversity. Experiments across 12 LLMs show that DOVE achieves superior predictive validity, attaining a 31.56% correlation with downstream tasks, while maintaining high reliability with as few as 500 samples per culture.
[2-14] RangeGuard: Efficient, Bounded Approximate Error Correction for Reliable DNNs (ISCA 2026)
고한음(석박통합과정)
As DRAM scales in density and adopts 3D integration, raw fault rates increase and multi-bit errors are no longer rare. Such errors can severely impact Deep Neural Networks (DNNs): although DNNs tolerate small numerical perturbations, random bit flips can create extreme outliers that propagate and sharply degrade accuracy. Large Language Models (LLMs) are particularly vulnerable because attention, residual, and normalization layers can amplify and preserve a single corrupted activation across many layers, destabilizing inference. This paper introduces RangeGuard, a metadata-centric error-correcting framework that provides strong reliability and high efficiency based on bounded approximate correction. Instead of protecting raw bits, RangeGuard encodes compact Range Identifiers (RIDs) that capture the numerical range of each value. These compact metadata enable efficient use of limited redundancy and concentrate protection on range changes, which indicate harmful semantic deviations, while ignoring benign intra-range variations. Upon detecting a range change, RangeGuard restores the correct range and substitutes a representative value, ensuring that error magnitudes are bounded within the range. Based on RIDs, RangeGuard can tolerate 64+ flipped bits using only 16 bits of parity available in GPU memories without a noticeable accuracy loss. By introducing semantic range protection, RangeGuard enables reliable DNN execution even under frequent memory errors and tight redundancy budgets.
[2-15] Control First, Robustness Next: Decoupled Representation Learning for Visual RL Generalization (NeurIPS 2026)
허찬용(박사과정)
Visual reinforcement learning (RL) is promising for real-world problems such as autonomous driving, robotic locomotion, and manipulation, but remains highly vulnerable to visual distribution shifts between training and deployment. Existing methods improve robustness through Q-consistency regularization, masking, and auxiliary objectives, yet most of them follow a coupled representation learning paradigm in which control-oriented and robustness-oriented representation learning are jointly optimized within the same training framework. We argue that this coupled structure can introduce harmful interference between
control learning and robustness learning, degrading original-environment control performance and limiting generalization under visual shifts. To address this issue, we propose Separate Then Align Representations (STAR), which reformulates visual RL generalization as a problem of decoupled representation learning. Specifically, we first learn a control-relevant reference representation through online RL under weak augmentation, and then perform an offline post-representation alignment stage that maps strongly augmented inputs back to this reference representation. Experiments on RL-ViGen benchmarks spanning DMControl and Robosuite show that STAR achieves strong generalization under visual distribution shifts while better preserving performance in the original environment than prior coupled approaches. Our source code is available in the supplementary material.
[2-16] UnPII: Unlearning Personally Identifiable Information with Quantifiable Exposure Risk (ICSE-SEIP 2026)
권유정(박사과정)
The ever-increasing adoption of Large Language Models in critical sectors like finance, healthcare, and government raises privacy concerns regarding the handling of sensitive Personally Identifiable Information (PII) during training. In response, regulations such as European Union’s General Data Protection Regulation (GDPR) mandate the deletion of PII upon requests, underscoring the need for reliable and cost-effective data removal solutions. Machine unlearning has emerged as a promising direction for selectively forgetting data points. However, existing unlearning techniques typically apply a uniform forgetting strategy that neither accounts for the varying privacy risks posed by different PII attributes nor reflects associated business risks. In this work, we propose UnPII, the first PII-centric unlearning approach that prioritizes forgetting based on the risk of individual or combined PII attributes. To this end, we introduce the PII risk index (PRI), a composite metric that incorporates multiple dimensions of risk factors: identifiability, sensitivity, usability, linkability, permanency, exposability, and compliancy. The PRI enables a nuanced evaluation of privacy risks associated with PII exposures and can be tailored to align with organizational privacy policies. To support realistic assessment, we systematically construct a synthetic PII dataset (e.g., 1,700 PII instances) that simulates realistic exposure scenarios. UnPII seamlessly integrates with established unlearning algorithms, such as Gradient Ascent, Negative Preference Optimization, and Direct Preference Optimization, without modifying their underlying principles. Our experimental results demonstrate that UnPII achieves the improvements of accuracy up to 11.8%, utility up to 6.3%, and generalizability up to 12.4%, respectively, while incurring a modest fine-tuning overhead of 27.5% on average during unlearning.
[2-17] ECO Decoding: Entropy-Based Control for Controllability and Fluency in Controllable Dialogue Generation (EMNLP 2025)
김두영(박사과정)
Controllable Dialogue Generation (CDG) enables chatbots to generate responses with desired attributes, and weighted decoding methods have achieved significant success in the CDG task. However, using a fixed constant value to manage the bias of attribute probabilities makes it challenging to find an ideal control strength that satisfies both controllability and fluency. To address this issue, we propose ECO decoding Entropy-based COntrol, which dynamically adjusts the control strength at each generation step according to the model’s entropy in both the language model and attribute classifier probability distributions. Experimental results on DailyDialog and MultiWOZ datasets show that our method achieves improved control accuracy while maintaining fluency and grammar, outperforming previous decoding methods across various models and settings. Furthermore, ECO decoding alleviates probability interpolation issues in multi-attribute generation, demonstrating its robust performance in both single and multi-attribute scenarios.
[2-18] Multi-scale Mixture of World Models for Embodied Agents in Evolving Environments (ECCV 2026)
장진우(석박통합과정)
Embodied agents operating in the real world require multi-scale reasoning and knowledge adaptation as conditions change. We identify two challenges in applying Mixture of Experts (MoE) to this setting: routing lacks an explicit notion of scale, preventing targeted updates at specific scales, and a uniform update policy cannot accommodate the different rates at which knowledge at each scale becomes outdated. We present MuSix, a framework that addresses both challenges through scale-aware world model mixture and evolution. A two-stage routing mechanism grounds scale selection in experiential distance, a measure of situational novelty inspired by Construal Level Theory: a meta-router first maps this quantity to a weight over continuous scale space, then per-scale base routers select world models within the identified scale. For adaptation, scale-dependent forgetting rates allow low-scale knowledge to refresh rapidly while high-scale abstractions persist, and gated inter-scale transfer maintains coherence across the hierarchy. Experiments on EmbodiedBench and HAZARD show that MuSix improves over state-of-the-art baselines on multi-scale reasoning and dynamic adaptation.
[2-19] REFLEX: Rewrite-Free Row-Aligned Sparse Attention for Efficient LLM Execution on PIM (DAC 2026)
박주홍(석박통합과정)
Large language models (LLMs) face decoding bottlenecks as attention repeatedly accesses the key-value (KV) cache. Sparse attention and processing-in-memory (PIM) each reduce data movement, but their naive integration produces irregular KV accesses that span multiple DRAM rows, leading to unnecessary activations and cache rewrites. We present REFLEX, a rewrite-free sparse attention framework that colocates required KV entries in a single DRAM row and applies activation-aware scheduling for PIM execution. REFLEX preserves accuracy without hardware changes, achieving up to 1.64× throughput and 1.36× energy efficiency on PIM, and 1.37× throughput in GPU-PIM systems.
[2-20] When Visual Frames Flip Mental-State Answers: An Item-Level Audit of Multimodal Theory-of-Mind Reasoning (CIKM 2026)
홍서현(석사과정)
Multimodal Theory-of-Mind (ToM) evaluation often assumes that visual frames provide useful evidence for answering questions about beliefs, goals, and intentions. We test this assumption with an item-level audit of MuMA-ToM, a benchmark of 900 multiple-choice questions about multi-agent household interactions. The audit compares six settings: question-and-answer only (QA-only), video frames only (Video-only), text only (Text-only), text paired with frames (Text+Frame), generic chain-of-thought prompting (CoT), and QLogic, a structured text-only prompt that makes the question logic explicit. Across GPT-4o, Gemini 2.5 Flash, and Qwen2.5-VL-7B-Instruct, Video-only outperforms QA-only, showing that frames can be useful. However, Text+Frame does not outperform Text-only for any model, and GPT-4o shows a significant paired drop. We therefore analyze answer transitions rather than aggregate accuracy alone. The main transition is Vision-Induced ToM Degradation, defined operationally as an item that is correct under Text-only but wrong under Text+Frame. On GPT-4o degradation cases, generic CoT recovers 14 of 78 items, while QLogic recovers 62 of 78; an exact McNemar test confirms the paired recovery difference (𝑝= 2.92 ×10−12). Ablations show that condition tracking and candidate comparison are the most informative QLogic components. The results support a narrow conclusion suited to short-form multimodal evaluation: added visual evidence should be audited for how it changes individual mental-state answers, not only for whether it improves average accuracy.
[2-21] Low-pass Personalized Subgraph Federated Recommendation (ICLR 2026)
심우석(박사과정)
Federated Recommender Systems (FRS) preserve privacy by training decentralized models on client-specific user-item subgraphs without sharing raw data. However, FRS faces a unique challenge: subgraph structural imbalance, where drastic variations in subgraph scale (user/item counts) and connectivity (item degree) misalign client representations, making it challenging to train a robust model that respects each client’s unique structural characteristics. To address this, we propose a Low-pass Personalized Subgraph Federated recommender system (LPSFed). LPSFed leverages graph Fourier transforms and low-pass spectral filtering to extract low-frequency structural signals that remain stable across subgraphs of varying size and degree, allowing robust personalized parameter updates guided by similarity to a neutral structural anchor. Additionally, we leverage a localized popularity bias-aware margin that captures item-degree imbalance within each subgraph and incorporates it into a personalized bias correction term to mitigate recommendation bias. Supported by theoretical analysis and validated on five realworld datasets, LPSFed achieves superior recommendation accuracy and enhances model robustness.
[2-22] Anchor-Regularized Adaptation for Generalizable AI-Generated Image Detection with DINOv3 (ACM MM 2026)
이주헌(석박통합과정)
Recent works in AI-generated image detection have shown that careful training data alignment can improve generalization by removing spurious correlations. However, linear probes on frozen DINOv3 representations achieve remarkably strong performance even when trained on misaligned datasets. Motivated by this result, we analyze the underlying rationale and the limits of this generalization. We find that frozen DINOv3 performs well because its decisions rely on features that faithfully represent the space of authentic images. At the same time, its final layer is less effective at capturing the subtle pixel-artifact cues that can be emphasized by aligned training data. We further observe that naively mixing aligned and misaligned data during adaptation improves sensitivity to such cues but at the cost of distorting the pre-trained representation, limiting generalization. To address this issue, we propose Anchor-Regularized Adaptation (ARA). We apply Low-Rank Adaptation to capture pixel-level artifacts while leveraging a frozen anchor classifier to avoid deviations from the original representation structure. This allows the model to exploit pixel-artifact cues without sacrificing generalization. Our method achieves state-of-the-art performance on nine diverse and challenging benchmarks, indicating that ARA enables complementary supervision from misaligned and aligned data for more effective detection.