论文检索

Jiaqi Weng, Han Zheng, Hanyu Zhang, Ej Zhou, Qinqin He, Jialing Tao, Hui Xue, Zhixuan Chu, Xiting Wang

Sparse autoencoders (SAEs) enable interpretability research by decomposing entangled model activations into monosemantic features. However, under what circumstances SAEs derive most fine-grained latent features for safety—a low-frequency concept domain—remains unexplored. Two key challenges exist: identifying SAEs with the greatest potential for generating safety domain-specific features, and the prohibitively high cost of detailed feature explanation. In this paper, we propose **Safe-SAIL**, a unified framework for interpreting SAE features in safety-critical domains to advance mechanistic understanding of large language models. Safe-SAIL introduces a pre-explanation evaluation metric to efficiently identify SAEs with strong safety domain-specific interpretability, and reduces interpretation cost by 55% through a segment-level simulation strategy. Building on Safe-SAIL, we train a comprehensive suite of SAEs with human-readable explanations and systematic evaluations for 1,758 safety-related features spanning four domains: pornography, politics, violence, and terror. Using this resource, we conduct empirical analyses and provide insights on the effectiveness of Safe-SAIL for risk feature identification and how safety-critical entities and concepts are encoded across model layers. All models, explanations, and tools are publicly released in an open-source toolkit at https://anonymous.4open.science/r/Safe-SAIL/.

Hengyi Feng, Zeang Sheng, Meiyi Qiang, Yang Li, Wentao Zhang

Despite the remarkable success of multimodal large language models (MLLMs) in generative tasks, we observe that they exhibit a counterintuitive deficiency in the zero-shot multimodal retrieval task. In this work, we investigate the underlying mechanisms that hinder MLLMs from being effective retrievers. With the help of sparse autoencoders (SAEs), we decompose MLLM output representations into interpretable semantic concepts to probe their intrinsic behavior. Our analysis reveals that the representation space of MLLMs is overwhelmingly dominated by textual semantics; and the visual semantics essential for multimodal retrieval only constitute a small portion. We find that this imbalance is compounded by the heavy focus of MLLMs on bridging image-text modalities, which facilitates generation but homogenizes embeddings and finally diminishes the discriminative power required for multimodal retrieval. We further discover that the specific feature components that contribute most to the similarity computations of MLLMs are actually distractors that greatly reduce retrieval performance. Building on these insights, we propose , a test-time adaptation approach that applies a whitening transformation to adjust the geometry of MLLM representation spaces. Empirical results show that this simple intervention consistently improves zero-shot multimodal retrieval performance across diverse MLLMs without fine-tuning efforts.

Chenye Zou, Difan Jiao, Lijie Hu

Large Language Models (LLMs) are increasingly deployed worldwide, yet they exhibit strong Western-centric biases, and the internal mechanisms governing their cultural behaviors remain poorly understood. Prior work has identified so-called cultural neurons, but individual neurons are often polysemous, conflating abstract cultural knowledge with surface-level lexical cues due to superposition. We apply Sparse Autoencoders (SAEs) to decompose intermediate LLM activations into sparse, interpretable feature representations that disentangle these factors. This analysis reveals culturally selective features that remain invariant across paraphrasing and task formats, indicating abstraction beyond lexical correlations. Through targeted feature ablation, we provide causal evidence that these features are necessary for cultural reasoning: their removal selectively degrades performance on culturally conditioned tasks. Furthermore, we show that steering model activations along these feature directions is sufficient to systematically modulate cultural-related knowledge generation, without retraining. Together, our results offer the first causal evidence that LLMs encode cultural knowledge as decoupled semantic structures rather than surface patterns, enabling a scalable pathway toward cultural alignment through mechanistic intervention. Code is available at https://github.com/IAN-YE/Cultural-features-SAE.

Hieu Man, Van-Cuong Pham, Nghia Trung Ngo, Franck Dernoncourt, Thien Huu Nguyen

Learning robust representations of authorial style is crucial for authorship attribution and AI-generated text detection. However, existing methods often struggle with content-style entanglement, where models learn spurious correlations between authors’ writing styles and topics, leading to poor generalization across domains. To address this challenge, we propose Explainable Authorship Variational Autoencoder (EAVAE), a novel framework that explicitly disentangles style from content through architectural separation-by-design. EAVAE first pretrains style encoders using supervised contrastive learning on diverse authorship data, then finetunes with a Variational Autoencoder (VEA) architecture using separate encoders for style and content representations. Disentanglement is enforced through a novel discriminator that not only distinguishes whether pairs of style/content representations belong to the same or different authors/content sources, but also generates natural language explanation for their decision, simultaneously mitigating confounding information and enhancing interpretability. Extensive experiments demonstrate the effectiveness of EAVAE. On authorship attribution, we achieve state-of-the-art performance on various datasets, including Amazon Reviews, PAN21, and HRS. For AI-generated text detection, EAVAE excels in few-shot learning over the M4 dataset.

Aastha A K Verma, Anwoy Chatterjee, Mehak Gupta, Tanmoy Chakraborty

Multilingual language models (LMs) organize representations for typologically and orthographically diverse languages into a shared parameter space, yet the nature of this internal organization remains elusive. In this work, we investigate which linguistic properties — abstract language identity or surface-form cues — shape multilingual representations. To do so, we analyze language-associated units across different model families and scales using the Language Activation Probability Entropy (LAPE) metric, and further decompose activations with Sparse Autoencoders. We find that these units are strongly conditioned on orthography: romanization induces near-disjoint representations that align with neither native-script inputs nor English, while word-order shuffling has limited effect on unit identity. Probing shows that typological structure becomes increasingly accessible in deeper layers, while causal interventions indicate that generation is most sensitive to units that are invariant to surface-form perturbations rather than to units identified by typological alignment alone. Overall, our results suggest that multilingual LMs organize representations around surface form, with linguistic abstraction emerging gradually without collapsing into a unified interlingua.

Jiwei Tang, Zhicheng Zhang, Shunlong Wu, Jingheng Ye, Lichen Bai, Zitai Wang, Tingwei Lu, Lin Hai, Yiming Zhao, Hai-Tao Zheng 等

Large Language Models (LLMs) have achieved remarkable performance across a wide range of Natural Language Processing (NLP) tasks. However, in long-context scenarios, they face two challenges: high computational cost and information redundancy. To address these challenges, we propose GMSA, an encoder-decoder context compression framework that generates a compact sequence of soft tokens for downstream tasks. GMSA introduces Group Merging to achieve more uniform aggregation, mitigating semantic dominance during autoencoder pretraining, and Layer Semantic Alignment (LSA) to bridge the semantic gap between high-level abstract semantics and low-level input semantics. We first pretrain GMSA as an autoencoder and then fine-tune it for downstream tasks. Experiments demonstrate that GMSA improves context reconstruction compared to existing soft prompt compression paradigm and outperforms baselines on multiple long-context question answering and summarization benchmarks across two backbone models, while maintaining low end-to-end latency.

Xiangchen Song, Aashiq Muhamed, Yujia Zheng, Lingjing Kong, Zeyu Tang, Mona T. Diab, Virginia Smith, Kun Zhang

Sparse Autoencoders (SAEs) are a prominent tool in mechanistic interpretability (MI) for decomposing neural network activations into interpretable features. However, the aspiration to identify a canonical set of features is challenged by the observed inconsistency of learned SAE features across different training runs, undermining reproducibility and complicating model comparison. We study run-to-run feature consistency in SAEs and argue that it should be reported as a standard evaluation axis alongside reconstruction and sparsity. We propose the Pairwise Dictionary Mean Correlation Coefficient (PW-MCC) as an assignment-based metric to quantify consistency and demonstrate that high levels are achievable (PW-MCC ≈ 0.80 for TopK SAEs on LLM activations) with appropriate architectural choices.Our contributions include: (i) theoretical grounding for strong consistency in the idealized setting of TopK SAEs; (ii) synthetic validation using a model organism, which verifies PW-MCC as a reliable proxy for ground-truth recovery; and (iii) empirical analysis on LLM activations, where PW-MCC correlates with the similarity of automatically generated natural-language feature explanations.

Deep Learning · Generative Models and Autoencoders

Erkan Turan, Ari Siozopoulos, Louis Martinez, Julien Gaubil, Emery Pierson, Maks Ovsjanikov

Continuous Normalizing Flows (CNFs) enable elegant generative modeling but remain bottlenecked by their iterative nature requiring costly sampling and lacking interpretability of the intermediate states. Recent approaches accelerate sampling by straightening trajectories or distilling endpoints, yet they treat the original generative process as a black box, discarding the teacher’s intermediate dynamics. We propose a fundamentally different perspective: globally linearizing flow dynamics via Koopman theory to achieve trajectory-preserving linearization. By lifting Conditional Flow Matching (CFM) into a higher-dimensional Koopman space, we represent its evolution with a single linear operator. Crucially, unlike boundary-only distillation, our method enforces infinitesimal consistency with the teacher's vector field along *the full generative path*. We derive a practical, simulation-free training objective that ensures this global alignment and yields two key benefits. First, sampling becomes one-step and parallelizable. Second, because the linearization is faithful to the dynamics, the Koopman operator provides unique insights on the generation. We demonstrate that this structure enables novel applications unavailable in prior approaches, including discovery of semantically coherent editing directions, inversion with a teacher-aligned linear operator and class-conditional spectral signatures. Empirically, our approach achieves competitive sample quality, while enabling spectral analysis and control of the *entire trajectories* of generative flows.

Deep Learning · Generative Models and Autoencoders

Zehua Chen, Yucheng Yang, Binjie Yuan, Kaiwen Zheng, Jun Liu, Jun Zhu

Guidance methods, e.g., classifier-free guidance (CFG) and auto-guidance (AG), have distinctively improved noise-to-data diffusion generation results. Recently, bridge models have been proposed, which present a data-to-data sampling process to exploit instructive information from clean prior representation, showing advantages on the tasks such as image-to-image translation. In this work, we design a custom guidance method for bridge models, named prior guidance (PG). Different from highlighting condition alignment (CFG) or score accuracy (AG), we training-freely construct an additional weak prior for the pre-trained bridge models, and extrapolate the estimation results to further encourage prior exploitation. Then, we analyze the underlying mechanism of prior exploitation in bridge process and design frequency-modulated prior guidance (FMPG), which tailors the guidance scale to low- and high-frequency bands coherent with bridge generative dynamics. Finally, considering the challenge of bridge models on image in-painting, we develop a cascaded guidance framework, CFG-FMPG, that first generates a coarse prior under global semantic condition and then refines it with FMPG, naturally fulfilling their complementary advantages along sampling trajectory. Experiments conducted on strong pre-trained bridge models, DDBM and DBIM, valid the consistent improvement achieved by our training-free design.

Deep Learning · Generative Models and Autoencoders

Ganggui Ding, Xiaogang Xu, Hao Chen, Chunhua Shen

Generative video diffusion models have shown strong robustness to large motion and occlusions for video frame interpolation (VFI). However, their inference efficiency lags significantly behind learning-based methods due to the structural redundancy of pairwise inference and the procedural latency of multi-step iterative denoising. To address these limitations, we propose SpeedVFI, a one-step diffusion framework that achieves dual efficiency improvements by interpolating the entire video sequence in a single forward pass to eliminate pairwise overhead, and distilling the generation trajectory into a one-step denoising process to bypass iterative latency. To support this high-efficiency architecture, we introduce temporal RoPE alignment to ensure temporal consistency across the unified sequence, and noise-centric partial attention to reduce computational overhead while preserving global context. Extensive experiments demonstrate that SpeedVFI accelerates diffusion-based VFI by orders of magnitude while maintaining competitive quantitative and visual quality.

Deep Learning · Generative Models and Autoencoders

Rajalaxmi Rajagopalan, Debottam Dutta, Yu-Lin Wei, Romit Roy Choudhury

Imagine Alice has a specific image $x^\ast$ in her mind, say, the view of the street in which she grew up during her childhood. To generate that exact image, she guides a generative model with multiple rounds of prompting and arrives at an image $x^{p*}$. Although $x^{p*}$ is reasonably close to $x^\ast$, Alice finds it difficult to close that gap using language prompts. This paper aims to narrow this gap by observing that even after language has reached its limits, humans can still tell when a new image $x^+$ is closer to $x^\ast$ than $x^{p*}$. Leveraging this observation, we develop **MultiBO** (Multi-Choice Preferential Bayesian Optimization) that carefully generates $K$ new images as a function of $x^{p*}$, gets preferential feedback from the user, uses the feedback to guide the diffusion model, and ultimately generates a new set of $K$ images. We show that within $B$ rounds of user feedback, it is possible to arrive much closer to $x^\ast$, even though the generative model has no information about $x^\ast$. Qualitative scores from $30$ users, combined with quantitative metrics compared across $5$ baselines, show promising results, suggesting that multi-choice feedback from humans can be effectively harnessed for personalized image generation.

Sen Ye, Jianning Pei, Mengde Xu, Shuyang Gu, Chunyu Wang, Liwei Wang, Han Hu

Most visual generative models compress images into a latent space before applying diffusion or autoregressive modelling. Yet, existing approaches such as VAEs and foundation model aligned encoders implicitly constrain the latent space without explicitly shaping its distribution, making it unclear which types of distributions are optimal for modeling. We introduce **Distribution-Matching VAE** (**DMVAE**), which explicitly aligns the encoder’s latent distribution with an arbitrary reference distribution via a distribution matching constraint. This generalizes beyond the Gaussian prior of conventional VAEs, enabling alignment with distributions derived from self-supervised features, diffusion noise, or other prior distributions. With DMVAE, we can systematically investigate which latent distributions are more conducive to modeling, and we find that SSL-derived distributions provide an excellent balance between reconstruction fidelity and modeling efficiency, reaching a gFID of 3.2 on ImageNet with only 64 training epochs. Our results suggest that choosing a suitable latent distribution structure (achieved via distribution-level alignment), rather than relying on fixed priors, is key to bridging the gap between easy-to-model latents and high-fidelity image synthesis.

Deep Learning · Generative Models and Autoencoders

Zixun Xia, Shuyu Guo, Boqian Li, jian Yang, Kai Wang, Yaxing Wang

Text-to-image models have advanced high-fidelity content generation, but their inability to maintain subject consistency hampers realistic applications. Existing training-based methods rely on heavy computation and large datasets; while training-free approaches demand excessive memory or complex auxiliary modules. In this paper, we first reveal a key property overlooked in prior works that the identity-relevant signals, termed Identity-Preserving Embeddings (*IPemb*), are implicitly encoded in textual embeddings of frame prompts. To address the consistent T2I generation with the *IPemb* embedding, we propose Boost Identity-Preserving Embedding (*BIPE*), a training-free yet plug-and-play framework that explicitly extracts and enhances the *IPemb*. Its core innovations are two complementary techniques: First, Adaptive Singular-Value Rescaling (*adaSVR*) applies singular-value decomposition to the joint embedding matrix of all frame prompts, amplifying identity-centric components while suppressing frame-specific noise. Second, Union Key (*UniK*) further reinforces consistency by aligning the T2I backbone’s image-text attention across the entire generation sequence. Experiments on the *ConsiStory+* benchmark demonstrate *BIPE* outperforms existing methods in both qualitative and quantitative metrics. To address the gap in evaluating a broader range of scenarios with diversified prompt templates, we introduce a *DiverStory* benchmark to further confirm our scalability.

Deep Learning · Generative Models and Autoencoders

Liyu Zhang, Kehan Li, Tao Zhou, Zeyi Huang, Chao LI, Jiming Chen

Reward feedback learning (ReFL) is effective for both text-to-image (T2I) and text-to-video (T2V) generation with image reward models (RMs). However, image RMs are misaligned with temporal objectives of T2V, motivating ReFL with video reward models. Nevertheless, directly deploying video RMs is impractical due to their large parameter size and the prohibitive cost of memory. To address this, we propose VELR: an efficient framework that employs ensemble latent reward models (LRMs) to predict rewards directly in latent space, bypassing expensive backpropagation through VAE decoders and video RMs. Specifically, we introduce the ensemble technique for the LRM, which enhances capacity, quantifies uncertainty, and mitigates reward hacking. VELR achieves a reduction of up to 150GB in memory, requiring as little as 12.4% of the memory compared to standard ReFL. Experiments on OpenSora-1.2, CogVideoX-1.5, and Wan-2.1 with large-scale video RMs demonstrate that VELR achieves comparable performance as standard ReFL and enables efficient and robust video RM-based ReFL at scales previously unattainable.

Tianhao Zhao, Youjia Zhang, Hang Long, Jinshen Zhang, Wenbing Li, Yang Yang, Gongbo Zhang, Jozef Hladký, Matthias Nießner, Wei Yang

In this paper, we introduce LATO, a novel topology-preserving latent representation that enables scalable, flow matching-based synthesis of explicit 3D meshes. LATO represents a mesh as a Vertex Displacement Field (VDF) anchored on surface, incorporating a sparse voxel Variational Autoencoder (VAE) to compress this explicit signal into a structured, topology-aware voxel latent. To decapsulate the mesh, the VAE decoder progressively subdivides and prunes latent voxels to instantiate precise vertex locations. In the end, a dedicated connection head queries the voxel latent to predict edge connectivity between vertex pairs directly, allowing mesh topology to be recovered without isosurface extraction or heuristic meshing. For generative modeling, LATO adopts a two-stage flow matching process, first synthesizing the structure voxels and subsequently refining the voxel-wise topology features. Compared to prior isosurface/triangle-based diffusion models and autoregressive generation approaches, LATO generates meshes with complex geometry, well-formed topology while being highly efficient in inference.

Deep Learning · Generative Models and Autoencoders

Binxu Wang, Jacob A Zavatone-Veth, Cengiz Pehlevan

Diffusion models trained on different, non-overlapping subsets of a dataset often produce strikingly similar outputs when given the same noise seed. We trace this consistency to a simple linear effect: the shared Gaussian statistics across splits already predict much of the generated images. To formalize this, we develop a random matrix theory (RMT) framework that quantifies how finite datasets shape the expectation and variance of the learned denoiser and sampling map in the linear setting. For expectations, sampling variability acts as a renormalization of the noise level through a self-consistent relation $\sigma^2\to\kappa(\sigma^2)$, explaining why limited data overshrink low-variance directions and pull samples toward the dataset mean. For fluctuations, our variance formulas reveal three key factors behind cross-split disagreement: \textit{anisotropy} across eigenmodes, \textit{inhomogeneity} across inputs, and overall scaling with dataset size. Extending deterministic-equivalence tools to fractional matrix powers further allows us to analyze entire sampling trajectories. The theory sharply predicts the behavior of linear diffusion models, and we validate its predictions on UNet and DiT architectures in their non-memorization regime, identifying where and how samples deviates across training data split. This provides a principled baseline for reproducibility in diffusion training, linking spectral properties of data to the stability of generative outputs.

Deep Learning · Generative Models and Autoencoders

Brett Levac, Jon Tamir, Marcelo Pereyra, Julián Tachella

Diffusion models (DMs) are a powerful framework for image generation and restoration. However, existing DMs are primarily trained in a supervised manner by using a large corpus of clean images. This poses fundamental challenges in many real-world scenarios, where acquiring noise-free data is hard or infeasible. While some methods are capable of training DMs using noisy data, they are effective only when the amount of noise is very mild or when additional noise-free data is available. In addition, existing methods for training DMs from incomplete measurements require access to multiple complementary acquisition processes, a significant practical limitation. Here we introduce the first approach for learning DMs for image restoration using only noisy measurement data from a single operator. First, we show that DMs, and more broadly minimum mean squared error denoisers, exhibit a weak form of scale equivariance linking rescaling in signal amplitude to changes in noise intensity. We then leverage this theoretical insight to develop a denoising score-matching strategy that generalizes robustly to noise levels below the training data, thereby enabling the learning of DMs from noisy measurements. For problems involving measurements both noisy and incomplete, we integrate our method with equivariant imaging, a complementary self-supervised learning framework that exploits the inherent invariants of imaging problems. This allows training DMs for image restoration from single-operator noisy measurements. We validate the effectiveness of our approach through extensive experiments on image denoising, demosaicing, inpainting, and MRI reconstruction along with comparisons with the state of the art.

Deep Learning · Generative Models and Autoencoders

Xiaofeng Lin, Seungbae Kim, Zhuoya Li, Zachary DeSoto, Charles Fleming, Guang Cheng

Deep generative models can help with data scarcity and privacy by producing synthetic training data, but they struggle in low-data, imbalanced tabular settings to fully learn the complex data distribution. We argue that striving for the full joint distribution could be overkill; for greater data efficiency, models should prioritize learning the conditional distribution $P(y\mid \bm{X})$, as suggested by recent theoretical analysis. Therefore, we overcome this limitation with \textbf{ReTabSyn}, a \textbf{Re}inforced \textbf{Tab}ular \textbf{Syn}thesis pipeline that provides direct feedback on feature correlation preservation during synthesizer training. This objective encourages the generator to prioritize the most useful predictive signals when training data is limited, thereby strengthening downstream model utility. We empirically fine-tune a language model-based generator using this approach, and across benchmarks with small sample sizes, class imbalance, and distribution shift, ReTabSyn consistently outperforms state-of-the-art baselines. Moreover, our approach can be readily extended to control various aspects of synthetic tabular data, such as applying expert-specified constraints on generated observations.

Deep Learning · Generative Models and Autoencoders

Kaiyuan Deng, Bo Hui, Gen Li, Jie Ji, Minghai Qin, Geng Yuan, Xiaolong Ma

The widespread adoption of text-to-image (T2I) diffusion models has raised concerns about their potential to generate copyrighted, inappropriate, or sensitive imagery learned from massive training corpora. As a practical solution, machine unlearning aims to selectively erase unwanted concepts from a pre-trained model without retraining from scratch. While most existing methods are effective for single-concept unlearning, they often struggle in real-world scenarios that require removing multiple concepts, since extending them to this setting is both non-trivial and problematic, causing significant challenges in unlearning effectiveness, generation quality, and sensitivity to hyperparameters and datasets. In this paper, we take a unique perspective on multi-concept unlearning by leveraging model sparsity and propose the Forget It All (FIA) framework. FIA first introduces Contrastive Concept Saliency to quantify each weight connection’s contribution to a target concept. It then identifies Concept-Sensitive Neurons by combining temporal and spatial information, ensuring that only neurons consistently responsive to the target concept are selected. Finally, FIA constructs masks from the identified neurons and fuses them into a unified multi-concept mask, where Concept-Agnostic Neurons that broadly support general content generation are preserved while concept-specific neurons are pruned to remove the targets. FIA is training-free and requires only minimal hyperparameter tuning for new tasks, thereby promoting a plug-and-play paradigm. Extensive experiments across three distinct unlearning tasks demonstrate that FIA achieves more reliable multi-concept unlearning, improving forgetting effectiveness while maintaining semantic fidelity and image quality.

Deep Learning · Generative Models and Autoencoders

Xuehui Yu, Fucheng Cai, Meiyi Wang, Xiaopeng Fan, Harold Soh

Inference-time guided sampling steers state-of-the-art diffusion and flow models without fine-tuning by interpreting the generation process as a controllable trajectory. This provides a simple and flexible way to inject external constraints (e.g., cost functions or pre-trained verifiers) for controlled generation. However, existing methods often fail when composing multiple constraints simultaneously, which leads to deviations from the true data manifold. In this work, we identify root causes of this off-manifold drift and find that local errors scale severely with multiple guidance misalignment. Building on these findings, we propose Conflict-Aware Additive Guidance ($g^\text{car}$), a lightweight and learnable method, which actively rectifies off-manifold drift by dynamically detecting and resolving gradient conflicts. We validate $g^\text{car}$ across diverse domains, ranging from synthetic datasets and image editing to generative decision-making for planning and control. Our results demonstrate that $g^\text{car}$ effectively rectifies off-manifold drift, surpassing baselines in generation fidelity while using light compute. Code is available at Anonymous Link.