论文检索

Moti Rattan Gupta, Anupam Sobti

Self Supervised Learning (SSL) has emerged as a prominent paradigm for label-efficient learning, and has been widely utilized by remote sensing foundation models (RSFMs). Recent RSFMs including SatMAE and DoFA primarily rely on masked autoencoding (MAE), contrastive learning or some combination of them. However, these pretext tasks often overlook the unique temporal characteristics of agricultural landscape, namely nature's cycle of sowing, growth, and harvest. Motivated by this gap, we propose three novel agriculture-specific pretext tasks, namely Time-Difference Prediction (TD), Temporal Frequency Prediction (FP), and Future-Frame Prediction (FF). Comprehensive evaluation on SICKLE dataset shows FF achieves 69.6% IoU on crop mapping and FP reduces yield prediction error to 30.7% MAPE, outperforming all baselines, and TD remains competitive on most tasks. Further, we also scale FF to the national scale of India, achieving 54.2% IoU outperforming all baselines on field boundary delineation on FTW India dataset.

Nirmalendu Prakash, Yeo Wei Jie, Amir Abdullah, Ranjan Satapathy, Erik Cambria, Roy Ka-Wei Lee

Refusal on harmful prompts is a key safety behaviour in instruction‑tuned large language models (LLMs), yet the internal causes of this behaviour remain poorly understood. We study two public instruction tuned models—Gemma‑2-2B‑IT and LLaMA‑3.1-8B‑IT using sparse autoencoders (SAEs) trained on residual‑stream activations. Given a harmful prompt, we search the SAE latent space for feature sets whose ablation flips the model from refusal to compliance, demonstrating causal influence and creating a jailbreak. Our search proceeds in three stages: 1. Refusal Direction - Finding a refusal mediating direction and collecting SAE features close to that direction, followed by 2. Greedy Filtering - to prune this set to obtain a minimal set and finally 3. Interaction Discovery - a factorization‑machine (FM) model that captures non‑linear interactions among the remaining active features and the minimal set. This pipeline yields a broad set of jailbreak-critical features, offering insight into the mechanistic basis of refusal. Moreover, we also find evidence of redundant features which remain dormant unless earlier features are suppressed. Our findings highlight the potential for fine-grained auditing and targeted intervention in safety behaviours by manipulating the interpretable latent space.

Chen Wang, Yue-Jiao Gong, Zhiguang Cao, Zeyuan Ma

To relieve intensive human-expertise required to design optimization algorithms, recent Meta-Black-Box Optimization (MetaBBO) researches leverage generalization strength of meta-learning to train neural network-based algorithm design policies over a predefined training problem set, which automates the adaptability of the low-level optimizers on unseen problem instances. Currently, a common training problem set choice in existing MetaBBOs is well-known benchmark suites CoCo-BBOB. Although such choice facilitates the MetaBBO's development, problem instances in CoCo-BBOB are more or less limited in diversity, raising the risk of overfitting of MetaBBOs, which might further results in poor generalization. In this paper, we propose an instance generation approach, termed as LSRE, which could generate diverse training problem instances for MetaBBOs to learn more generalizable policies. LSRE first trains an autoencoder which maps high-dimensional problem features into a 2-dimensional latent space. Uniform-grid sampling in this latent space leads to hidden representations of problem instances with sufficient diversity. By leveraging a genetic-programming approach to search function formulas with minimal L2-distance to these hidden representations, LSRE reverse engineers a diversified problem set, termed as Diverse-BBO. We validate the effectiveness of LSRE by training various MetaBBOs on Diverse-BBO and observe their generalization performances on either synthetic or realistic scenarios. Extensive experimental results underscore the superiority of Diverse-BBO to existing training set choices in MetaBBOs. Further ablation studies not only demonstrate the effectiveness of design choices in LSRE, but also reveal interesting insights on instance diversity and MetaBBO's generalization.

Omar Claflin

Current sparse autoencoder (SAE) approaches to neural network interpretability assume that activations can be decomposed through linear superposition into sparse, interpret-able features. Despite high reconstruction fidelity, SAEs consistently fail to eliminate polysemanticity and exhibit pathological behavioral errors. We propose that neural networks encode information in two complementary spaces com-pressed into the same substrate: feature identity and feature integration. To test this dual encoding hypothesis, we develop sequential and joint-training architectures to capture identity and integration patterns simultaneously. Joint training achieves 41.3% reconstruction improvement and 51.6% reduction in KL divergence errors. This architecture spontaneously develops bimodal feature organization: low squared norm features contributing to integration pathways and the rest contributing directly to the residual. Small nonlinear components (3% of parameters) achieve 16.5% standalone improvements, demonstrating parameter-efficient capture of computational relationships crucial for behavior. Additionally, intervention experiments using 2×2 factorial stimulus designs demonstrated that integration features exhibit selective sensitivity to experimental manipulations and produce systematic behavioral effects on model outputs, including significant interaction effects across semantic dimensions. This work provides systematic evidence for (1) dual encoding in neural representations, (2) meaningful nonlinearly encoded feature interactions, and (3) introduces an architectural paradigm shift from post-hoc feature analysis to integrated computational design, establishing foundations for next-generation SAEs.

Zhenliang Zhang, Xinyu Hu, Xiaojun Wan

Large language models sometimes inadvertently reproduce passages that are copyrighted, exposing downstream applications to legal risk. Most existing studies for inference-time defences focus on surface-level token matching and rely on external blocklists or filters, which add deployment complexity and may overlook semantically paraphrased leakage. In this work, we reframe copyright infringement mitigation as intrinsic semantic-space control and introduce SCOPE, an inference-time method that requires no parameter updates or auxiliary filters. Specifically, the sparse autoencoder (SAE) projects hidden states into a high-dimensional, near-monosemantic space; benefiting from this representation, we identify a copyright-sensitive subspace and clamp its activations during decoding. Experiments on widely recognized benchmarks show that SCOPE mitigates copyright infringement without degrading general utility. Further interpretability analyses confirm that the isolated subspace captures high-level semantics.

Shuyi Zhang, Wei Shi, Sihang Li, Jiayi Liao, Tao Liang, Hengxing Cai, Xiang Wang

Large language models (LLMs) have been widely deployed across numerous fields. Reinforcement Learning from Human Feedback (RLHF) leverages reward models (RMs) as proxies for human preferences to align LLM behaviors with human values, making the accuracy, reliability, and interpretability of RMs critical for effective alignment. However, traditional RMs lack interpretability, offer limited insight into the reasoning behind reward assignments, and are inflexible toward user preference shifts. While recent multidimensional RMs aim for improved interpretability, they often fail to provide feature-level attribution and require costly annotations. To overcome these limitations, we introduce the Sparse Autoencoder-Enhanced Reward Model (SARM), a novel architecture that integrates a pretrained Sparse Autoencoder (SAE) into a reward model. SARM maps the hidden activations of LLM-based RM into an interpretable, sparse, and monosemantic feature space, from which a scalar head aggregates feature activations to produce transparent and conceptually meaningful reward scores. Empirical evaluations demonstrate that SARM facilitates direct feature-level attribution of reward assignments, allows dynamic adjustment to preference shifts, and achieves superior alignment performance compared to conventional reward models.

Daipeng Zhang, Wenhuan Lu, Xianghu Yue, Hongcheng Zhang, Jianguo Wei

Dysarthric speech reconstruction (DSR) aims to enhance the intelligibility of dysarthric speech. Compared with normal speech, the dysarthric speech is characterized by its pathological features, including discontinuous pronunciation, slow speech, hoarseness, and improper pauses. Significant disparities in the feature space between normal and dysarthric speech may result in suboptimal speech reconstruction, thereby degrading speech intelligibility. To enhance the reconstruction ability of speech feature spaces, this paper proposes a DSR model named the Encoding-Aligned Variational Autoencoder (EA-VAE). By incorporating alignment modules of frame-level embedding features, prior distributions, and duration into the encoder of the VAE, the model explicitly aligns the dysarthric speech encoding with a representation of the parallel normal speech. A shared decoder is then used to generate speech with improved intelligibility. Experimental results on the UASpeech benchmark confirm that EA-VAE achieves state-of-the-art performance, with a 31.7% relative word error rate reduction and the highest subjective MOS score (4.48), thoroughly validating the effectiveness and advancements of the proposed method in dysarthric speech reconstruction.

Xinwei Wu, Heng Liu, Xiaohu Zhao, Yuqi Ren, Linlong Xu, Longyue Wang, Deyi Xiong, Weihua Luo, Kaifu Zhang

Large Language Models (LLMs) frequently exhibit strong translation abilities, even without task-specific fine-tuning. However, the internal mechanisms governing this innate capability remain largely opaque. To demystify this process, we leverage Sparse Autoencoders (SAEs) and introduce a novel framework for identifying task-specific features. Our method first recalls features that are frequently co-activated on translation inputs and then filters them for functional coherence using a PCA-based consistency metric. This framework successfully isolates a small set of "translation initiation" features. Causal interventions demonstrate that amplifying these features steers the model towards correct translation, while ablating them induces hallucinations and off-task outputs, confirming they represent a core component of the model's innate translation competency. Moving from analysis to application, we leverage this mechanistic insight to propose a new data selection strategy for efficient fine-tuning. Specifically, we prioritize training on "mechanistically hard" samples—those that fail to naturally activate the translation initiation features. Experiments show this approach significantly improves data efficiency and suppresses hallucinations. Furthermore, we find these mechanisms are transferable to larger models of the same family. Our work not only decodes a core component of the translation mechanism in LLMs but also provides a blueprint for using internal model mechanism to create more robust and efficient models.

Zhouxing Tan, Hanlin Xue, Yulong Wan, Ruochong Xiong, Xu Chu, Xiang Li, Junfei Liu

Large language models (LLMs) suffer from a lack of decision-making transparency, limiting their deployment in high-stakes domains such as healthcare. We propose a mechanistic interpretability framework that introduces two novel paradigms: Medical Fine-Tuning with Frozen Attention Layers (FTFA) and Posterior Adaptation Transcoders (PAT). FTFA freezes attention layers while fine-tuning only feed-forward network (FFN) parameters, enabling PAT to efficiently adapt pre-trained transcoders on the same data. This approach achieves over 1000× efficiency improvement compared to training transcoders from scratch. We theoretically justify this methodology and demonstrate its cost-effectiveness for cross-domain transfer. Transcoders are sparse autoencoders that replace MLP layers to provide interpretable feature representations. By substituting MLP layers of both base Gemma2-2b and its medical fine-tuned variant with per-layer transcoders, we enable feature-level attribution analysis. Through systematic pruning and node merging of resulting attribution graphs, we construct human-interpretable decision pathways. Our analysis reveals that LLMs employ two parallel mechanisms for medical diagnosis: pattern matching and multi-hop reasoning, with fine-tuned models demonstrating enhanced correct reasoning patterns. This work provides a practical framework for training transcoders on fine-tuned models at minimal cost, enabling broader application of mechanistic interpretability across domains and potentially guiding model training through transcoder-based analysis.

Ruike Song, Zeen Song, Huijie Guo, Wenwen Qiang

External reasoning systems combine language models with process reward models (PRMs) to select high-quality reasoning paths for complex tasks such as mathematical problem solving. However, these systems are prone to reward hacking, where high-scoring but logically incorrect paths are assigned high scores by the PRMs, leading to incorrect answers. From a causal inference perspective, we attribute this phenomenon primarily to the presence of confounding semantic features. To address it, we propose Causal Reward Adjustment (CRA), a method that mitigates reward hacking by estimating the true reward of a reasoning path. CRA trains sparse autoencoders on the PRM’s internal activations to recover interpretable features, then corrects confounding by using backdoor adjustment. Experiments on math solving datasets demonstrate that CRA mitigates reward hacking and improves final accuracy, without modifying the policy model or retraining PRM.

Dengcan Liu, Jiahao Li, Zheren Fu, Yi Tu, Jiajun Li, Zhendong Mao, Yongdong Zhang

Reward models (RMs) are a core component in the post-training of large language models (LLMs), serving as proxies for human preference evaluation and guiding model alignment. However, training reliable RMs under limited resources remains challenging due to the reliance on large-scale preference annotations and the high cost of fine-tuning LLMs. To address this, we propose SparseRM, which leverages Sparse Autoencoder (SAE) to extract preference-relevant information encoded in model representations, enabling the construction of a lightweight and interpretable reward model. SparseRM first employs SAE to decompose LLM representations into interpretable directions that capture preference-relevant features. The representations are then projected onto these directions to compute alignment scores, which quantify the strength of each preference feature in the representations. A simple reward head aggregates these scores to predict preference scores. Experiments on three preference modeling tasks show that SparseRM achieves superior performance over most mainstream RMs while using less than 1% of trainable parameters. Moreover, it integrates seamlessly into downstream alignment pipelines, highlighting its potential for efficient alignment.

Andrey V. Galichin, Alexey Dontsov, Polina Druzhinina, Anton Razzhigaev, Oleg Rogov, Elena Tutubalina, Ivan Oseledets

Recent LLMs like DeepSeek-R1 have demonstrated state-of-the-art performance by integrating deep thinking and complex reasoning during generation. However, the internal mechanisms behind these reasoning processes remain unexplored. We observe reasoning LLMs consistently use vocabulary associated with human reasoning processes. We hypothesize these words correspond to specific reasoning moments within the models' internal mechanisms. To test this hypothesis, we employ Sparse Autoencoders (SAEs), a technique for sparse decomposition of neural network activations into human-interpretable features. We introduce ReasonScore, an automatic metric to identify active SAE features during these reasoning moments. We perform manual and automatic interpretation of the features detected by our metric, and find those with activation patterns matching uncertainty, exploratory thinking, and reflection. Through steering experiments, we demonstrate that amplifying these features increases performance on reasoning-intensive benchmarks (+2.2%) while producing longer reasoning traces (+20.5%). Using the model diffing technique, we provide evidence that these features are present only in models with reasoning capabilities. Our work provides the first step towards a mechanistic understanding of reasoning in LLMs.

Xi Chen, Aske Plaat, Niki Van Stein

Chain‑of‑thought (CoT) prompting boosts Large Language Models accuracy on multi‑step tasks, yet whether the generated ``thoughts'' reflect the true internal reasoning process is unresolved. We present the first feature‑level causal study of CoT faithfulness. Combining sparse autoencoders with activation patching, we extract monosemantic features from Pythia‑70M and Pythia‑2.8B while they tackle GSM8K math problems under CoT and plain (noCoT) prompting. Swapping a small set of CoT‑reasoning features into a noCoT run raises answer log‑probabilities significantly in the 2.8B model, but has no reliable effect in 70M, revealing a clear contrast for these two scales. CoT also leads to significantly higher activation sparsity and feature interpretability scores in the larger model, signalling more modular internal computation. For example, the model's confidence in generating correct answers improves from 1.2 to 4.3. We introduce patch‑curves and random‑feature patching baselines, showing that useful CoT information is not only present in the top-K patches but widely distributed. Overall, our results indicate that CoT can induce more interpretable internal structures in high-capacity LLMs, validating its role as a structured prompting method.

Haozhuo Zheng, Cheng Wang, Yang Liu

The de novo generation of molecules with desirable properties is a critical challenge, where diffusion models are computationally intensive and autoregressive models struggle with error propagation. In this work, we introduce the Graph VQ-Transformer (GVT), a two-stage generative framework that achieves both high accuracy and efficiency. The core of our approach is a novel Graph Vector Quantized Variational Autoencoder (VQ-VAE) that compresses molecular graphs into high-fidelity discrete latent sequences. By synergistically combining a Graph Transformer with canonical Reverse Cuthill-McKee (RCM) node ordering and Rotary Positional Embeddings (RoPE), our VQ-VAE achieves near-perfect reconstruction rates. An autoregressive Transformer is then trained on these discrete latents, effectively converting graph generation into a well-structured sequence modeling problem. Crucially, this mapping of complex graphs to high-fidelity discrete sequences bridges molecular design with the powerful paradigm of large-scale sequence modeling, unlocking potential synergies with Large Language Models (LLMs). Extensive experiments show that GVT achieves state-of-the-art or highly competitive performance across major benchmarks like ZINC250k, MOSES, and GuacaMol, and notably outperforms leading diffusion models on key distribution similarity metrics such as FCD and KL Divergence. With its superior performance, efficiency, and architectural novelty, GVT not only presents a compelling alternative to diffusion models but also establishes a strong new baseline for the field, paving the way for future research in discrete latent-space molecular generation.

Huiliang Zhai, Xiangyi Teng, Jing Liu

Multiplex graphs are widely used to model multi-relational complex systems and play an important role in various real-world scenarios, such as financial systems and social networks. Hence, detecting anomalous samples in multiplex graph becomes crucial to ensure cybersecurity and stability. Although existing homogeneous graph anomaly detection (GAD) methods can be applied to deal with multiplex graphs, they still face two major challenges: 1) Due to the multiplicity and complexity of relations in multiplex graphs, homogeneous GAD models fail to effectively capture anomalous behaviors that correlate with diverse relational patterns. 2) In real-world applications, malicious entities usually disguise themselves through various camouflage strategies, making it difficult to capture subtle anomalous features via single-relation analysis. To address these challenges, we propose a novel unsupervised anomaly detection method for multiplex graphs based on Similarity-constrained Fusion Graph Autoencoder (SFGA). In SFGA, we design a multiplex graph autoencoder and introduced a cross-plex attention module at the model bottleneck to achieve comprehensive modeling of cross-relation anomaly patterns. Then, a similarity balancing strategy is proposed to constrain node representations at the bottleneck from both local and global perspectives, enhancing the discriminative power against camouflaged anomalies of autoencoder and enabling more effective identification of anomalous nodes with overlapping or deceptive patterns. Extensive experiments are conducted on both synthetic and real-world datasets at varying scales, and the results demonstrate that our proposed method outperforms state-of-the-art approaches by a large margin.

Shangshang Yang, Xuewen Duan, Xiaoshan Yu, Ziwen Wang, Haiping Ma, Xingyi Zhang

Cognitive diagnosis (CD), inferring student knowledge mastery based on historical response records, is crucial for personalized educational services such as adaptive practice and learning path planning. Existing CD models were built based on the assumption that student's response data is integral, overlooking the nonrandom missingness of data caused by student answering exercises selectively. This missingness generally leads to biased and incomplete observations, where confounders, such as selection bias and exposure bias, significantly undermine the accuracy of student knowledge modeling. To address missingness, we propose a Debiased Cognitive Diagnosis (DBCD) framework through the perspective of counterfactual modeling to remove exogenous confounders from the response data. Specifically, the proposed DBCD achieves debiasing for CD by applying the idea of contrastive learning to constrain the model's prediction distributions on both factual and counterfactual data. For a student, the factual data is his/her original response records, while the counterfactual data is generated by sampling the same number of exercises from all exercises of each concept through a similarity-based counterfactual sampling strategy. Considering the difficulty of directly removing the exogenous confounders for student, we devise a β-Variational Autoencoder to model their exogenous confounders within the latent representations of knowledge proficiency by leveraging exercise priors and student response patterns. Then, the learned representations are further combined with the vanilla student's ability embedding via a gating mechanism-based fusion for final diagnosis prediction of the model. Extensive experiments on real-world educational datasets demonstrate that the proposed DBCD effectively mitigates confounders and even outperforms existing methods, thereby validating the feasibility and effectiveness of the DBCD framework.

Yinghui Xing, Xiaoting Su, Shizhou Zhang, Donghao Chu, Di Xu

Infrared imaging plays a critical role in low-light and adverse weather conditions. However, due to the distinct characteristics of infrared images, existing foundation models such as Masked Autoencoder (MAE) trained on visible data perform suboptimal in infrared image interpretation tasks. To bridge this gap, an infrared foundation model known as InfMAE was developed and pre-trained on large-scale infrared datasets. Despite its effectiveness, InfMAE still faces several limitations, including the omission of informative tokens, insufficient modeling of global associations, and neglect of non-uniform noise. In this paper, we propose a Dual-domain Guided Infrared foundation model based on MAE (DuGI-MAE). First, we design a deterministic masking strategy based on token entropy, preserving only high-entropy tokens for reconstruction to enhance informativeness. Next, we introduce a Dual-Domain Guidance (DDG) module, which simultaneously captures global token relationships and adaptively filters non-uniform background noise commonly present in infrared imagery. To facilitate large-scale pretraining, we construct Inf-590K, a comprehensive infrared image dataset encompassing diverse scenes, various target types, and multiple spatial resolutions. Pretrained on Inf-590K, DuGI-MAE demonstrates strong generalization capabilities across various downstream tasks, including infrared object detection, semantic segmentation, and small target detection. Experimental results validate the superiority of the proposed method over both supervised and self-supervised comparison methods.

Jiezhong Wu, Jack O'Brien, Jennifer Li, M. S. Krafczyk, Ved G. Shah, Amanda R. Wasserman, Daniel W. Apley, Gautham Narayan, Noelle I. Samia

The discovery rate of optical transients will explode to 10 million public alerts per night once the Vera C. Rubin Observatory’s Legacy Survey of Space and Time comes online, overwhelming the traditional physics-based inference pipelines. A continuous-time forecasting AI model is of interest because it can deliver millisecond-scale inference for thousands of objects per day, whereas legacy MCMC codes need hours per object. In this paper, we propose SELDON, a new continuous-time variational autoencoder for panels of sparse and irregularly time-sampled (gappy) astrophysical light curves that are nonstationary, heteroscedastic, and inherently dependent. SELDON combines a masked GRU-ODE encoder with a latent neural ODE propagator and an interpretable Gaussian-basis decoder. The encoder learns to summarize panels of imbalanced and correlated data even when only a handful of points are observed. The neural ODE then integrates this hidden state forward in continuous time, extrapolating to future unseen epochs. This extrapolated time series is further encoded by deep sets to a latent distribution that is decoded to a weighted sum of Gaussian basis functions, the parameters of which are physically meaningful. Such parameters (e.g., rise time, decay rate, peak flux) directly drive downstream prioritization of spectroscopic follow-up for astrophysical surveys. Beyond astronomy, the architecture of SELDON offers a generic recipe for interpretable and continuous-time sequence modeling in any time domain where data are multivariate, sparse, heteroscedastic, and irregularly spaced.

Yiming Wang, Kaiyan Zhao, Xu Li, Yan Li, Jiayu Chen, Steven Morad, Leong Hou U

Control in high-dimensional action spaces remains a fundamental challenge in reinforcement learning (RL), primarily due to inefficient exploration of the action space. While recent methods attempt to guide exploration, they often fall short of achieving the agility and coordination exhibited in biological motor control. Inspired by how organisms exploit muscle synergies for efficient movement, we propose Explore to Learn (ETL), a two-stage framework that first discovers fundamental synergy patterns and then leverages them for task-specific policy learning. In the first stage, ETL discovers underlying synergy patterns by deploying a targeted exploration policy. These patterns are modeled as latent directions in a low-dimensional space, along which the agent is guided to collect diverse and structured muscle activation trajectories. A variational autoencoder (VAE) is then trained to encode high-dimensional actions into a latent space whose dimensions correspond to the synergy patterns. In the second stage, the policy is trained entirely in this synergy-aware latent space, producing synergy coefficients that the decoder maps back to full-dimensional muscle actions. This structured representation significantly reduces the complexity of learning, while the decoder is further fine-tuned to enhance expressiveness and generalization across downstream tasks. Extensive experiments across musculoskeletal environments and the DMControl suite demonstrate that ETL consistently outperforms prior methods in both exploration efficiency and control performance, achieving superior scalability and generalization in overactuated control tasks.

Md Shahriar Rahim Siddiqui, Moshe Eliasof, Eldad Haber

Flow matching casts sample generation as learning a continuous-time velocity field that transports noise to data. Existing flow matching networks typically predict each point's velocity independently, considering only its location and time along its flow trajectory, and ignoring neighboring points. However, this pointwise approach may overlook correlations between points along the generation trajectory that could enhance velocity predictions, thereby improving downstream generation quality. To address this, we propose Graph Flow Matching (GFM), a lightweight enhancement that decomposes the learned velocity into a reaction term -- any standard flow matching network -- and a diffusion term that aggregates neighbor information via a graph neural module. This reaction-diffusion formulation retains the scalability of deep flow models while enriching velocity predictions with local context, all at minimal additional computational cost. Operating in the latent space of a pretrained variational autoencoder, GFM consistently improves Fréchet Inception Distance (FID) and recall across five image generation benchmarks (LSUN Church, LSUN Bedroom, FFHQ, AFHQ-Cat, and CelebA-HQ at 256 × 256), demonstrating its effectiveness as a modular enhancement to existing flow matching architectures.