Embedding
-
Aligning MLLMs with the Latent Structure of Human Cognition via Behavior-Derived Semantic Dimensions
2026-10-30 16:06:08Ning E, Changde Du, Yizhuo Lu, Huiguang HeAbstract
Multimodal Large Language Models (MLLMs) have achieved impressive performance across a wide range of reasoning tasks, yet it remains unclear whether their internal representations and decision criteria align with the latent structure of human cognition. In this work, we propose a novel framework to align MLLMs with the multidimensional mental representations underlying human similarity judgements. We build on a previously established behavior-derived semantic embedding space, in which objects are represented along 66 interpretable dimensions derived from 4.70 million human Odd-One-Out (O1O) judgements. For each triplet, we infer the most salient decision dimension and convert it into a dimension-grounded linguistic rationale. This enables MLLMs to learn both the final choice and a justification consistent with the inferred behavioral criterion. Experimental results demonstrate that our approach improves model consistency with human similarity judgements while retaining competitive performance on standard multimodal benchmarks. Furthermore, using searchlight Representational Similarity Analysis (RSA) on an independent fMRI Natural Scenes Dataset (NSD), we observe increased representational alignment between model and human. Ultimately, our findings point to a data-driven route for incorporating human cognitive structure into MLLMs alignment. The code is available \href{this https URL}{here}.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41951
PDF
-
Automated Keyword Generation Using LLMs in a High-Volume Data Virtualization Service
2026-10-30 15:34:59Saiful Khan, Yiwen Xing, Joyraj Chakraborty, Afrah Farea, Philip Beaucamp, Samuel Jackson, Min ChenAbstract
In modern data infrastructure, a data virtualization service (DVS) allows users to create many versions of source datasets for different processing tasks. For example, a multivariate or multi-modality dataset can be used for different machine learning tasks (e.g., classification, detection, segmentation, forecasting, and so on). Data selection, participation, integration, and preprocessing in different tasks result in many virtual datasets [Khan et al., SSC 2026]. Meanwhile, different users need to discover existing virtual datasets frequently and easily to avoid duplicate efforts. To enable an effective search engine in a DVS, it is desirable to automatically generate keywords for each virtual dataset. In this work, we present a solution for generating keywords for virtual datasets using large language models (LLMs). We propose five strategies for selecting different collections of text data for processing by the LLMs. We applied the five strategies to virtual datasets created in five application case studies with different source datasets. We conducted experiments with three LLM generators and two embedding techniques. The experiments show that the performances of the five strategies depend on the quality of the text descriptions. The results help us adopt a multi-strategy implementation of a new keyword generation service (KGS) as part of an extended DVS. The work demonstrates that the capability of automated keyword generation is indispensable for a high-volume DVS that is dataset-rich and frequently accessed.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41893
PDF
-
Text as an Index for Zero-Shot Cross-Domain Knowledge Tracing
2026-10-30 15:29:08Yin Yi, Haiping Ma, Siyu Song, Ziwen Wang, Shangshang YangAbstract
Knowledge tracing (KT) predicts a student's next response from their history. New platforms often have question texts but no response logs for training, motivating zero-shot cross-domain KT. Existing transfer and cold-start methods use question semantics, but cross-platform semantic discrepancies limit transferability. We propose TIZSKT, which separates behavioral sequence modeling from domain-local historical evidence construction and uses text as an index rather than a directly transferable predictive representation. The Mastery-Dynamics Encoder learns correctness dynamics without question identifiers or text embeddings. Domain-Local Question Addressing builds fixed question addresses using a frozen text encoder and an unsupervised per-domain projection. The Text-Addressing Kernel uses these addresses to aggregate the current student's previous outcomes into a question-specific historical correctness summary. Cross-Segment Continuation retains access to earlier responses without extending the behavioral encoder's input sequence. Evidence-Volume Read-Out preserves the total kernel mass discarded by normalization and combines its log-compressed value with the behavioral state and historical correctness summary to predict the next response. All trainable parameters are learned from source domains; target inference uses public question texts and the current student's observed history without target-domain training. Leave-one-domain-out experiments on three public datasets show that TIZSKT achieves the highest mean AUC and accuracy among evaluated baselines on all target domains. Ablations confirm the contributions of text addressing and cross-segment continuation.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41883
PDF
-
Self-Supervised Object-Centric Representation for Multi-Task Audio Analysis
2026-10-30 15:20:22Parthasaarathy Sudarsanam, Tuomas VirtanenAbstract
This paper studies the capability of an object-centric audio representation to support multiple downstream tasks within a unified framework. The object-centric model is trained in a self-supervised setup using slot attention by reconstructing Audio Spectrogram Transformer (AST) patch embeddings from a set of latent slots. It is trained on synthetic audio mixtures, which provides controlled multi-source audio scenes. These slot representations are evaluated for the classification, detection, and separation of sound events by training task specific output layers. On the classification task, the model achieves an F1-score of 0.8, comparable to clip-level AST embeddings which achieve 0.7. For sound event detection, it reaches a macro F1-score of 0.7, outperforming clip-level AST embeddings at 0.4 and matching AST patch embeddings at 0.7, which encode finer time-frequency information. In source separation, it achieves an SI-SDR improvement (SI-SDRi) of 18.4 dB. Furthermore, in a zero-shot source separation setting, the model achieves an SI-SDRi of 14.5 dB. These results demonstrate that slot-based representations capture transferable and compositional structure in audio scenes, enabling generalization across tasks and datasets.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41870
PDF
-
Iterated Integrals as Lexical Fingerprints: Path Signature Embeddings for Lexical Similarity
2026-10-30 15:11:18Raj Shekhar Singh, Pratham AgarwalAbstract
While sentence embeddings are primarily designed for semantic generalization, their latent geometry implicitly encodes surface-level lexical properties, a relationship that remains poorly understood. Shallow pooling methods discard word order, while deep Transformer models obscure these lexical alignments behind computational complexity and uninterpretable spaces. We propose constructing sentence embeddings via the path signature of Word2Vec trajectories, leveraging the iterated integrals of rough path theory to create representations that explicitly capture sequential and lexical structure in a geometrically tractable manner. Through extensive ablations spanning Word2Vec dimensions, signature depths, and PCA compression, we correlate embedding distances against five lexical string metrics: Levenshtein, Jaro-Winkler, word Jaccard, phoneme Levenshtein, and character TF-IDF. We reveal that while average pooling almost entirely fails to capture character-level edit distances, path signatures strongly align with them, with depth 2 signatures optimally balancing character and word-level sensitivities. Furthermore, we demonstrate that the critical lexical variance is concentrated in a highly compact PCA subspace, and unlike pooling, signatures structurally preserve and quantify word order. These findings establish path signatures not as semantic generalizers, but as rigorous mathematical lenses that explicitly encode discrete lexical structures and word order, offering an interpretable alternative to opaque deep learning for lexical tracking.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41855
PDF
-
Hard Sample-Conditioned Decorrelated Attention for Robust Face Recognition
2026-10-30 14:21:43Tahar Chettaoui, Guray Ozgur, Eduarda Caldeira, Naser Damer, Fadi BoutrosAbstract
Multi-head attention in Vision Transformers (ViTs) allows independent attention patterns across heads, but without explicit regularization, heads may learn redundant patterns rather than complementary roles. Correlation-based penalties encourage head disentanglement, yet apply uniform regularization across all training samples. In face recognition (FR), hard samples whose embeddings lie close to competing identities are critical for optimization and generalization. Motivated by this, we propose ViT-HarDis, which conditions attention head disentanglement on sample difficulty: samples with high cosine similarity to negative class centers receive stronger head decorrelation, enforcing specialization where identity ambiguity is highest. Experiments show improvements over both the baseline and uniform disentanglement on most benchmarks. Our analysis shows that ViT-HarDis reduces the similarity of hard samples to competing class centers while decreasing inter-head attention similarity, and that applying disentanglement only to the final transformer layers outperforms regularizing the entire network. Trained on WebFace4M, ViT-HarDis achieves SOTA results among ViT-based methods on IJB-B, IJB-C, and TinyFace Rank-1, while achieving the best IJB-B and TinyFace results when trained on MS1MV3, improving FR without architectural modifications or additional inference cost. Code: this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41789
PDF
-
Peeling the Matryoshka: Extracting Ontological Structure from Retrieval Embeddings
2026-10-30 14:19:53Eunchan Kim, Inwhee JoeAbstract
Taxonomy construction is essential for knowledge-intensive NLP yet remains costly: clustering-based methods (i) need careful per-corpus tuning and yield unstable hierarchies, while LLM-based pipelines (ii) depend on expensive task-specific training that does not transfer across domains. We report an empirical finding that bears on both: in Matryoshka Representation Learning (MRL) embeddings, where each leading prefix is itself a usable, coarser vector, prefix length aligns with the coarse-to-fine granularity of gold taxonomies; shuffled- and reverse-dimension controls isolate the alignment to the learned early prefix dimensions. The pattern holds across three architecturally distinct MRL models. To test whether this regularity is usable, we introduce PEEL (Peeling Embeddings to Extract Latent ontologies): a minimal probe over these frozen embeddings, trained once with an order-embedding objective that represents hierarchy as vector containment; a minimum spanning arborescence then decodes taxonomic DAGs. Using only a concept vocabulary (no documents), PEEL recovers taxonomic DAGs with high structural fidelity on two OLLM benchmarks, transferring to a new domain with no target-edge supervision. The end-to-end OLLM pipeline, which jointly discovers concepts and induces structure from document corpora, is reported as a cross-task reference. Pre-trained retrieval embeddings encode a decodable granularity gradient; a minimal probe renders it as explicit parent-child (is-a) structure.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41785
PDF
-
From Fragmentation to Fusion: A Survey and Roadmap for Omnimodal Embeddings
2026-10-30 14:07:34Ziyin Zhang, Yuyu Guo, Ying Zhang, Siqiao Xue, Zihan Liao, Jin Qin, Wenjie Yang, Siyuan Yang, Shijie Li, Cheng Chen, Hao Wu, Zhaoyang Zhu, Ruipeng Li, Zhaokun Jiang, Lei Lei, Rui Wang, Hang Yu, Rui WangAbstract
Embedding models underpin virtually every modern AI system, from search engines to recommendation systems. Despite shared mathematical foundations, research on embedding models remains fragmented by modality: the communities develop largely in isolation, seldom exchanging insights. We argue that this fragmentation obscures a deeper convergence. Driven by contrastive learning, Transformer architectures, and a granularity shift, embedding research across modalities is converging toward a common set of principles, even as modality-specific challenges remain underexplored. To reason across this convergence without conflating it, we propose a framework of three progressively harder capabilities: coverage, comparability, and compositionality. Three paths toward unification satisfy these in turn: transformation broadens coverage, alignment enables comparability, and joint embedding, the terminus we advocate, aims at compositionality. This position paper traces the resulting evolution, providing the first unified, cross-cutting treatment of omnimodal embedding research.
Abstract (translated)
URL
https://arxiv.org/abs/2611.00172
PDF
-
Latent-Lens: Visual Perception in Small Language Models Through Latent Communication
2026-10-30 13:53:44Mo Shahdloo, Faegheh Sardari, Mathew SalvarisAbstract
Bridging independent vision-language models (VLMs) and small language models (SLMs) in a modular multi-agent system typically forces them to exchange information as discrete text, which creates a lossy bottleneck for spatial and fine-grained details. This constraint is especially acute for SLMs, whose limited capacity makes them more reliant on the quality of the input signal. Recent latent-communication methods exchange states between language models. How a VLM can pass visual information to a separately pretrained text-only SLM though remains unexplored. We propose \textbf{Latent-Lens}, a lightweight codec that bridges the representation gap between a frozen VLM and a frozen SLM directly in latent space, enabling them to communicate purely through hidden states. The codec consists of a domain encoder that projects hidden states from the VLM into the SLM's embedding space, and an attention decoder formalised as a LoRA that teaches the SLM to attend to these visual prefixes, while training only ${\sim}1.6$M parameters (0.2\% of the combined model size). On visual question answering (CLEVR, GQA), Latent-Lens achieves $81.0\pm7.9$\% and $65.4\pm5.4$\%, compared with $58.3\pm5.7$\% and $61.8\pm1.0$\% for text communication with the same LoRA budget (+30.2--69.8pp relative to frozen text communication), while on image captioning (Flickr8k) it improves over the same baseline by +16.5 (+18.8 frozen). We also evaluate Latent-Lens with TinyLlama-1.1B and InternVL2-1B, which use different architectures, vocabularies, and hidden dimensions.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41751
PDF
-
Task Vectors Can Be Linearly Decomposed into Interpretable and Controllable Components
2026-10-30 13:48:48Yukiya Horiba, Hakaze Cho, Takahiro YakohAbstract
Task Vectors (TVs) provide a compact way to steer large language models, but they are typically treated as atomized directions merely extracted from hidden states, leaving their internal functional structure poorly understood. Motivated by prior work that decomposes in-context learning into Task Recognition (TR) and Task Learning (TL), we ask whether a single TV can likewise be decomposed into functionally meaningful components. We formulate a TR/TL-based decomposition from candidate-label unembeddings, separating each TV into TR, TL, and residual components, and try to optimize the TR and TL coefficients using Bayesian optimization for more effective intervention. We evaluate this formulation across multiple model families, classification tasks, and both instruction- and demonstration-based TVs. Across these settings, TR/TL-based optimization identifies intervention directions that outperform the original TVs. Although optimization over random directions can also improve performance, TR/TL-based search yields more reliable gains, particularly for instruction-based TVs. Moreover, isolated interventions on the decomposed components produce distinct logit-level effects that are consistent with their intended functional roles. Together, these results suggest that TVs are not monolithic steering directions, but can be decomposed into functionally meaningful components that support both interpretation and control.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41743
PDF
-
Advancing Clinical Trial Outcome Prediction with Domain-Specific Large Language Models
2026-10-30 13:17:00Burak Bakkar, Hana Moalla, Mohamed Salem Jedidi, Haifa Ben Messaoud, Nicolas Lopez CarranzaAbstract
Clinical trial failures are a major barrier to efficient drug development, incurring substantial costs and delaying patient access to effective therapies. Structured approaches such as the Hierarchical Interaction Network (HINT) predict trial outcomes by explicitly modelling interactions between drugs, diseases and protocols, but make limited use of the free text in trial registrations. We evaluate two domain-specific language models, BioBERT (108M parameters) and BioMistral (7B parameters), for phase-level outcome prediction on the TOP benchmark (12,477 trials), representing each trial as text built from ten this http URL fields. We compare fine-tuning (full fine-tuning for BioBERT, low-rank adaptation for BioMistral) with probing, in which the pretrained model is frozen and only a lightweight classifier is trained on its embeddings. On the held-out test set of each phase, all four language-model variants achieved higher ROC-AUC (0.613--0.772) and PR-AUC (0.649--0.901) than the published HINT results, corresponding to relative PR-AUC gains of 10--20\%. F1 differences were small and mixed, and no model, including HINT, clearly exceeded the F1 of a trivial classifier that predicts success for every trial. Under the configurations tested, probing performed comparably to fine-tuning: fine-tuned models scored highest on ROC-AUC and PR-AUC in Phases I and II, and frozen BioMistral embeddings in Phase III (ROC-AUC 0.772; PR-AUC 0.901), although most differences between variants were below 0.03 and were not tested for significance. These results suggest that pretrained biomedical representations already encode signal relevant to trial outcome, and that probing is a computationally cheaper alternative to fine-tuning for this task.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41710
PDF
-
Connectivity-Aware Client Selection in Decentralized Federated Learning
2026-10-30 11:01:58Usevalad Milasheuski, Monica Nicoli, Stefano SavazziAbstract
Decentralized federated learning (DFL) has emerged as a scalable alternative to centralized federated learning, removing the single point of failure and communication bottleneck at the server. However, most existing decentralized approaches rely on full-neighborhood broadcasting, where each client transmits its model update to all of its neighbors in each communication round. While effective, this results in significant communication overhead and high energy consumption. This paper applies spectral graph tools as lightweight client selection methods that leverage the graph topology by analyzing the spectral properties of the graph Laplacian over a fixed communication graph. Specifically, we exploit eigenvector embeddings and heat/diffusion kernel similarities to guide each client in selecting only a subset of structurally relevant neighbors, i.e., those that are most beneficial for the convergence of the federation process. This topology aware randomization matches or outperforms uniform neighborhood subsampling and approaches the accuracy of full broadcasting. The largest gains, about $3$ percentage points, emerge under strong data heterogeneity on larger, structurally rich graphs, while the total number of transmitted messages is reduced by $50\%$ compared to naive broadcasting.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41570
PDF
-
From the Zoo to the BULLPEN: Sample-Efficient LLM Profiling via Multi-Channel Model Embeddings
2026-10-30 07:41:44Russell Ro, Qiming Cui, Reza Abbasi-AslAbstract
Choosing among the hundreds of thousands of available large language models (LLMs) requires a thorough understanding of each candidate model's properties. In practice, each property is measured by its own benchmark suite, so profiling a model on many properties means paying for many separate evaluations. To address this, model embeddings compress each model into a fixed-length vector from which many properties can be predicted. Existing methods, such as EmbedLLM, build these vectors from correctness patterns and question features, and often need large probe batteries to do so. However, many behaviors that are important when choosing a model, such as its verbosity, safety and style, show up more in how it answers than in whether it is correct. We introduce BULLPEN (Behavioral Understanding of Large Language models via Profile ENcodings), a framework that extends EmbedLLM in three ways. First, BULLPEN organizes the downstream tasks that prior work uses and adds 20 task benchmarks we built on behavior, style and provenance. Second, it is the first framework to read a model's own answer text jointly with its correctness and the question text. Third, it encodes models from a few hundred questions instead of the full 24,118-question pool. Across 260 open-weight models, our answer-text encoder scores 0.48 +/- 0.10 on the new tasks against 0.35 +/- 0.10 for the best prior encoder. Our encoders that read answer text keep at least 91% of their full-pool skill from only 400 questions (1.3% of the inference compute of the full question pool), while a correctness-only baseline keeps 83% with the same 400 questions. Because these embeddings need only model outputs, they support behavioral auditing and low-cost profiling, and could extend to API-only models. Code and data: this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2610.41375
PDF
-
SciGraphSelect: Auditable Scientific Data Selection through Knowledge--Reasoning Graphs
2026-10-30 05:36:49MiHyeon Kim, Ji-Eun Han, Eunkyeong Lee, Jeehyun Lim, Hwijung RyuAbstract
Scientific data selection is typically driven by relevance or utility scores, providing limited insight into why a document is retained. We introduce SciGraphSelect, a reference-grounded framework for auditable scientific data selection. SciGraphSelect organizes domain-specific knowledge, reusable reasoning operations, and reference-attested relations into a typed knowledge--reasoning graph, then scores candidate documents using concept coverage, relation and path co-support, and directional alignment. To avoid over-interpreting document embeddings, rich relational structure is preserved on the reference side while candidate-side evidence is treated conservatively without assuming relation entailment. By exposing the reference-grounded evidence behind each selection, SciGraphSelect brings traceability to the upstream construction of scientific training corpora and complements downstream scientific verification.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41280
PDF
-
Differential rank preservation manifold learning
2026-10-30 00:56:05Henry Han, Luyu Niu, Diane Li, Jiacun WangAbstract
Manifold learning methods like t-SNE and UMAP preserve local neighborhoods, but they do not preserve the ranking of the neighbors inside them. Two embeddings can return the same neighbor set while disagreeing on which point is closest. This loss of order degrades downstream applications that value the order in the embedding. We propose Smooth Rank t-SNE (SR-tSNE), which adds a differentiable rank term to the standard t-SNE loss. To enable efficient training, this term uses a sigmoid rank approximation to explicitly force a point's true high-dimensional neighbors to outrank unrelated intruders and retain their original order. We also introduce Ordinal DLP (O-DLP), an extension of the Degree of Locality Preservation (DLP) metric that credits a neighbor only when it returns at its original rank. Across eight benchmarks spanning image, text, speech, remote-sensing and single-cell data, SR-tSNE outperforms 12 baselines in rank preservation and neighborhood overlap. Ablation studies demonstrate our rank term alone boosts O-DLP by 5% to 28%, a gain not achievable using standard pairwise loss. This advantage transfers to downstream applications. SR-tSNE achieves the best label-transfer performance across seven vision datasets. For cell-type discovery across six single-cell datasets, it ranks first in 17 of 18 dataset-metric combinations, whether clustering resolution is selected with or without cell-type labels. In intraday trading, SR-tSNE achieves higher returns, lower risk, and lower turnover than the baselines. To our knowledge, SR-tSNE is the first t-SNE-family method with explicit order preservation.
Abstract (translated)
URL
https://arxiv.org/abs/2610.41107
PDF
-
CAMEL: Cross-Architecture Masked Embedding Learning for Ultra-Lightweight Retinal Foundation Models
2026-10-29 20:33:32Andre Guarnier De Mitri, Giovani Decico Lucafo, Miller Matheus Lima Anacleto Rocha, Diego Furtado SilvaAbstract
Recent progress in deep learning has substantially advanced automated grading of ophthalmologic pathologies, supported by large annotated datasets and improved representation learning. However, a significant gap remains between state-of-the-art performance and clinical scalability. In particular, foundation models have reshaped image-based medical screening and diagnosis. These large-scale architectures trained on extensive datasets learn rich, generalizable feature representations that can be adapted to a wide range of downstream tasks. In ophthalmology, RETFound, one of the most relevant foundation models for retinal images, performs strongly across classification, segmentation, and risk prediction tasks. Despite their effectiveness, foundation models are often computationally prohibitive, hindering their deployment in real-world clinical environments where hardware constraints and latency are critical. To bridge this gap, we introduce CAMEL (Cross-Architecture Masked Embedding Learning), a novel cross-architecture distillation framework that uses an asymmetric masking strategy during embedding learning to effectively transfer knowledge from massive teacher foundation models to compact, lightweight student models. Our experimental evaluation across four datasets shows that CAMEL preserves accuracy while drastically reducing computational complexity, enabling lightweight yet highly accurate models suitable for resource-limited clinical settings.
Abstract (translated)
URL
https://arxiv.org/abs/2610.40954
PDF
-
CoFiBreak: A Coarse-to-Fine Diversity Reward for Effective Multi-Turn Jailbreak
2026-10-29 19:51:36Jialin Song, Xiaodong Liu, Weiwei Yang, Mingqian Feng, Chengquan Guo, Qiyao Ma, Carolyn Buractaon, Anson Ho, Wuyang Chen, Christopher WhiteAbstract
Automated red teaming trains an attacker large language model (LLM) to elicit harmful responses from a victim LLM, but reinforcement learning can collapse attackers onto a narrow set of prompts when attack success is the only reward. Such a collapse of attack diversity can bias safety evaluation toward repeatedly testing near-duplicate attack patterns rather than the underlying harmful intents. Existing methods often trade off attack success or inference cost for greater content or strategy diversity. We propose CoFiBreak, a coarse-to-fine diversity reward for Group Relative Policy Optimization (GRPO) that improves attack diversity while preserving strong attack effectiveness and low inference cost. The coarse reward broadens which strategies the attacker uses by encouraging strategies underrepresented within each GRPO group, while the fine reward broadens how those strategies are realized by discouraging reuse of the same personas, framings, and output formats after harmful-intent information is removed from prompt embeddings. Diversity rewards apply only to attacks that already exceed an effectiveness threshold, so diversity cannot be gained through ineffective attacks. Since both rewards are computed only during training, CoFiBreak also retains inference efficiency. Trained with Llama3.1-8B-Instruct as the victim, CoFiBreak discovers 1.6-2.0x as many distinct successful fine-grained strategies as the strongest baseline across frontier victims on HarmBench. It reaches 98.8% ASR@10 (attack success rate with 10 generations per intent) on AdvBench with the highest diversity score among all compared methods, which is 2.1x that of the strongest-attacking baseline. The resulting attacks transfer to unseen frontier LLMs, reaching 71.0% ASR@10 on AdvBench with the unseen victim Claude-Opus-4.6. The code will be released.
Abstract (translated)
URL
https://arxiv.org/abs/2610.40926
PDF
-
Sparse Autoencoders for Expert-Assisted Interpretation of Histopathology Foundation Model Embeddings
2026-10-29 19:26:15\v{S}imon \v{S}m\'ida, Wanda BenesovaAbstract
Histopathology foundation models learn rich embedding representations, but their internal structure remains difficult to inspect. Concept-level interpretability is particularly relevant in computational pathology because pathologists reason through recurring morphological and tissue-level patterns, including tissue architecture, cellular morphology, staining variation, and artifacts. In this work, we investigate whether sparse autoencoders (SAEs) can decompose embeddings from histopathology foundation models into sparse latent features suitable for expert-assisted interpretation. We train vanilla and Batch-TopK SAE variants on NCT-CRC embeddings, using CONCH as the primary foundation model and UNI2-h as an additional robustness evaluation, and characterize the resulting latent features through reconstruction fidelity, sparsity, coherence, label enrichment, and downstream class-signal preservation. The learned sparse representations preserve most downstream tissue-class information, and their latent features retrieve more coherent and label-concentrated patch groups than raw embedding dimensions or PCA components. Discussions with pathology experts further motivated a WSI-level latent visualization workflow that places feature activations into larger tissue context. Overall, the additional UNI2-h evaluation suggests that the observed behavior is not specific to CONCH, supporting the use of sparse autoencoders for expert-assisted analysis of histopathology foundation-model embeddings, while emphasizing that latent features should be treated as candidate concepts requiring pathological assessment rather than automatically validated histopathological concepts.
Abstract (translated)
URL
https://arxiv.org/abs/2610.40905
PDF
-
TAPE: Learning General Audio Effect Representations
2026-10-29 17:17:15Phillip Long, Yongyi Zang, Qihui Yang, Christodoulos Benetatos, Junfei Liu, Randal Leistikow, Julian McAuley, Perry R. CookAbstract
Musicians and mixing engineers often can decipher the signal processing chain they hear on a reference recording, then adapt it to new material. We present TAPE (Transformations of Audio Parameter Embeddings), which estimates this processing from the processed reference audio alone. An encoder trained to identify processors in eight major effect families maps the reference to an embedding. A differentiable digital signal processing (DSP) chain decodes the embedding into editable parameters in one forward pass and applies them to a different unprocessed recording. Its modules are equalization, compression, waveshaping, a small residual network, and reverberation. TAPE achieves lower spectral error than DeepAFx-ST, the previous state of the art for single-step regression, and lower spectral error than per-example search over a chain approximating ST-ITO's, while running 275$\times$ faster than that search. Because the decoder outputs DSP parameters, its chain can also run causally: a C++ implementation with the trained weights processed 128-sample buffers 1.8$\times$ faster than real time, with no missed deadlines.
Abstract (translated)
URL
https://arxiv.org/abs/2610.40759
PDF
-
Retrieving from Your Own Mistakes: On-Policy Training of Embedding Models for Multi-Hop Retrieval-Augmented Generation
2026-10-29 16:40:02Chenjie Qiu, Zixuan Zeng, Shaobin Shi, Bingfeng Chen, Dongsheng Luo, Boyan XuAbstract
Complex questions in retrieval-augmented generation require evidence spread over several documents, and an answer is reliable only when the retriever finds all of it. Chain embedding retrievers collect such evidence step by step, re-encoding the question together with the passages retrieved so far, and we find that they are already strong: trained only on gold histories, a small chain embedding retriever that calls no language model outperforms graph-based systems and is competitive with a search agent. This training, however, leaves a gap. At test time the retriever conditions on its own retrievals, which soon contain mistakes: by the third step, about half or more of the states it visits hold a wrong passage, and these passages lower its chance of retrieving a missing gold passage well beyond what the difficulty of the states explains. We propose on-policy state training, which trains the encoder on the states it reaches itself. At every epoch the current encoder retrieves chains for the training questions; each state it visits, including those built on its own mistakes, is labeled with the earliest gold passage still missing while the other missing gold passages are masked, so that no correct continuation is penalized; and the encoder is updated contrastively on these states mixed with teacher-forced ones, so that its training distribution follows it as it learns. The resulting retriever retrieves complete evidence more often than teacher-forced training and the systems we compare, and it transfers without further training to harder long-chain variants of HotpotQA and 2WikiMultiHopQA that we construct. On the same states, it recovers from its own mistakes far more often, and controlled comparisons attribute the gain to training on the states the encoder currently reaches.
Abstract (translated)
URL
https://arxiv.org/abs/2610.40409
PDF
