Embedding
-
Selecting What Matters: Semantic Compression-Guided Selective Pooling for Long-Context Embeddings
2026-09-29 15:16:48Zifeng Cheng, Jie Zheng, Zhiwei Jiang, Shuwen Wang, Fei Shen, Shiping Ge, Qing GuAbstract
Large language models (LLMs) have shown strong potential as training-free text encoders for long-context embeddings. Existing approaches primarily improve information flow under causal attention and typically construct embeddings by uniformly averaging all token representations. However, for long documents, such mean pooling can dilute salient semantic information with abundant redundant or weakly informative content. To this end, we propose SCSP, a training-free framework that leverages semantic compression for informative token selection in long-context embedding. Specifically, SCSP first partitions a document into sentence-aware chunks and appends a semantic compression prompt to each chunk. A prompt-isolated attention mask preserves information flow among document tokens while restricting each prompt to its corresponding local context. We then use the attention patterns elicited by these prompts to estimate token importance, select informative tokens, and aggregate their intermediate-layer representations into the final embedding. Extensive experiments on long-context embedding benchmarks demonstrate that SCSP can be integrated into both zero-shot and fine-tuned models in a plug-and-play manner, consistently improving their performance.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37782
PDF
-
Spatiotemporal Hyperedges for EEG Seizure Detection and Prediction
2026-09-29 14:53:14Hyunju Kim, Sheo Yon Jhin, Noseong Park, Nabil ImamAbstract
Seizure detection and prediction from EEG are clinically important but challenging because seizures are rare, temporally localized, and propagate as coordinated events across multiple channels. Recent dynamic graph neural networks model this by running a temporal model over a sequence of per-time-step pairwise channel edges. However, this pairwise construction misses the spatiotemporal coupling that constitutes a seizure, at substantial training cost. We propose HyBrain, which summarizes spatiotemporal EEG evidence through a small set of soft hyperedges rather than pairwise edges. A per-channel Mamba backbone produces one token per (channel, second), and a spatiotemporal hyperedge block pools these tokens into E_h shared group embeddings through soft memberships and broadcasts them back. The same encoder serves three downstream tasks: window-based detection, one-second point-wise detection, and preictal seizure prediction. On TUSZ and CHB-MIT, HyBrain achieves the best AUROC on every reported setting against ten baselines, with the largest gap on long-clip preictal prediction. It also matches the most efficient baselines in training time and peak GPU memory. A qualitative analysis shows that even a single learned hyperedge cleanly captures the preictal -> ictal -> postictal trajectory on a real seizure clip.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37730
PDF
-
Beyond Selection: Token Parameterization for Extreme Visual Token Compression
2026-09-29 14:19:02Rui Zhong, Yu Li, Zheyu Yan, Cheng ZhuoAbstract
Visual-token compression is effective for improving the efficiency of vision-language models, but under extreme compression budgets, token pruning can break visual grounding while learned resamplers increase parameter count, attention cost, and training complexity. We revisit compression through a token parameterization lens, separating (i) basis transformation and structured truncation (retained subspace/compressibility) from (ii) coordinate organization (optimization and cross-modal alignment). This view yields two coupled objectives, compressibility and learnability, which we formalize as unified functionals. Guided by these objectives, we design Braco, a lightweight four-step coder that combines transform-basis truncation, input-independent basis-coordinate embeddings, budget-dependent orthogonal re-parameterization, and learned spatial residual tokens from lightweight pooling. Experiments show that Braco forms the favorable empirical accuracy-efficiency frontier under $23\times$--$64\times$ compression and remains competitive at $144\times$, reaching 95.2% accuracy while reducing prefill FLOPs by 84.2%--86.7% relative to the uncompressed upper bound. Against prior methods, Braco matches or improves accuracy while achieving up to approximately 36% end-to-end speedup and using $16.6\times$/$78.8\times$ lower compressor latency/FLOPs.
Abstract (translated)
URL
https://arxiv.org/abs/2609.35232
PDF
-
Flattening the Connectome Spectrum: A Spectral Filter for FC Induces a Pretraining Target for fMRI Encoders
2026-09-29 14:12:01Giovanni Marraffini (UNITO), Victoria Shevchenko (UNITO), Carlo Alberto Barbano (UNITO), Demian Wassermann (MIND)Abstract
Self-supervised pretraining reshaped prediction in language and vision, and brain foundation models (BFMs) inherited its promise. Representations learned from large unlabelled corpora should capture individual functional dynamics and generalise across cohorts. However, kernel ridge regression (KRR) fitted on functional connectivity (FC) matrices still predicts individual phenotypes more accurately than any BFM we tested. In this paper, we show that KRR is weighted by the eigenvalues of the FC which are miscalibrated for phenotype prediction. We apply an efficient spectral filter to recalibrate the eigenvalues of each subject's FC matrix, enabling the model to exploit more inter-individual variance. Across the 5 datasets, 11 parcellations and 6 prediction targets we tested, we match or exceed the KRR baseline. Based on this finding, we then pretrain a small encoder model on about 4,000 hours of fMRI from 162 open datasets, whereby we align the pairwise similarities between the embeddings of recording snippets with those between the recalibrated connectomes. Our model performs on par with the best of the 6 published BFMs we tested while having an order of magnitude fewer parameters. Our encoder performs better than FC on short scans and in smaller cohorts, especially in fingerprinting. We release the pretrained model weights, the code and the pretraining data, preprocessed and parcellated.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37642
PDF
-
Mathematical Transfer in LLMs Follows Reasoning Approach More Than Topic
2026-09-29 13:50:02Sajad Goudarzi, Samaneh Zamanifard, Seyed Amin Seyed Haeri, Moloud Nasiri, Hamed RahimianAbstract
When selecting mathematical training data for LLMs, a natural organizing principle is topic: probability examples for probability targets. An alternative is reasoning approach: worked solutions that share a solution method with the target, even when the mathematical domain differs. We ask which relation produces greater transfer after fine-tuning. We evaluate two counterbalanced $2\times2$ designs: probability and combinatorics crossed with invariant reasoning and double counting (2,000 problems), and number theory and geometry crossed with complement and pigeonhole reasoning (800 problems). In each design, every cell serves as the held-out target in turn: same-approach (SA) sources share the target's method but change the topic, while same-topic (ST) sources share the topic but change the method. Every source appears once in each role, so additive source-quality effects cancel from the equally weighted aggregate contrast. Across five base models and three training seeds per design, SA outperforms ST in all 40 seed-pooled model--target comparisons. Model-level advantages range from 8.2 to 16.2 percentage points in the primary design (mean: 10.8) and from 12.0 to 16.0 in the second design (mean: 14.3); all ten model-level 95% confidence intervals exclude zero. In both designs, ST sources are more similar to targets under embedding and lexical measures, so the SA advantage runs opposite to the measured ordering of statement-level resemblance. These findings identify reasoning approach as a more effective matching criterion than topic for mathematical transfer across the evaluated topic--approach combinations.
Abstract (translated)
URL
https://arxiv.org/abs/2610.00331
PDF
-
XU-RS: Explaining Credal Width in Random-Set Language Models
2026-09-29 13:48:33David Achara, Maryam Sultana, Alexander D. Rast, Fabio CuzzolinAbstract
Uncertainty estimates tell us how unsure a model is, but not why. Without knowing which parts of an input influences a model's uncertainty, we cannot tell whether that uncertainty score depends on input features that are relevant for the task. We study this problem in randomset classifiers built using pretrained language models. These classifiers assign probability to individual answers and to groups of answers, producing lower and upper probabilities for each answer; The difference between these probabilities, called credal width, is used to represent epistemic uncertainty about an answer arising from limited training data. We propose XU-RS, a framework that attributes an answer's credal width to the input tokens (words or word pieces) supplied to a language model. XU-RS uses Expected Gradients (a standard feature attribution method) to estimate how input tokens contribute to credal width. The proposed framework is evaluated on a MedQA dataset using SmolLM3-3B and Llama-2-7B models, demonstrating that setting the embedding of a token ranked highly by XU-RS to zero (zero-masking) causes larger changes in credal width than zero-masking randomly selected tokens. In addition, we show that normalisation can cause other answer groups to influence an answer's width, reveal how token attribution can mask numerical errors, and provide diagnostic checks to verify whether a token ranked highly by XU-RS meaningfully explains model uncertainty.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37594
PDF
-
Quantifying Behavioral Tails in Black-Box Language Models
2026-09-29 13:41:26Elsayed Eshra, Ali Al-Lawati, Dongwon Lee, Suhang WangAbstract
We introduce RareTrap, a framework for estimating the probability of severe behaviors in black box large language models (LLMs). A key challenge for probability estimation is defining a tractable distribution over the input space. To accomplish that, RareTrap uses a surrogate LLM and constructs a geometry-aware mapping from a lower-dimensional latent reference space into its token-embedding space to induce an explicit and reproducible distribution over input prompts. A response-level performance function is utilized on the response to quantify behavior severity. This enables sequential rare event simulation that concentrates evaluations on progressively more severe behaviors while preserving probability under the induced prompt distribution, which would otherwise be prohibitive to measure. Across 10 open-weight and two frontier models (GPT-5.4 and Claude Sonnet 4.6), we find that RareTrap successfully induces severe resource consumption behaviors and computes their probability with as few as 200 evaluations. RareTrap provides model developers a principled approach for evaluating language models under a common distribution, and prioritizing alignment effort to improve safety and mitigate risks.
Abstract (translated)
URL
https://arxiv.org/abs/2609.33638
PDF
-
RunyaNER: Auxiliary Language Selection for Runyankore NER
2026-09-29 13:22:15Prosper Arineitwe Asiimwe, Francois Meyer, Jan BuysAbstract
Cross-lingual zero-shot transfer and multilingual fine-tuning are promising approaches for NLP tasks such as Named Entity Recognition (NER) in low-resource languages, but in the absence of target language benchmarks, it is unclear which auxiliary language selection strategy leads to the best transfer. We introduce RunyaNER, the first publicly available NER benchmark for the East African language Runyankore, and use it to investigate the choice of which languages to use for transfer. Created with a semi-automated pipeline and fully manually verified, RunyaNER contains over 237k annotated words across 30k sentences. We benchmark pretrained models on RunyaNER, establishing that our dataset is of sufficient quality and size to produce effective Runyankore NER models. We then use RunyaNER to investigate auxiliary language selection in cross-lingual zero-shot and multilingual fine-tuning settings. Our experiments show that while transfer performance is highly sensitive to auxiliary language selection, embedding-based measures computed from labelled training spans correlate more strongly with downstream transfer performance than traditional linguistic features based on metadata or typology. By releasing RunyaNER and providing a systematic analysis of auxiliary language selection strategies, this work contributes both a new benchmark resource and practical insights for multilingual transfer in low-resource settings.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37543
PDF
-
Hierarchical Compression of Vision-Language Model Benchmarks
2026-09-29 13:08:46Hyunjong Ok, Seunggu Kang, Jaeho LeeAbstract
Thorough evaluation of vision-language models (VLMs) has become prohibitively expensive, as benchmarks span an ever-broader spectrum of capabilities and new models arrive at a relentless pace. Benchmark compression methods that preserve model rankings at a fraction of the cost are well studied for language models, but for VLMs the question remains under-explored. We present PRIMEBench (Pruning Redundant Items for Multimodal Evaluation), a vision-aware hierarchical benchmark compression framework that substantially reduces evaluation cost while preserving model rankings. This hierarchical framework operates in four stages: data cleaning to remove items answerable without the image and all-correct items, category representative selection to pick one benchmark per capability category, item pruning with Vision-Aware Variance (VAW), and category-count pruning. VAW combines inter-model variance with a vision-dependence score computed from multimodal embeddings alone, while encouraging coverage of diverse items within each benchmark. On models held out from item selection, it has the highest mean fidelity at the released 5% retention. The hierarchical design lets practitioners stop at any stage to match their compute budget; the released suite removes over 97% of items while preserving model rankings. Beyond compression, our analyses show how VLM evaluation behaves as model panels grow and evolve, providing guidance for designing future benchmarks that are more efficient, robust to model turnover, and explicit about the limits of evaluation-side pruning.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37515
PDF
-
LazySloth: Bounded LLM-based Lazy Tree Search for Fast Long Video Comprehension
2026-09-29 12:53:57Arka Mukherjee, Kaleen Shrestha, Larissa Zhu, Maja Matari\'cAbstract
Modern vision-language models (VLMs) have shown promising results in long-video understanding due to the rich semantic information they can capture. However, most methods focus on coarse captioning of extracted image frames that are computationally inefficient and require models with large context windows. While past work has explored efficient methods through multimodal retrieval-augmented generation (RAG), they rely on lossy embeddings that lose temporal context and fine-grained detail. Few works to date have investigated how VLM-based query-relevant information retrieval can be optimized. We introduce LazySloth, an efficient tree-based search method that speeds up video comprehension and retrieval tasks 2.9-8.3x (compared to existing agentic methods) through bounded captioning of portions of the video considered irrelevant by a VLM of the video. Compared to contemporary specialized video-understanding VLMs and RAG-based methods, LazySloth achieved similar or better final task accuracy across two recent open-source base VLMs--Gemma 4 31B and Qwen3.6 27B--across four benchmarks. LazySloth reduced the gap between the base open-source model and a closed-source model, GPT-4o. Ablations showed that replacing VLM scene understanding with CLIP-based retrieval cost 8.8-19.9% in accuracy, while lazy tree construction matches eager construction at a fraction of the captioning cost. With LazySloth, we demonstrate the possibility of faster long-video comprehension without substantial loss in performance.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37426
PDF
-
Technical note on: Zero-Training Feature-Space Alignment via Information Geometry
2026-09-29 11:39:56Behraj Khan, Tahir Qasim Syed, Syed Ahmad Chan BukhariAbstract
Deep vision models often degrade under distribution shift. Test-time adaptation can improve robustness but typically requires iterative optimization, hyperparameter tuning, and multiple forward-backward passes. We propose Zero-Training Fisher Geometry Alignment (ZFGA), a closed-form method that improves robustness under covariate shift without modifying model parameters. ZFGA is based on the observation that distribution shifts distort feature-space geometry. It estimates the Fisher information matrix of the predictive distribution with respect to feature embeddings and applies a linear transformation that aligns test-feature Fisher geometry with a reference geometry computed from clean data. This provides a natural-gradient-inspired preconditioning step in feature space. We evaluate ZFGA on CIFAR-10-C and ImageNet-C using ResNet-50, DINO ViT-S/16, and CLIP ViT-B/32. ZFGA consistently improves over zero-shot inference across all three models, although it is not the strongest method for every model. Covariance whitening performs better on ResNet-50, while Fisher whitening is statistically indistinguishable from ZFGA on CLIP. Across six training-free and gradient-based alternatives (covariance whitening, Fisher whitening, TENT, T3A, LAME, and AdaNPC), ZFGA is the only method that does not substantially harm any of the three model families. The Fisher geometry distortion is also positively correlated with ZFGA gain (Pearson r = 0.366, p = 0.017), providing preliminary evidence that geometric misalignment contributes to robustness degradation. ZFGA requires only forward passes and matrix operations at inference time, offering a lightweight and deterministic alternative to optimization-based test-time adaptation.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37302
PDF
-
ResComEmb: Effective and Efficient Multimodal Embedding via Residual Homogeneity Compression
2026-09-29 10:39:46Zijing Cai, Yuzhe Wang, Jingxian Zhu, Fengbin Zhu, Richang HongAbstract
Multimodal large language models (MLLMs) have shown strong potential for universal multimodal representation learning. However, existing methods either compress each input into a single vector, limiting fine-grained expressiveness, or retain long sequences of visual-token vectors, incurring substantial storage and interaction costs. To resolve this trade-off, we propose ResComEmb, a trainable framework for effective and efficient universal multi-vector multimodal embedding. ResComEmb first encodes each input at native dynamic resolution into ordered global, intermediate, and fine-grained views. After MLLM contextualization and embedding projection, a trainable Residual Homogeneity Compression (RHC) module reduces within-granularity redundancy and cross-granularity repetition under explicit visual token budgets. Then, ResComEmb introduces a length-adaptive Bidirectional Late-Interaction Matching mechanism for robust query-document scoring, which averages the strongest token-level matches in each direction and combines the two scores using a weight based on how many valid tokens each side has. Extensive experiments on MMEB, ViDoRe V1, and ViDoRe V2 show that ResComEmb produces higher-quality universal multimodal embeddings than VLM2Vec-V2, and outperforms ColQwen2.5 in visual document retrieval using only 37.5% of its full visual token budget, demonstrating a favorable effectiveness-efficiency trade-off.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37225
PDF
-
Beyond Semantic Narrowing: Robust and Efficient LLM Watermarking with Hamming Neighborhoods
2026-09-29 10:34:13Zewen Sun, Tongyang Zhao, Liyao Xiang, Mingxuan Ma, Lingzhe Wang, Zhiyuan LiAbstract
Semantic watermarking improves robustness against watermark removal attacks by embedding detectable signals into sentence-level representations. However, existing watermarking methods typically impose watermark-specific semantic preferences on generated sentences without explicitly accounting for the highly non-uniform and context-dependent semantic preference of LLM generation. When these two preferences are poorly aligned, many natural continuations become incompatible with the watermark, causing semantic narrowing: reduced semantic freedom, increased resampling cost, and potential degradation on tasks with strict semantic requirements. To alleviate this problem, we propose HammingMark, which uses the semantic hash of the preceding sentence as a dynamic center and accepts candidates whose hashes fall within its Hamming neighborhood. Defining watermark validity over a Hamming neighborhood in compact hash space retains a larger fraction of naturally likely semantic continuations. The coarse many-to-one hash mapping further allows diverse semantic realizations to remain watermark-valid. Experiments on C4 and BookSum show that HammingMark achieves strong robustness, high detectability, and near-unwatermarked generation quality, requiring only 2.2 sampled candidates per accepted sentence,a 72.8% reduction compared with the most sampling-efficient existing method. On more complex tasks with strict semantic constraints, HammingMark achieves the highest detection rates with the highest or tied-highest ROUGE-L scores, demonstrating its effectiveness in balancing watermark detectability and generation quality under constrained generation settings.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37218
PDF
-
V-Engram: Trigger-Indexed External Memory for Modular Text-to-Image Personalization
2026-09-29 10:17:23Haoran He, Runyuan Cai, Yiming Wang, Lin Yu, Xiaodong ZengAbstract
Pretrained text-to-image models contain broad visual knowledge, yet they cannot reliably acquire or refine a specific visual identity from only a few references while preserving compositional control. Token-embedding methods are compact but often underfit identity, whereas adapter-based methods improve fidelity through persistent weight updates that can be costly to store and interfere when concepts are composed. We introduce V-Engram, a trigger-indexed external memory mechanism for Stable Diffusion 3.5. Each concept is assigned an explicit trigger that retrieves concept-specific memory, whose gated directions enter frozen text-encoder and MMDiT context states as relative residuals. Separating this memory from backbone adaptation enables prompt-selective and multi-concept access without merging model updates. Experiments show that V-Engram broadly matches DreamBooth-LoRA in overall subject fidelity while showing advantages in settings such as contextual subject preservation. Prompt-matched loading retrieves only matched entries, reducing most additional adaptation-state loading for a single-concept query. Qualitative results further demonstrate paired-trigger composition and same-class separation, while prompts without registered entries retain the frozen model's base behavior. Together, these results establish trigger-indexed memory as a modular interface for adding targeted visual evidence without rewriting the generator.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37198
PDF
-
From Sky to Soil: A Morphing Aerial-Ground Robot for Seed Deployment
Abstract
Aerial seed broadcasting can reach remote restoration sites, but provides limited control over seed placement within the soil. This paper presents a geometry-assisted, tri-functional morphing robot that combines aerial access, ground locomotion, and controlled-depth seed embedding in a fly-drive-plant architecture. After landing, the platform reconfigures into a four-wheeled planting configuration: an electronically coupled dual-motor drive folds the rear arms outward to form ground wheels, while a descending front tray engages the propulsion motors with a drill gear train. Reusing the propulsion motors for drilling eliminates a dedicated drill drive. The planting sequence forms a hole, dispenses a seed, and allows the vehicle to reposition on the ground or return to its flight configuration. Ground mobility supports repeated planting without requiring a separate flight between adjacent sites. A companion controller issues reconfiguration, tray, and seed-gate commands, while a dedicated autopilot handles flight control. The morphing and planting mechanisms are validated using a hardware prototype that demonstrates ground repositioning and seed embedding. This proof of concept establishes a hardware basis for aerial-ground seed embedding through coordinated reconfiguration and actuator reuse.
Abstract (translated)
URL
https://arxiv.org/abs/2609.37072
PDF
-
$lambda$-JEPA Spectral Anti-Collapse Regularization for Self-Supervised Learning
2026-09-29 09:02:47Berker Demirel, Cl\'ementine Domin\'e, Valentino Maiorca, Marco Fumero, Marco Mondelli, Francesco LocatelloAbstract
Joint-embedding self-supervised learning typically combines an invariance objective across augmented views with additional mechanisms to prevent representational collapse. These objectives are often applied after a projection head, while downstream tasks use the backbone representation before the projector. We find that this mismatch does not necessarily prevent dimensional collapse in the backbone, which can retain low effective rank and potentially limit downstream transfer. To address this, we introduce SACReg, a spectral anti-collapse regularizer motivated by an analysis of $\lambda$-balance, which captures the relative scale of weight matrices across layers. In a two-layer linear network, we show that (i) $\lambda$-balance prevents collapse, and (ii) our regularizer applied to the backbone induces $\lambda$-balance. In the nonlinear case, this regularizer leads to anti-collapse as well and, in realistic architectures on ImageNet100, it empirically increases the representations' ranks. We apply SACReg to JEPA and propose $\lambda$-JEPA, which improves over LeJEPA and VISReg on ImageNet-1k classification and in average linear-probe transfer performance across eight downstream image datasets. On video self-supervised learning, $\lambda$-JEPA improves over LeVJEPA and V-JEPA 2 on the Something-Something-v2 and Kinetics-400 benchmarks. Code is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2609.35288
PDF
-
ImbalancE: Inference-Time Latent Search Against Degree Imbalance in Link Prediction
2026-09-29 08:26:37Alberto Bernardi, Luca Costabello, Christophe GueretAbstract
Knowledge Graph Embedding models have been extensively used to learn representations of entities and relations in Knowledge Graphs for predicting missing links. However, the quality of the learned representations varies a lot across different areas of the graph. If previous research has loosely linked the problem to relation types or degree bias, we show that it is more widespread and it correlates with the degree imbalance of the entities in test triples. In particular, the prediction of a target entity that has a degree much smaller than the degree of the anchor entity is extremely problematic. This is critical in recommender systems and other use cases, where these triples represent important corner cases. To address this issue, we propose an inference-time latent search optimization method capable of significantly improving model predictions on the most imbalanced triples. Built on top of a pre-trained model, it explores the embedding space at evaluation time, blending known and out-of-band information to mitigate the degree imbalance bias. We show the value of our approach on imbalanced triples from common benchmark datasets, where we outperform conventional methods, opening the door to the successful adoption of Knowledge Graph Embedding models on these critical corner cases.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36996
PDF
-
Structured Visual Target Learning For Cross-Subject eeg-to-image retrieval
2026-09-29 08:07:20Salini Yadav, Taveena Lotey, Micka\"el Coustaty, Pravendra Singh, Partha Pratim RoyAbstract
Cross-subject EEG-to-image retrieval requires a neural represen- tation trained on source subjects to remain aligned with a visual embedding space for an unseen subject. Whereas existing methods primarily focus on the EEG side, we address this problem from the perspective of the visual target. Our approach preserves the spatial information of the Perception Encoder, converts its patch grid into a compact set of learned visual views, and aggregates them for each image with a block-structured, content-dependent router. The target is learned jointly with the EEG encoder through contrastive learning with MMD regularization across source subjects. For deployment, we propose a training-free representation refinement that aligns frozen embeddings without updating either encoder. Under leave- one-subject-out evaluation on THINGS-EEG2, the structured target achieves 35.3%/65.6% Top-1/Top-5 accuracy, the best among com- pared methods. Refinement raises this to 48.1%/77.1%, an 18.5% Top-1 gain over the strongest compared method, improving all ten held-out subjects.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36971
PDF
-
ER-JEPA: Experience Replay Improves Joint-Embedding Predictive Learning in Language Models
2026-09-29 07:56:01Jingnan Pu, Zi-En Fan, Feng LianAbstract
Large language models (LLMs) excel at token-level generation but may learn undesirable abstract semantics and lack comprehensive perception. LLM-JEPA mitigates this by aligning different views of the same underlying knowledge via a joint-embedding predictive architecture (JEPA). However, strong alignment does not necessarily lead to accurate, stable predictions. To address this, we propose ER-JEPA, which adds an episodic replay path to LLM-JEPA. ER-JEPA stores training pairs in a memory. At each step, it stores and retrieves relevant data to provide additional supervision. This enables learning from both the current batch and stored training pairs, providing additional supervision for token prediction and representation alignment. Experiments across multiple datasets (NL-RX, GSM8K, Spider, and NQ-Open) demonstrate that ER-JEPA consistently outperforms LLM-JEPA.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36952
PDF
-
Safer Content or Firmer Refusals? A Hybrid Perturbation Defense for Alignment under Harmful Fine-tuning
2026-09-29 07:04:31Muhammad Zeeshan Akram, Mufid Kamel Marican, Anvesh Reddy Yenugu, Ali Zain Kaimkhani, Minghong FangAbstract
Fine-tuning-as-a-service lets users adapt a safety-aligned language model to their own data, but it also creates a harmful fine-tuning attack surface: a small amount of harmful data mixed into an otherwise benign fine-tuning set can degrade the model's alignment. Two recent alignment-stage defenses address this problem at different levels of the model. Vaccine improves the robustness of hidden embeddings to the representation shifts induced by harmful fine-tuning, whereas Booster simulates harmful weight updates and attenuates their effect during alignment. We investigate whether these mechanisms are complementary and propose VaccineBooster, a single alignment procedure that combines embedding perturbation and weight-level gradient attenuation within each training step. On Llama-2-7B aligned with BeaverTails and then attacked through poisoned fine-tuning, VaccineBooster achieves the lowest OpenAI moderation score among the compared defenses, 0.315, while a Booster-Only variant retains the highest post-attack refusal rate, 50%. Together with ablations over the embedding-perturbation and gradient-attenuation strengths, these results indicate a trade-off: embedding perturbation primarily reduces flagged harmful content, whereas gradient attenuation primarily preserves explicit refusal behavior. Because our evaluation uses ten prompts and a single unseeded run per configuration, we report this trade-off as an observed pattern rather than a statistically resolved effect. These results provide practical guidance for prioritizing content safety or refusal retention when aligned models are exposed to untrusted fine-tuning.
Abstract (translated)
URL
https://arxiv.org/abs/2609.36862
PDF
