Scaling video diffusion transformers (DiTs) is limited by their quadratic 3D attention, even though most of the attention mass concentrates on a small subset of positions. We turn this observation into VSA, a trainable, hardware-efficient sparse attention that replaces full attention at both training and inference. In VSA, a lightweight coarse stage pools tokens into tiles and identifies high-weight critical tokens; a fine stage computes token-level attention only inside those tiles subjecting to block computing layout to ensure hard efficiency. This leads to a single differentiable kernel that trains end-to-end, requires no post-hoc profiling, and sustains 85% of FlashAttention3 MFU. We perform a large sweep of ablation studies and scaling-law experiments by pretraining DiTs from 60M to 1.4B parameters. VSA reaches a Pareto point that cuts training FLOPS by 2.53 with no drop in diffusion loss.

This paper proposes the Dual-Stage Sparse Attention (DuSA) mechanism for attention acceleration of transformers. In the first stage, DuSA performs intrablock sparse attention to aggregate local inductive biases. In the second stage, DuSA performs interblock sparse attention to obtain long-range dependencies. Both stages have low computational complexity and can be further accelerated by memory acceleration attention mechanisms directly, which makes DuSA faster than some extremely fast attention mechanisms. The dual-stage sparse attention design provides a lower error in approximating vanilla scaled-dot product attention than the basic single-stage sparse attention mechanisms and further advances the basic sparse attention mechanisms to match or even outperform vanilla scaled-dot product attention. Even in some plug and play situations, DuSA can still maintain low performance loss. DuSA can be used in both training and inference acceleration. DuSA achieves leading performance in different benchmarks: long range arena, image classification, semantic segmentation, object detection, text to video generation, and long context understanding, and accelerates models of different sizes.

Weextend our analysis to the more general class of vector symbolic architectures (VSA), which compute withhigh-dimensional vectors(hypervectors) thatarenotnecessarily binary.

Despite their capabilities, Large Language Models (LLMs) remain opaque with limited understanding of their internal representations. Current interpretability methods either focus on input-oriented feature extraction, such as supervised probes and Sparse Autoencoders (SAEs), or on output distribution inspection, such as logit-oriented approaches. A full understanding of LLM vector spaces, however, requires integrating both perspectives, something existing approaches struggle with due to constraints on latent feature definitions. We introduce the Hyperdimensional Probe, a hybrid supervised probe that combines symbolic representations with neural probing. Leveraging Vector Symbolic Architectures (VSAs) and hypervector algebra, it unifies prior methods: the top-down interpretability of supervised probes, SAE's sparsity-driven proxy space, and output-oriented logit investigation. This allows deeper input-focused feature extraction while supporting output-oriented investigation. Our experiments show that our method consistently extracts meaningful concepts across LLMs, embedding sizes, and setups, uncovering concept-driven patterns in analogy-oriented inference and QA-focused text generation. By supporting joint input-output analysis, this work advances semantic understanding of neural representations while unifying the complementary perspectives of prior methods.

Multiple Instance Learning (MIL) tasks impose a strict logical constraint: a bag is labeled positive if and only if at least one instance within it is positive. While this iff constraint aligns with many real-world applications, recent work has shown that most deep learning-based MIL approaches violate it, leading to inflated performance metrics and poor generalization. We propose a novel MIL framework based on Vector Symbolic Architectures (VSAs), which provide a differentiable mechanism for performing symbolic operations in high-dimensional space. Our method encodes the MIL assumption directly into the model's structure by representing instances and concepts as nearly orthogonal high-dimensional vectors and using algebraic operations to enforce the iff constraint during classification. To bridge the gap between raw data and VSA representations, we design a learned encoder that transforms input instances into VSA-compatible vectors while preserving key distributional properties. Our approach, which includes a VSA-driven MaxNetwork classifier, achieves state-of-the-art results for a valid MIL model on standard MIL benchmarks and medical imaging datasets, outperforming existing methods while maintaining strict adherence to the MIL formulation. This work offers a principled, interpretable, and effective alternative to existing MIL approaches that rely on learned heuristics.

Complex Query Answering (CQA) over incomplete Knowledge Graphs (KGs), typically formalized as reasoning with Existential First-Order predicate logic with one free variable (EFO\textsubscript{1}), faces a fundamental tradeoff between logic fidelity and computational efficiency. This work establishes a Grounding-Skolemization dichotomy to systematically analyze this challenge and motivate a paradigm shift in CQA. While Grounding-based methods inherently suffer from combinatorial explosion, most Skolemization-based methods neglect to explicitly model Skolem functions and compromise logical consistency. To address these limitations, we propose the Logic-constrained Vector Symbolic Architecture (LVSA), a neuro-symbolic framework that unifies a differentiable Skolemization module and a neural negator, as well as a logical constraint-driven optimization protocol to harmonize geometric and logical requirements. Theoretically, LVSA guarantees universality for all EFO\textsubscript{1} queries with low computational complexity. Empirically, it outperforms state-of-the-art Skolemization-based methods and reduces inference costs by orders of magnitude compared to Grounding-based baselines.