Institution profile

Max Planck Institute for Informatics

Academic institutioneurope · de
Official website
Research library369linked papers
Opportunities0open roles
Selected work

Representative Papers

Sports-QA: A Large-Scale Video Question Answering Benchmark for Complex and Professional Sports

Jan 03, 2024arXiv.org

Existing VideoQA datasets lack fine-grained modeling of professional sports actions, hindering effective reasoning for descriptive, temporal, causal, and counterfactual questions. To address this, we introduce Sports-QA—the first video question answering benchmark tailored to professional sports scenarios—covering multiple sports disciplines and four categories of complex reasoning tasks. Methodologically, we propose the Auto-Focus Transformer (AFT), which employs an attention-driven dynamic focusing mechanism to adaptively model multi-scale temporal information and integrates joint video–language representation learning. Extensive experiments demonstrate that AFT achieves state-of-the-art performance on Sports-QA, substantially outperforming general-purpose VideoQA models. This work constitutes the first systematic validation of an architecture explicitly designed for fine-grained sports action understanding and dynamic logical reasoning, establishing a new foundation for domain-specific VideoQA research.

10 citations2 influentialRead paper

Manipulating Feature Visualizations with Gradient Slingshots

Jan 11, 2024arXiv.org

This work exposes a critical credibility vulnerability in feature visualization (FV) for deep neural network interpretability: FV outputs are susceptible to stealthy manipulation, leading to erroneous attribution of neuron semantics. To address this, we propose the first model-architecture-agnostic targeted FV manipulation method. Our approach integrates gradient redirection (via Slingshot optimization), adversarial latent-space perturbations, and neuron-activation-constrained regularization to achieve “semantic masking”—i.e., seamless substitution of a target neuron’s original FV explanation with an arbitrary user-specified semantic concept. Experiments across CNNs and Vision Transformers demonstrate successful concealment of functionally critical neurons: model accuracy degrades by less than 0.3%, yet FV-based auditing yields a 92% false-negative rate in detecting manipulated neurons. These results underscore the fragility of prevailing FV techniques and establish a new paradigm for robust model auditing and interpretability governance.

6 citationsRead paper

Explaining the Inherent Tradeoffs for Suffix Array Functionality: Equivalences between String Problems and Prefix Range Queries

Oct 22, 2025

This paper addresses the problem of efficiently accessing suffix arrays (SAs) when they cannot be stored explicitly. It establishes, for the first time, a *bidirectional equivalence*—in space, query time, and construction efficiency—between SA access and prefix selection, unifying the complexity analysis of fundamental string indexing operations. Through systematic reductions, the authors identify six pairs of intrinsically equivalent problems and prove that nearly all optimal SA representations can be realized via prefix selection structures. Leveraging this equivalence, they design a data structure supporting sublinear construction: for binary text, it achieves *O(n)* bits of space, *O(n/√log n)* preprocessing time, and *O(log^ε n)* query time—*matching and closing a long-standing complexity gap* in the field.

2 citationsRead paper

Boosting Segment Anything Model to Generalize Visually Non-Salient Scenarios

Jan 02, 2026IEEE Transactions on Image Processing

This work addresses the significant performance degradation of the Segment Anything Model (SAM) in visually non-salient scenarios, where foreground and background exhibit low contrast. To mitigate this limitation, the authors propose VNS-SAM, which incorporates a Mask-Edge Token Interactive decoder and a Non-Salient Feature Mining module. These components effectively enhance SAM’s perception of non-salient objects with minimal additional parameters and computational overhead, while preserving its zero-shot generalization capability. The study also introduces VNS-SEG, the first unified benchmark dataset dedicated to multi-class non-salient segmentation. Experimental results demonstrate that VNS-SAM achieves superior performance across diverse non-salient segmentation tasks, notably outperforming baseline methods under zero-shot settings. The added parameters can be optimized within four hours, and both the model and dataset are publicly released.

1 citationsRead paper

Beating Meet-in-the-Middle for Subset Balancing Problems

Nov 13, 2025

This paper studies the Subset Balancing problem: given an integer vector and a constant-size coefficient set (C), find a nonzero integer solution vector (x) such that the dot product of (x) and the input vector is zero. The problem unifies classical NP-hard problems including Subset Sum, Partition, and Equal Subset Sum. Addressing the long-standing worst-case time complexity bottleneck—the Meet-in-the-Middle lower bound of (O(|C|^{n/2}))—this work achieves the first worst-case improvement. We introduce a novel hybrid representation technique, extend the Howgrave-Graham-Joux framework, incorporate flexible input encoding and pseudosolution cancellation, and design an efficient compatible solution-pair recovery algorithm. Our main result is a deterministic (O(|C|^{(0.5-varepsilon)n})) time algorithm for some constant (varepsilon > 0). This yields exponential speedups for Equal Subset Sum and related problems, significantly surpassing prior state-of-the-art algorithms, which only achieve such bounds in average-case settings.

1 citationsRead paper
Recent publications

Latest Papers

Stuffed IBLTs: Optimal Linear Multiset Sketches

Sep 15, 2026

该论文提出了一种名为Stuffed IBLT的线性草图方法，用于精确恢复原始向量。此方法在保持信息理论最优空间使用的同时，支持高效的更新和解码操作，解决了多集合协调问题。

0 citationsRead paper