Institution profile

Capgemini

Industry researcheurope · fr
Official website
Research library20linked papers
Opportunities0open roles
Selected work

Representative Papers

Efficient Sequential Neural Network with Spatial-Temporal Attention and Linear LSTM for Robust Lane Detection Using Multi-Frame Images

Feb 03, 2026

This work addresses the challenges of achieving accuracy, robustness, and real-time performance in visual lane detection under complex scenarios such as occlusion and strong illumination, where existing methods often fail to exploit the spatiotemporal saliency of critical regions. To this end, we propose a lightweight encoder-decoder sequential network that integrates a spatial-temporal attention mechanism with a linear LSTM module. By leveraging multi-frame inputs, our model effectively captures spatiotemporal dependencies and adaptively focuses on salient lane features. Extensive experiments demonstrate that the proposed method outperforms state-of-the-art approaches on three large-scale public benchmarks, particularly excelling in challenging conditions, while significantly reducing both model parameters and computational cost (measured in MACs), thereby enabling efficient and robust end-to-end multi-frame lane detection.

1 citationsRead paper

Regional Explanations: Bridging Local and Global Variable Importance

Apr 13, 2026

Existing local attribution methods, such as Local Shapley Values and LIME, remain prone to erroneously assigning importance to irrelevant features even under ideal conditions, leading to unreliable explanations. This work proposes R-LOCO, a novel approach that partitions the input space into regions where feature importance is approximately homogeneous and applies a global attribution strategy within each region to construct instance-specific yet stable local explanations. By innovatively integrating local and global explanation paradigms, R-LOCO satisfies established axiomatic principles of sound attribution. Empirical evaluations demonstrate that R-LOCO significantly outperforms Local Shapley Values and LIME in both fidelity and stability, offering more trustworthy and consistent interpretability for individual predictions.

0 citationsRead paper

ToolFlood: Beyond Selection -- Hiding Valid Tools from LLM Agents via Semantic Covering

Mar 14, 2026

This work addresses the vulnerability of tool-augmented large language model (LLM) agents during the embedding-based retrieval phase and proposes ToolFlood, a novel attack method that stealthily disrupts legitimate tool selection by injecting a small number of malicious tools with broad semantic coverage. These adversarial tools prevent valid candidates from appearing in the top-k retrieval results, thereby compromising the agent’s tool-calling capability. ToolFlood leverages LLM-generated diverse tool metadata, a greedy iterative selection algorithm based on cosine distance, and geometric analysis in the embedding space to achieve efficient semantic coverage. Experiments on benchmarks such as ToolBench demonstrate a 95% attack success rate with only a 1% tool injection ratio. Furthermore, this study uncovers and provides the first theoretical analysis of the retrieval saturation phenomenon.

0 citationsRead paper
Recent publications

Latest Papers

Regional Explanations: Bridging Local and Global Variable Importance

Apr 13, 2026

Existing local attribution methods, such as Local Shapley Values and LIME, remain prone to erroneously assigning importance to irrelevant features even under ideal conditions, leading to unreliable explanations. This work proposes R-LOCO, a novel approach that partitions the input space into regions where feature importance is approximately homogeneous and applies a global attribution strategy within each region to construct instance-specific yet stable local explanations. By innovatively integrating local and global explanation paradigms, R-LOCO satisfies established axiomatic principles of sound attribution. Empirical evaluations demonstrate that R-LOCO significantly outperforms Local Shapley Values and LIME in both fidelity and stability, offering more trustworthy and consistent interpretability for individual predictions.

0 citationsRead paper

ToolFlood: Beyond Selection -- Hiding Valid Tools from LLM Agents via Semantic Covering

Mar 14, 2026

This work addresses the vulnerability of tool-augmented large language model (LLM) agents during the embedding-based retrieval phase and proposes ToolFlood, a novel attack method that stealthily disrupts legitimate tool selection by injecting a small number of malicious tools with broad semantic coverage. These adversarial tools prevent valid candidates from appearing in the top-k retrieval results, thereby compromising the agent’s tool-calling capability. ToolFlood leverages LLM-generated diverse tool metadata, a greedy iterative selection algorithm based on cosine distance, and geometric analysis in the embedding space to achieve efficient semantic coverage. Experiments on benchmarks such as ToolBench demonstrate a 95% attack success rate with only a 1% tool injection ratio. Furthermore, this study uncovers and provides the first theoretical analysis of the retrieval saturation phenomenon.

0 citationsRead paper

Scalar-Measurement Attitude Estimation on $\mathbf{SO}(3)$ with Bias Compensation

Mar 02, 2026

This work addresses the challenge that traditional attitude estimation methods rely on full vector measurements, whereas many practical scenarios provide only scalar observations. To overcome this limitation, the paper proposes a nonlinear deterministic observer on $\mathbf{SO}(3)$ that jointly estimates attitude and gyroscope bias using scalar measurements alone. Theoretical analysis establishes, for the first time, that attitude observability is guaranteed with merely two scalar measurements under sufficient excitation, and three suffice in static conditions. The method is validated on the BROAD dataset, demonstrating robust performance with low estimation error even under severely degraded measurement configurations, thereby highlighting its strong stability and practical applicability.

0 citationsRead paper