Institution profile

AITRICS

Industry researchasia · kr
Official website
Research library44linked papers
Opportunities0open roles
Selected work

Representative Papers

Learning Beyond What You Sample: Off-Policy-Aware Cross-Model Trajectory Exchange for RLVR

Sep 29, 2026

This study addresses the absence of gradient signals in completely failed groups caused by limited sampling in reinforcement learning by proposing the GRAFT framework. Without requiring a designated teacher model, this approach leverages the complementarity of multiple models to exchange trajectories across them, enhancing policy learning via a gated replacement mechanism. Furthermore, it introduces sequence-level compatibility weighting and token-level importance ratio clipping to effectively mitigate distribution shift. Experimental results demonstrate that GRAFT achieves an average improvement of 2.1 points on mathematical reasoning benchmarks, with gains reaching up to 4.5 points, thereby enabling efficient teacher-free mutual learning.

0 citationsRead paper

ChronoFlow: Hierarchical Flow Matching for Irregular Time Series Generation

Sep 27, 2026

This study addresses the challenge that existing methods struggle to natively generate irregular time series encompassing sampling structures, observation frequencies, and feature dependencies. To this end, it proposes a unified hierarchical flow matching framework that decomposes the complex joint generation process into structurally aligned subproblems. By sequentially generating observation counts, timestamps, patterns, and feature values from coarse to fine statistical granularity, the approach effectively preserves multi-scale dependencies. Furthermore, this work introduces novel evaluation metrics that capture both sample fidelity and dependency preservation. Extensive experiments across five benchmark datasets demonstrate that the proposed method achieves significantly superior generation fidelity compared to existing baselines, thereby validating the effectiveness of the hierarchical decomposition strategy.

0 citationsRead paper

Domain Generalization under Sampling Pattern Shifts in Irregular Time Series

Sep 27, 2026

This study addresses the poor domain generalization robustness and model shortcut reliance caused by sampling pattern shifts in irregular time series. To this end, we propose PRISM, a novel framework that first constructs HAR-C, the inaugural controlled sampling shift benchmark. Methodologically, PRISM employs unsupervised dual-view representation learning to disentangle intrinsic features from sampling representations. This is further combined with an adversarial and diversity-enhanced supervised training strategy to eliminate shortcut dependencies on specific sampling patterns. Experimental results demonstrate that the proposed approach significantly improves predictive robustness against unseen sampling shifts across both controlled and real-world benchmarks, outperforming existing domain generalization methods.

0 citationsRead paper

Looks the Same, Answers Differently: Flip-Direction Steering for Robust Vision-Language Reasoning

Sep 23, 2026

This study addresses the vulnerability of vision-language models to subtle image perturbations, which can cause reasoning trajectory deviations and answer flipping. To this end, we propose FlipDir, a novel training-free direction-guided technique that locates flip-inducing activation subspaces via low-rank subspace estimation and contrastive learning, and selectively steers hidden states during decoding through a margin-based gating mechanism to achieve robust inference. Furthermore, we construct the VisFlip benchmark to systematically evaluate prediction recovery and stability preservation capabilities. Extensive experiments demonstrate that our method consistently outperforms existing baselines across 18 settings, significantly enhancing model robustness against visual perturbations.

0 citationsRead paper
Recent publications

Latest Papers

Learning Beyond What You Sample: Off-Policy-Aware Cross-Model Trajectory Exchange for RLVR

Sep 29, 2026

This study addresses the absence of gradient signals in completely failed groups caused by limited sampling in reinforcement learning by proposing the GRAFT framework. Without requiring a designated teacher model, this approach leverages the complementarity of multiple models to exchange trajectories across them, enhancing policy learning via a gated replacement mechanism. Furthermore, it introduces sequence-level compatibility weighting and token-level importance ratio clipping to effectively mitigate distribution shift. Experimental results demonstrate that GRAFT achieves an average improvement of 2.1 points on mathematical reasoning benchmarks, with gains reaching up to 4.5 points, thereby enabling efficient teacher-free mutual learning.

0 citationsRead paper

ChronoFlow: Hierarchical Flow Matching for Irregular Time Series Generation

Sep 27, 2026

This study addresses the challenge that existing methods struggle to natively generate irregular time series encompassing sampling structures, observation frequencies, and feature dependencies. To this end, it proposes a unified hierarchical flow matching framework that decomposes the complex joint generation process into structurally aligned subproblems. By sequentially generating observation counts, timestamps, patterns, and feature values from coarse to fine statistical granularity, the approach effectively preserves multi-scale dependencies. Furthermore, this work introduces novel evaluation metrics that capture both sample fidelity and dependency preservation. Extensive experiments across five benchmark datasets demonstrate that the proposed method achieves significantly superior generation fidelity compared to existing baselines, thereby validating the effectiveness of the hierarchical decomposition strategy.

0 citationsRead paper

Domain Generalization under Sampling Pattern Shifts in Irregular Time Series

Sep 27, 2026

This study addresses the poor domain generalization robustness and model shortcut reliance caused by sampling pattern shifts in irregular time series. To this end, we propose PRISM, a novel framework that first constructs HAR-C, the inaugural controlled sampling shift benchmark. Methodologically, PRISM employs unsupervised dual-view representation learning to disentangle intrinsic features from sampling representations. This is further combined with an adversarial and diversity-enhanced supervised training strategy to eliminate shortcut dependencies on specific sampling patterns. Experimental results demonstrate that the proposed approach significantly improves predictive robustness against unseen sampling shifts across both controlled and real-world benchmarks, outperforming existing domain generalization methods.

0 citationsRead paper

Looks the Same, Answers Differently: Flip-Direction Steering for Robust Vision-Language Reasoning

Sep 23, 2026

This study addresses the vulnerability of vision-language models to subtle image perturbations, which can cause reasoning trajectory deviations and answer flipping. To this end, we propose FlipDir, a novel training-free direction-guided technique that locates flip-inducing activation subspaces via low-rank subspace estimation and contrastive learning, and selectively steers hidden states during decoding through a margin-based gating mechanism to achieve robust inference. Furthermore, we construct the VisFlip benchmark to systematically evaluate prediction recovery and stability preservation capabilities. Extensive experiments demonstrate that our method consistently outperforms existing baselines across 18 settings, significantly enhancing model robustness against visual perturbations.

0 citationsRead paper