Selected work

Representative Papers

ChronoStitch: Training-Free Composition of Visual KV Memories for Long-Horizon Temporal Reasoning

Jul 21, 2026

This work addresses the challenge of cross-segment temporal reasoning in long-form video question answering, where existing methods suffer from temporal phase conflicts in visual key-value (KV) caches due to local positional encodings. To overcome this limitation without requiring retraining, the authors propose an efficient approach that reconstructs the KV cache using a global triaxial multimodal RoPE coordinate system, preserving spatiotemporal structural consistency and avoiding spatial-temporal confusion caused by one-dimensional reindexing. Additionally, the method selectively recomputes high-bias visual tokens to compensate for missing contextual information. Notably, this is the first work to employ three-dimensional RoPE for cross-segment temporal alignment, achieving significant performance gains over baselines on Qwen2.5-VL-3B and TempCompass, with improved event ordering accuracy and 3.3× faster inference compared to full re-prefilling.

0 citationsRead paper

BLUE: A Stale-Pixel Optical-Flow Compositor for Entropy-Efficient Surveillance Video Encoding

Jun 27, 2026

This work addresses the inefficiency in storing static surveillance videos caused by repeatedly encoding invariant backgrounds. The authors propose BLUE, a pre-encoding synthesizer that integrates background pixel substitution with synthesis prior to compression: by maintaining a background seed frame, it replaces corresponding regions in input frames with static pixels, prompting the encoder to generate near-zero-bitrate SKIP macroblocks while preserving foreground visual fidelity. A lightweight VMAF-aware activation gating mechanism is introduced to balance compression gains against perceptual quality. Compatible with both x264 and x265 encoders, BLUE achieves average bitrate savings of 34.6% (x264) and 39.4% (x265) on the VIRAT dataset at CRF 28, improves foreground PSNR in 60.7% of clips, and incurs negligible full-frame quality degradation (ΔVMAF ≤ −2.9) in 40% of cases.

0 citationsRead paper

Recent publications

Latest Papers

ChronoStitch: Training-Free Composition of Visual KV Memories for Long-Horizon Temporal Reasoning

Jul 21, 2026

This work addresses the challenge of cross-segment temporal reasoning in long-form video question answering, where existing methods suffer from temporal phase conflicts in visual key-value (KV) caches due to local positional encodings. To overcome this limitation without requiring retraining, the authors propose an efficient approach that reconstructs the KV cache using a global triaxial multimodal RoPE coordinate system, preserving spatiotemporal structural consistency and avoiding spatial-temporal confusion caused by one-dimensional reindexing. Additionally, the method selectively recomputes high-bias visual tokens to compensate for missing contextual information. Notably, this is the first work to employ three-dimensional RoPE for cross-segment temporal alignment, achieving significant performance gains over baselines on Qwen2.5-VL-3B and TempCompass, with improved event ordering accuracy and 3.3× faster inference compared to full re-prefilling.

0 citationsRead paper

BLUE: A Stale-Pixel Optical-Flow Compositor for Entropy-Efficient Surveillance Video Encoding

Jun 27, 2026

This work addresses the inefficiency in storing static surveillance videos caused by repeatedly encoding invariant backgrounds. The authors propose BLUE, a pre-encoding synthesizer that integrates background pixel substitution with synthesis prior to compression: by maintaining a background seed frame, it replaces corresponding regions in input frames with static pixels, prompting the encoder to generate near-zero-bitrate SKIP macroblocks while preserving foreground visual fidelity. A lightweight VMAF-aware activation gating mechanism is introduced to balance compression gains against perceptual quality. Compatible with both x264 and x265 encoders, BLUE achieves average bitrate savings of 34.6% (x264) and 39.4% (x265) on the VIRAT dataset at CRF 28, improves foreground PSNR in 60.7% of clips, and incurs negligible full-frame quality degradation (ΔVMAF ≤ −2.9) in 40% of cases.

0 citationsRead paper