#SongZe
LightNews LightNews
arxiv-cs-cv.bsky.social
Bojun Yang, Haochen Zhou, Zhifang Zhang, Haobo Wang, Songze Li, Lei Feng
Purifying Backdoored Large Vision-Language Models by Removing Hijacked Directions
https://arxiv.org/abs/2610.09941
October 10, 2026 at 7:48 AM
cscv-bot.bsky.social
Bojun Yang, Haochen Zhou, Zhifang Zhang, Haobo Wang, Songze Li, Lei Feng: Purifying Backdoored Large Vision-Language Models by Removing Hijacked Directions https://arxiv.org/abs/2610.09941 https://arxiv.org/pdf/2610.09941 https://arxiv.org/html/2610.09941
October 8, 2026 at 6:41 AM
ultraeric.bsky.social
A street view on Songze Road

Shot on July 3rd, 2026

#shanghai
October 2, 2026 at 4:03 AM
csma-bot.bsky.social
Yapeng Li, Songze Li, Shuang Yu, Jing Yu, Zhixin Liu, Liqiang Wen, Tonghua Su: MASTraceBench: Diagnosing Collaboration Gains through Proposal Trajectories in LLM-Based Multi-Agent Systems https://arxiv.org/abs/2609.34496 https://arxiv.org/pdf/2609.34496 https://arxiv.org/html/2609.34496
September 29, 2026 at 6:42 AM
dailyrunway.bsky.social
chen duling for cosmopolitan china 33rd anniversary issue, photographed by tu songze
August 7, 2026 at 11:19 PM
nanoatp.bsky.social
[28/30] 187 Upvotes, 2 Comments, 2 Posts, arXiv:2607.17423

🆕TimeLens2: Generalist Video Temporal Grounding with Multimodal LLMs

Yuhan Zhu, Changlian Ma, Xiangyu Zeng, Xinhao Li, Zhiqiu Zhang, Songze Li
Video multimodal large language models (MLLMs) can describe what happens in a video, but rarely identify when the supporting evidence occurs. We study generalist video temporal grounding, in which one model predicts a variable-cardinality set of evidence intervals across video lengths, domains, query forms, and viewpoints. Existing training strategies are misaligned with this set-valued task: long-video labels often rely on brittle one-pass annotation, while reinforcement-learning rewards either fail to distinguish non-overlapping predictions or require fragile segment matching. TimeLens2 treats temporal evidence as an interval set throughout supervision and optimization. TimeLens2-93K constructs reliable multi-span supervision through caption-derived proposals, independent localization, cross-agent consensus, semantic verification, and boundary refinement. Our temporal Wasserstein reward computes exact one-dimensional \(W_1\) between uniform distributions over merged interval supports, providing dense, matching-free feedback under unequal cardinalities and equivalent fragmentation; temporal IoU complements it with precise-overlap feedback. Across seven benchmarks, TimeLens2-2B outperforms all size-matched baselines on every benchmark, while the 4B and 8B variants achieve state-of-the-art performance, surpassing open-source models with up to 397B parameters. The 2B, 4B, and 8B variants improve over their Qwen3-VL backbones by 14.2, 13.0, and 18.1 mIoU points, respectively.
July 24, 2026 at 12:04 AM
nanoatp.bsky.social
[29/30] 179 Upvotes, 2 Comments, 2 Posts, arXiv:2607.17423

🆕TimeLens2: Generalist Video Temporal Grounding with Multimodal LLMs

Yuhan Zhu, Changlian Ma, Xiangyu Zeng, Xinhao Li, Zhiqiu Zhang, Songze Li
Video multimodal large language models (MLLMs) can describe what happens in a video, but rarely identify when the supporting evidence occurs. We study generalist video temporal grounding, in which one model predicts a variable-cardinality set of evidence intervals across video lengths, domains, query forms, and viewpoints. Existing training strategies are misaligned with this set-valued task: long-video labels often rely on brittle one-pass annotation, while reinforcement-learning rewards either fail to distinguish non-overlapping predictions or require fragile segment matching. TimeLens2 treats temporal evidence as an interval set throughout supervision and optimization. TimeLens2-93K constructs reliable multi-span supervision through caption-derived proposals, independent localization, cross-agent consensus, semantic verification, and boundary refinement. Our temporal Wasserstein reward computes exact one-dimensional \(W_1\) between uniform distributions over merged interval supports, providing dense, matching-free feedback under unequal cardinalities and equivalent fragmentation; temporal IoU complements it with precise-overlap feedback. Across seven benchmarks, TimeLens2-2B outperforms all size-matched baselines on every benchmark, while the 4B and 8B variants achieve state-of-the-art performance, surpassing open-source models with up to 397B parameters. The 2B, 4B, and 8B variants improve over their Qwen3-VL backbones by 14.2, 13.0, and 18.1 mIoU points, respectively.
July 23, 2026 at 12:05 AM
arxiv-cs-cv.bsky.social
Yuhan Zhu, Changlian Ma, Xiangyu Zeng, Xinhao Li, Zhiqiu Zhang, Songze Li, Jun Zhang, Tianxiang Jiang, Yuandong Yang, Ziang Yan, Zikang Wang, Xinyu Chen, Haoran Chen, Shaowei Zhang, Limin Wang
TimeLens2: Generalist Video Temporal Grounding with Multimodal LLMs
https://arxiv.org/abs/2607.17423
July 21, 2026 at 10:25 PM
cscr-bot.bsky.social
Licheng Yu, Aiwei Liu, Songze Li: Toward Stronger Code Watermarking: A Grammar-Driven Approach to Optimizing the Trade-off Between Quality and Detectability https://arxiv.org/abs/2607.10210 https://arxiv.org/pdf/2607.10210 https://arxiv.org/html/2607.10210
July 14, 2026 at 6:39 AM
arxiv-daily-bot.bsky.social
Toward Stronger Code Watermarking: A Grammar-Driven Approach to Optimizing the Trade-off Between Quality and Detectability

Licheng Yu, Aiwei Liu, Songze Li

#arXiv #cs.CR #cs.CL
Toward Stronger Code Watermarking: A Grammar-Driven Approach to Optimizing the Trade-off Between Quality and Detectability
With the rapid development of Large Language Models (LLMs), text watermarking has emerged as a crucial technique for identifying machine-generated content. However, directly applying existing logits-based watermarking methods to code generation remains challenging, since the low-entropy nature of c…
arxiv.org
July 14, 2026 at 5:23 AM
oyatsumaster.bsky.social
2026年7月4日
#考古学のおやつ 1/2
中国、江蘇省（Jiangsu） 蘇州市（Suzhou） 呉中区（Wuzhong）・塘北遺跡（塘北遗址：Tangbei）の、崧沢文化（Songze）～西周（Western Zhou）時代・紀元前4000年～前700年の土器片141点を分析。
鉄分に富み石灰分を含まない在地の粘土を常に使用。混和材は崧沢文化では植物。馬橋文化（Maqiao culture）以降は、特に調理用の土器では砂の混和材が支配的。印文のある硬陶はシリカに富んだ胎土を用い、高温で焼成。→
July 4, 2026 at 1:04 PM
arxiv-cs-cr.bsky.social
Chenqing Zhu, Yanbo Dai, Yulong Tian, Qingming Li, Songze Li
When the Aggregator Cheats: Data-Free Backdoors in Federated LLM-based QA Systems
https://arxiv.org/abs/2606.27511
June 29, 2026 at 8:10 PM
cscr-bot.bsky.social
Chenqing Zhu, Yanbo Dai, Yulong Tian, Qingming Li, Songze Li: When the Aggregator Cheats: Data-Free Backdoors in Federated LLM-based QA Systems https://arxiv.org/abs/2606.27511 https://arxiv.org/pdf/2606.27511 https://arxiv.org/html/2606.27511
June 29, 2026 at 6:39 AM
physicsoptics-bot.bsky.social
Yu Wang, Benquan Wang, Eng Aik Chan, Songze Li, Zhenyu Guo, Xi Xie, Masako Kishida, Che-Chin Chen, Yijie Shen, Jun-Yu Ou: Alignment-Free Nanometric Optical Metrology Enabled by Structured Light https://arxiv.org/abs/2606.22366 https://arxiv.org/pdf/2606.22366 https://arxiv.org/html/2606.22366
June 23, 2026 at 6:59 AM
cscr-bot.bsky.social
Chunjiang Che, Songze Li, Xuechao Wang: OptChain: Achieving Optimal Throughput of Permissionless Blockchains https://arxiv.org/abs/2606.23405 https://arxiv.org/pdf/2606.23405 https://arxiv.org/html/2606.23405
June 23, 2026 at 6:40 AM
cslg-bot.bsky.social
Zhitong Wang, Songze Li, Hao Peng, Shuzheng Si, Yi Wang, Maosong Sun, Juanzi Li: EnvRL: Learn from Environment Dynamics in Agentic Reinforcement Learning https://arxiv.org/abs/2606.17680 https://arxiv.org/pdf/2606.17680 https://arxiv.org/html/2606.17680
June 17, 2026 at 6:43 AM
cscr-bot.bsky.social
Jin Xie, Songze Li: OCELOT: Inference-Leakage Budgets for Privacy-Preserving LLM Agents https://arxiv.org/abs/2606.12341 https://arxiv.org/pdf/2606.12341 https://arxiv.org/html/2606.12341
June 11, 2026 at 6:40 AM
arxiv-cs-cv.bsky.social
Tianxiang Jiang, Linquan Wu, Sheng Xia, Songze Li, Ziang Yan, Haoyu Yang, Yu Qiao, Yi Wang
Imagine Before You Predict: Interleaved Latent Visual Reasoning for Video Event Prediction
https://arxiv.org/abs/2606.05769
June 6, 2026 at 12:35 AM
cscv-bot.bsky.social
Tianxiang Jiang, Linquan Wu, Sheng Xia, Songze Li, Ziang Yan, Haoyu Yang, Yu Qiao, Yi Wang: Imagine Before You Predict: Interleaved Latent Visual Reasoning for Video Event Prediction https://arxiv.org/abs/2606.05769 https://arxiv.org/pdf/2606.05769 https://arxiv.org/html/2606.05769
June 5, 2026 at 6:40 AM
cscr-bot.bsky.social
Mingxuan Zhang, Jiahui Han, Dadi Guo, Songze Li, Guanchu Wang, Na Zou, Dongrui Liu, Xia Hu: PrivacyPeek: Auditing What LLM-Based Agents Acquire, Not Just What They Say https://arxiv.org/abs/2606.00152 https://arxiv.org/pdf/2606.00152 https://arxiv.org/html/2606.00152
June 2, 2026 at 6:39 AM
cscr-bot.bsky.social
Kai Wang, Jiale Zhang, Chengcheng Zhu, Chuang Ma, Songze Li: Awakening the Hydra: Stabilizing Multi-Concept Backdoor Injection in Text-to-Image Diffusion Models https://arxiv.org/abs/2605.19698 https://arxiv.org/pdf/2605.19698 https://arxiv.org/html/2605.19698
May 20, 2026 at 6:40 AM
arxiv-cs-cr.bsky.social
Zehan Sun, Dingfan Chen, Songze Li
When Efficiency Backfires: Cascading LLMs Trigger Cascade Failure under Adversarial Attack
https://arxiv.org/abs/2605.17288
May 20, 2026 at 12:36 AM
cscr-bot.bsky.social
Zehan Sun, Dingfan Chen, Songze Li: When Efficiency Backfires: Cascading LLMs Trigger Cascade Failure under Adversarial Attack https://arxiv.org/abs/2605.17288 https://arxiv.org/pdf/2605.17288 https://arxiv.org/html/2605.17288
May 19, 2026 at 6:41 AM
cscv-bot.bsky.social
Jiacheng Chen, Songze Li, Han Fu, Baoquan Zhao, Wei Liu, Yanyan Liang, Li Qing, Xudong Mao: Delta-Adapter: Scalable Exemplar-Based Image Editing with Single-Pair Supervision https://arxiv.org/abs/2605.07940 https://arxiv.org/pdf/2605.07940 https://arxiv.org/html/2605.07940
May 11, 2026 at 6:42 AM
cscr-bot.bsky.social
Runhe Wang, Li Bai, Haibo Hu, Songze Li: Cross-Modal Backdoors in Multimodal Large Language Models https://arxiv.org/abs/2605.07490 https://arxiv.org/pdf/2605.07490 https://arxiv.org/html/2605.07490
May 11, 2026 at 6:40 AM