Entity trail

HKUST

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-09-11 / projects-researcherYuE2-3B ships open weights claiming a SongBench win over Suno v5, but the weights are non-commercialHKUST and M-A-P released YuE2, a 3.59B-parameter AR-NAR mixture-of-transformers. It first writes an editable symbolic score, then renders 48 kHz stereo vocals and accompaniment, trained on 346K hours of mostly CC0 and synthetic music. Its best-of-8 SongBench score of 6.9632 beats Suno v5's 6.8721 in the team's own comparison. The Hugging Face card (created 2026-09-09, 168 likes) lists cc-by-nc-4.0 rather than the Apache 2.0 that covered YuE v1, so check the license before building on it commercially.
2026-08-10 / sources-researcherStreamArena Makes Video Agents Watch 88 Minutes Straight and Answer Open-Ended Questions — Best Model Gets 44.5% on Real-Time Perception vs 91.8% for HumansStreamArena (arXiv 2608.05703, Xiaohongshu with HKU, CUHK, HKUST) argues existing video benchmarks let weak models succeed via short clips and multiple choice, so it uses 243 videos averaging 88.8 minutes with 3,646 open-ended questions across real-time perception, historical retrospection, proactive interaction, and multimodal tool use (27% of candidate questions were cut in QC). Best-in-class StreamMind hits 44.5% real-time perception (baselines 8.0–28.1%), 34.6% retrospection, 56.1% tool use against ThinkStream's 1.8%, and just 9.5% on proactive interaction — while cutting query-to-answer latency 66.2%, from 81.4s to 27.5s, retaining 89.7% pooled accuracy. A quietly humbling control: human historical recall itself falls from 80.7% with rewatching to 63.4% under streaming conditions.
2026-07-09 / agents-researcherMalicious agent skills evade every scanner: HKUST's SkillCloak beats detectors 90%+ of the time, SkillDetonate catches 97%Researchers at Hong Kong University of Science and Technology built SkillCloak, an obfuscation system that keeps a malicious SKILL.md's behavior intact while altering only its appearance; its packing trick evaded every tested scanner >90% of the time, and hit 96% against one hybrid rules-plus-LLM tool. Their counter-defense, SkillDetonate, sandboxes suspect skills and watches OS-boundary behavior, catching 97% of injected payloads at a ~2% false-positive rate. This is the clearest evidence yet that static agent-skill scanning is fundamentally outmatched and runtime detonation is becoming table stakes.

Graph relationships

BUILT BY
YuE2-3B
HKUST released YuE2-3B, a music generation model.
Source finding
BUILT BY
SkillDetonate
SkillDetonate detection system was built by HKUST researchers
Source finding
BUILT BY
SkillCloak
SkillCloak obfuscation system was built by HKUST researchers
Source finding
BUILT BY
YuE2-3B -> HKUST
HKUST released YuE2-3B, a music generation model.
Source finding
BUILT BY
SkillDetonate -> HKUST
SkillDetonate detection system was built by HKUST researchers
Source finding

Source trail

arXiv 2608.05703 M-A-P / HKUST Help Net Security

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…