AI
Artificial intelligence: LLMs, agents, diffusion, vision, NLP and the latest from top labs.
593 published articles
AI MUSIC
The melody-first trick that lets Qwen-Music beat Suno on 13 metrics
Qwen-Music separates melody planning from full-song generation via Melody-CoT. The 3-billion-parameter model, trained on 5 million+ hours of multilingual data, achieves state-of-the-art results against Suno V5 and MiniMax Music 2.6 in both objective metrics and human preference.
2026-07-31
Medical AI Research
ClinPRISM: a 4B-parameter model answers irregular clinical data in 0.15 seconds
ClinPRISM is a cost-effective multimodal LLM reasoning framework designed for question answering over irregular clinical time series. Using just 16 tokens and a 4B-parameter backbone, it achieves state-of-the-art performance with an average latency of 0.15 seconds per question.
2026-07-31
AI Strategy
China's open-weight gambit splits Silicon Valley
As Chinese open-weight models like Kimi K3 match US frontier systems on benchmarks while being given away for free, Silicon Valley is split: a 41-company coalition defends openness while Anthropic and others hold back, and a rogue-model safety incident only sharpens the divide.
2026-07-31
GUI Agents
Alibaba's Qwen-UI-Agent scores higher on real phones than in sandboxes
Alibaba Tongyi Lab's Qwen-UI-Agent claims state-of-the-art mobile scores (97.5% AndroidDaily) and competitive computer-use results against Opus 4.8, Gemini 3.1 Pro, and GPT-5.6 Sol. Its real-device benchmark score beats its sandbox score, while 40% partial progress on OSWorld-v2 is the honest limit.
2026-07-31
LLM distillation
LLM distillation stalls when context goes static. Flux-OPD makes it evolve
Contexts can carry task preferences into LLM training, but once distilled into a student they add little supervision. Peking University's Flux-OPD keeps the context moving with the student and uses a conflict term to weight teacher corrections. The abstract reports gains over existing OPD paradigms without naming numbers.
2026-07-31
Frontis-MA1 / OpenMLE
The AI that improves AI now runs on one RTX 4090
FrontisAI's OpenMLE stack and 35B Frontis-MA1 agent turn recursive self-improvement into a reproducible experiment. On a single RTX 4090, the model scores 71.21% on MLE-Bench Lite, ahead of GPT-5.5 + Codex and close to much larger frontier models.
2026-07-31
AI research
Over half your AI's reasoning is froth, and nobody noticed until now
LLMs often generate reasoning chains that are correct but padded with unnecessary steps. A new diagnostic benchmark shows current evaluators miss this inefficiency entirely, and half of human-written reasoning steps may be compressible.
2026-07-31
AI Research
Memory Decoder: a 6.9B bolt-on memory makes a 410M model beat Pythia-12B
Memory Decoder at Scale splits long-term memory from reasoning in LLMs. A 6.9B pretrained memory lifted Pythia-410M past Pythia-12B on 17 benchmarks with 39% fewer parameters; 1.7B domain memories added 9+ points to Qwen3 Base at every scale tested.
2026-07-31
Scientific literature search goes claim-level
Chemists search papers; AskChem answers with 2.4 million claims
AskChem indexes 2.4 million claims from 147,000 papers, each grounded in a source DOI and a verbatim quote, served to scientists and AI agents via web app, REST, SDK, and MCP. On AskChem-Bench, DOI resolvability for a GPT-5.5 reader went from 88.3% to 100%.
2026-07-31
Research Digest
The simple topology trick that beats prompt magic in LLM induction tasks
A new methodology called Hourglass reasoning enforces strict context isolation between induction, deduction, and implementation stages, passing only a compressed symbolic rule between them. On ARC-AGI-2, it raises best-of-5 accuracy by up to 14 points over iterative refinement; on ChipBench Verilog synthesis, it nearly doubles accuracy with GPT-5.5. Ablations confirm the topology itself drives the improvement.
2026-07-31
Computer Vision, World Models & AI Research
PhiZero: teaching video AI to think in physics before it renders
PhiZero, a CASIA world model, learns a compact discrete "physical language" from raw video and uses it to reason about how a scene will evolve before rendering frames. The authors argue this reason-then-render design produces more physically coherent video than direct pixel prediction.
2026-07-31
AI Research
Video generation gets 120x faster on one GPU, inside Nvidia's hybrid attention breakthrough
Nvidia Research's SANA-Video 2.0 combines linear and periodic softmax attention in a 3:1 ratio to run 120x faster than Wan 2.2 on one H100. The hybrid design recovers full-rank expressiveness while keeping inference at 13 seconds for a 720p clip. The caveats: the benchmarks stop at 720p and bundle the attention change with proprietary optimizations.
2026-07-31