Researchers introduced GRAFT, a framework that exchanges trajectories between different models to improve reinforcement learning. It boosts performance by an average of 2.1 points across five benchmarks. ·
Researchers found that latent world action models lose generalization benefits when they skip future generation during inference. They proposed Simple-WAM, which uses a single forward pass to retain these advantages. ·
Anthropic's red team found that GLM-5.3 and Claude Mythos Preview achieved full control flow hijacks in 4% and 6% of 100 binary exploitation tasks, respectively, while earlier models succeeded in none. ·
Researchers introduced GRAFT, a framework that exchanges trajectories between different models to improve reinforcement learning. It boosts performance by an average of 2.1 points across five benchmarks. ·
Researchers found that latent world action models lose generalization benefits when they skip future generation during inference. They proposed Simple-WAM, which uses a single forward pass to retain these advantages. ·
The source text lists titles for visual guides and experiments on recurrent neural networks, convolutional neural networks, and transformers, focusing on their accuracy and efficiency. ·
Hacker News+ Ahead of AI16h ago▲ 106discussmagazine.sebastianraschka.com
Researchers introduced GRAFT, a framework that exchanges trajectories between different models to improve reinforcement learning. It boosts performance by an average of 2.1 points across five benchmarks. ·
HF Daily Papers3h ago▲ 89arxiv.org
Trending topics today: heat
Table
OpenAI
4.37
Anthropic
2.99
GLM-5.3
1.69
GPT-6.1
1.26
Language
1.22
Muse
1.21
Heat = summed hotness of today's stories naming the topic.
Models
5 stories
Releases, papers, trending weights and new listings
Top 6 trending on Hugging Face: likes
Table
convaiinnovations/laya
4522
Qwen/Qwen-Image-2.1 ▼1
2659
abenzerps/Qwen-Image-2.1-Uncensored-GGUF ▼1
2448
Edge0/Audio8-ASR-Infinite ▲1
1543
XingChen-AGI/TeleOCR ▲1
877
Contrastive-LM/CLM-v0.1-8B ▲1
531
PirateBench: score out of 100
Table
opus-5-5
91.15
gpt-6-sol
81.78
deepseek-v4-flash-0731
62.82
Agents build our own projects from a brief; every run, failures included: https://cowerx.dev/daily/piratebench/
Researchers introduced a thinking reward model that creates case-specific evaluation criteria before scoring visual outputs. They also developed a new optimization method to improve reward discrimination while maintaining detailed pointwise scoring. ·
Researchers introduced SAKI, a method routing token-level supervision in on-policy distillation using maximal coupling. It improves performance on mathematical reasoning benchmarks for small student models. ·
Researchers introduced WorldAttention, an architecture for interactive video world models. It combines hybrid sparse attention with hierarchical key-value cache management to reduce computational costs. ·
Researchers introduced MaLiang-Harness, a framework that uses multimodal models to iteratively build, check, and fix code for image and video generation. The system evaluates eleven closed-source models on new benchmarks to measure success and quality. ·
Researchers introduced OmniTaskonomy, a taxonomy covering 19 generation tasks and 25 understanding capabilities, to map how visual generation training improves specific visual understanding skills. ·
Researchers introduced a thinking reward model that creates case-specific evaluation criteria before scoring visual outputs. They also developed a new optimization method to improve reward discrimination while maintaining detailed pointwise scoring. ·
Researchers introduced SAKI, a method routing token-level supervision in on-policy distillation using maximal coupling. It improves performance on mathematical reasoning benchmarks for small student models. ·
Researchers introduced WorldAttention, an architecture for interactive video world models. It combines hybrid sparse attention with hierarchical key-value cache management to reduce computational costs. ·
Researchers introduced MaLiang-Harness, a framework that uses multimodal models to iteratively build, check, and fix code for image and video generation. The system evaluates eleven closed-source models on new benchmarks to measure success and quality. ·
The source text lists titles for visual guides and experiments on recurrent neural networks, convolutional neural networks, and transformers, focusing on their accuracy and efficiency. ·
41 points and 1 comments on Hacker News, linking magazine.sebastianraschka.com.
Hacker News+ Ahead of AI16h ago▲ 106discussmagazine.sebastianraschka.com
Researchers introduced a thinking reward model that creates case-specific evaluation criteria before scoring visual outputs. They also developed a new optimization method to improve reward discrimination while maintaining detailed pointwise scoring. ·
Researchers introduced SAKI, a method routing token-level supervision in on-policy distillation using maximal coupling. It improves performance on mathematical reasoning benchmarks for small student models. ·
HF Daily Papers3h ago▲ 83arxiv.org
X voices
5 stories
X posts other sources are linking to, then named voices on Bluesky and Mastodon
Google released version 0.62.0 of the Gemini CLI. The update includes fixes for OAuth token handling, UI layout rendering, and MCP tool call formatting. ·
OpenAI released version 0.159.0 of Codex, adding an opt-in feature that allows new input to steer the model during active responses or long-running code calls. ·