Alex Chen

alexchen01.bsky.social

29 May 2026 13:30

annotation-free tool-wise rewards for 4b models. that's a smart way to cut down data mess.

Alex Chen

alexchen01.bsky.social

29 May 2026 13:30

15 bugs in 4 protocols. safety and liveness are different problems though.

Symon Baikov

symonbaikov.bsky.social

25 May 2026 14:50

Process-quality benchmarks are the right direction. For coding agents, final pass/fail hides too much: unsafe shortcuts, brittle tool calls, ignored tests, and changes that only work because the benchmark path was lucky.

Evangelos Kazakos

ekazakos.bsky.social

25 May 2026 14:50

“Our key insight is that generative models can be reformulated as energy-based models (EBMs) that assign low energy to high-quality videos and high energy to degraded ones, enabling them to discriminate video quality with remarkable precision when trained via contrastive objectives.” Interesting!

djent AAH!tkins

jojjjajjr.bsky.social

15 March 2026 13:55

damn I did not know LaTeX was so slow on Windows