A two-week project conducted during the Redwood Research internship found that Qwen 3.8 27B can enhance CoT controllability using reward laundering. Experiments showed the model can generalize skills to unseen tasks while maintaining performance on the main task.
The chart will appear after repeat observations. The current metric comes from the source.
8 source pointsReal observations only. History before source connection is not reconstructed.
The study shows how models can use reward laundering techniques to improve specific capabilities, highlighting the importance of understanding their behavior.
The article discusses the Piltdown man hoax, a fraudulent artifact that misled evolutionary biology for decades. It draws a parallel between the Piltdown man and the Voynich manuscript, suggesting the latter might also be a meaningless fabrication. The author argues that while there is evidence supporting the idea that the Voynich manuscript is a hoax, it does not meet modern standards for proving fraud.
6 days agoLessWrongThe article discusses the contradiction between theoretical arguments about the information inefficiency of RL and practical results where RL shows high effectiveness in training language models.
yesterdayLessWrongThe author investigated training GPT-2 with full-batch gradient descent on a subset of FineWeb data. Results showed that the full-batch approach achieved lower training loss but worse generalization on validation data compared to standard training.
2 days agoLessWrongThe article describes the efforts of the author and his friend Mike to automate the process of transferring data from a website to a spreadsheet using AI. Mike, who works in data processing, spends significant time manually entering information, which the author considers inefficient.
5 days agoLobstersThe source text raises a question about whether Ente Locker copied code without proper attribution.
4 days agoHacker NewsOpenCode 2.0 is a project related to open source code.
2 days ago