@schaul.bsky.social, I was not clear from reading your paper on Socratic learning whether you envision that process as actually occurring through Socratic dialogue between different agents or just a single agent with itself. Are the games between many players? or solo? arxiv.org/abs/2411.16905 - ThreadSky

andrescorrada.bsky.social • 181 days ago

@schaul.bsky.social, I was not clear from reading your paper on Socratic learning whether you envision that process as actually occurring through Socratic dialogue between different agents or just a single agent with itself. Are the games between many players? or solo? https://arxiv.org/abs/2411.16905

Comments

schaul.bsky.social•181 days ago

Either one or many players. For "improvement" to be well-defined, one agent must be special (see footnote 6), but the multi-agent setting has many benefits.

andrescorrada.bsky.social•181 days ago

By "special" you mean the agent producing the language games used to evaluate the learning agent? If so, would it be possible to have them play games with each other? They all take turns playing each other's games - an academy where learners interact. Logical consistency could be guaranteed then.

andrescorrada.bsky.social•181 days ago

Your thought or gedanken experiment of a learner with no inputs is like the one for AI safety. A monitoring system is the paranoid detective that suspects everyone in the closed room. The detective susses out the suspects without witnessing the crime. How? Logical consistency of disagreeing agents.

andrescorrada.bsky.social•181 days ago

With a logic of unsupervised evaluation, intelligent agents could construct tests for each other. Say, multiple choice exams, although essay questions are fine too as I will soon make clear. Every test now has responses from all of them. They could check if any of them is misaligned with the group.

andrescorrada.bsky.social•181 days ago

The paper only treats binary classification or equivalently, multiple choice exams with two responses. But I have figured out the math/logic for 3 or more classes. The NTQR package implements it now as shown in this Jupyter documentation page. https://ntqr.readthedocs.io/en/latest/notebooks/LogicalAlarmsMisalignedClassifiers.html

andrescorrada.bsky.social•181 days ago

So essay exams are possible because you then add another grading layer where you ask the agents to grade their essays in whatever numerical scale you want - a multiple choice exam of their ability to grade themselves! Then use the logical alarms to chuck out any agent that is misaligned in grading.

Comments

Posting Rules

Reply