BEGIN:VCALENDAR
VERSION:2.0
CALSCALE:GREGORIAN
PRODID:adamgibbons/ics
METHOD:PUBLISH
X-PUBLISHED-TTL:PT1H
BEGIN:VEVENT
UID:Ic7Rf2M9S6US3PIw_-9Un
SUMMARY:Scaling LLM-RL for the age of agents
DTSTAMP:20260513T113355Z
DTSTART:20260522T114500Z
DESCRIPTION:Description:\nReinforcement learning (RL) has emerged as the pr
	imary paradigm for scaling base models into autonomous agents across softw
	are engineering\, research\, and beyond. Yet scaling RL poses fundamentall
	y different challenges than pre-training. Its asynchronous and online natu
	re introduces new problems\, from curriculum design and synthetic task gen
	eration to off-policy training dynamics and large-scale sandboxed rollout 
	execution. In this talk\, Konstantin Dunas traces how RL scaling has evolv
	ed from early approaches and why this shift creates new infrastructural bo
	ttlenecks. He will examine challenges such as training stability\, trainin
	g–inference policy mismatch\, and off-policy learning\, and how prime-rl\,
	 Prime Intellect’s open-source training framework\, enables others to appl
	y these approaches to train their own agents. The talk will also outline w
	hat comes next for large-scale LLM-RL.\n--------------------------------\n
	\nSpeaker:\n- Konstantin Dunas\n\n--------------------------------\n\nTalk
	 details:\n- Link to the Big Techday website: https://bigtechday.com/en/ta
	lks#7Bzb87sZbHksbnpcYOf6fK\n
LOCATION:Stellwerk
DURATION:PT50M
END:VEVENT
END:VCALENDAR