Aviral Kumar on X: "🚨🚨 New paper on flow-matching value functions Last year, we showed training RL value functions with a flow-matching loss achieved SOTA results. But why does it work? And what could it possibly tell us about other things that have nothing to do with VFs or even RL? Short https://t.co/x5trTotUO4" / X

To view keyboard shortcuts, press question mark View keyboard shortcuts Home Explore 1 Notifications Chat Grok Premium Bookmarks Creator Studio Articles Profile More Post sean lee @infinitefun_ Post See new posts Conversation Aviral Kumar @aviral_kumar2 New paper on flow-matching value functions Last year, we showed training RL value functions with a flow-matching loss achieved SOTA results. But why does it work? And what could it possibly tell us about other things that have nothing to do with VFs or even RL? Short answer: iterative compute used correctly can address feature plasticity in continual learning! 1:38 PM · Mar 5, 2026 · 59.3K Views 3 79 481 397 Relevant View quotes Post your reply Reply Aviral Kumar @aviral_kumar2 · Mar 5 Background: last year we developed floq https:// x.com/aviral_kumar2/ status/1965396309377966490 …. floq uses flow matching to make scalar predictions. Key idea: Transform state/action scalar value with an integration process that integrate

@aviral_kumar2: New paper on flow-matching value functions Last year, we showed training RL value functions with a flow-matching loss achieved SOTA results. But why does it work? And what could it possibly tell us about other things that have nothing to do with VFs or even RL? Short answer: iterative compute used correctly can address feature plasticity in continual learning! @aviral_kumar2: Background: last year we developed floq https:// x.com/aviral_kumar2/ status/1965396309377966490 …. floq uses flow matching to make scalar predictions. Key idea: Transform state/action scalar v

saved by

• Sean Lee

related reading

• Seohong Park on X: "We scaled up an "alternative" paradigm in RL: *divide and conquer*. Compared to Q-learning (TD learning), divide and conquer can naturally scale to much longer horizons. Blog post: https://t.co/xtXBzya0bI Paper: https://t.co/nqYkLucsWu ↓ https://t.co/XCdgUzaLxF" / Xx.com
• Flow Matching Policy Gradientsflowreinforce.github.io
• How I Understand Flow Matchingyoutube.com
• You can’t imitation-learn how to continual-learn — LessWronglesswrong.com
• Feature-wise transformationsdistill.pub
• The Continual Learning Problemjessylin.com
• Is Frontier Asynchronous RL Solved? — Luke J. Huangluk-huang.github.io
• RL is even more information inefficient than you thoughtdwarkesh.com
• [2510.13786] The Art of Scaling Reinforcement Learning Compute for LLMsarxiv.org
• The 4-bitter Lesson | humans&humansand.ai
• In (highly contingent!) defense of interpretability-in-the-loop ML training — AI Alignment Forumalignmentforum.org
• Features as Rewards: Using Interpretability to Reduce Hallucinationsgoodfire.ai