Premium Only Content
ActInf GuestStream 115.1 ~ Energy-Based Transformers and the Future of Scaling
"Energy-Based Transformers are Scalable Learners and Thinkers"
Alexi Gladstone, Ganesh Nanduru, Md Mofijul Islam, Peixuan Han, Hyeonjeong Ha, Aman Chadha, Yilun Du, Heng Ji, Jundong Li, Tariq Iqbal
https://arxiv.org/abs/2507.02092
Inference-time computation techniques, analogous to human System 2 Thinking, have recently become popular for improving model performances. However, most existing approaches suffer from several limitations: they are modality-specific (e.g., working only in text), problem-specific (e.g., verifiable domains like math and coding), or require additional supervision/training on top of unsupervised pretraining (e.g., verifiers or verifiable rewards). In this paper, we ask the question "Is it possible to generalize these System 2 Thinking approaches, and develop models that learn to think solely from unsupervised learning?" Interestingly, we find the answer is yes, by learning to explicitly verify the compatibility between inputs and candidate-predictions, and then re-framing prediction problems as optimization with respect to this verifier. Specifically, we train Energy-Based Transformers (EBTs) -- a new class of Energy-Based Models (EBMs) -- to assign an energy value to every input and candidate-prediction pair, enabling predictions through gradient descent-based energy minimization until convergence. Across both discrete (text) and continuous (visual) modalities, we find EBTs scale faster than the dominant Transformer++ approach during training, achieving an up to 35% higher scaling rate with respect to data, batch size, parameters, FLOPs, and depth. During inference, EBTs improve performance with System 2 Thinking by 29% more than the Transformer++ on language tasks, and EBTs outperform Diffusion Transformers on image denoising while using fewer forward passes. Further, we find that EBTs achieve better results than existing models on most downstream tasks given the same or worse pretraining performance, suggesting that EBTs generalize better than existing approaches. Consequently, EBTs are a promising new paradigm for scaling both the learning and thinking capabilities of models.
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Computer Vision and Pattern Recognition (cs.CV)
Cite as: arXiv:2507.02092 [cs.LG]
(or arXiv:2507.02092v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2507.02092
Active Inference Institute information:
Website: https://www.activeinference.institute/
Activities: https://activities.activeinference.institute/
Discord: https://discord.activeinference.institute/
Donate: http://donate.activeinference.institute/
YouTube: https://www.youtube.com/c/ActiveInference/
X: https://x.com/InferenceActive
Active Inference Livestreams: https://video.activeinference.institute/
-
54:40
Active Inference Institute
12 days agoActInf GuestStream 122.1 ~ Radical Relation, Scaling without Hierarchies, Models without the Meat
18 -
LIVE
LFA TV
11 hours agoLIVE & BREAKING NEWS! | WEDNESDAY 10/29/25
8,250 watching -
1:24:42
Crypto Power Hour
11 hours agoOG Anonymous & NFT Expert Jeremy Ryan aka NFT Demon
28.5K6 -
26:01
Lady Decade
17 hours agoIs the Atari Jaguar worth playing in 2025 !?
11.8K8 -
11:08
ThinkStory
11 hours agoWELCOME TO DERRY Episode 1 Biggest Theories & Hidden Details
13K2 -
4:33
Buddy Brown
17 hours ago $0.01 earnedWatch EBT Users Threaten to Come to Our Homes & STEAL FOOD on Thanksgiving! | Buddy Brown
12.1K17 -
LIVE
BEK TV
23 hours agoTrent Loos in the Morning - 10/29/2025
220 watching -
15:09
BlackDiamondGunsandGear
11 hours agoTop 5 Pistols UNDER $400
56.5K15 -
LIVE
Flex011
6 hours agoCan We All Survive the Night? 😱 | UNTIL DAWN Horror Adventure LIVE!
157 watching -
8:31
Hollywood Exposed
14 hours agoJoe Rogan Just EXPOSED Who’s REALLY Behind the “No Kings” Protests
18K4