The LeJEPA Debate: From Doubt to LeCun's Clarification

Episode 1 · LeJEPA Feud: Berton Questions JEPA Results, LeCun Fires Back (2026-11-07, 8 posts)

A dispute over the self-supervised image learning method LeJEPA is brewing in the AI research community: researcher Gabriele Berton publicly stated that models with JEPA in the title generally underperform, that LeJEPA falls far short of DINO in practice—beautiful paper, elegant math, weak results—with Meta's V-JEPA 2.1 the only possible exception. The discussion drew Yann LeCun's attention and turned into a small-scale skirmish in the AI world.

Confirmed

• Berton and Google DeepMind researcher Luca Eyre (giffmana) both pointed out that existing comparisons are unfair: LeJEPA's public weights were trained only on ImageNet-1k or 21k, while DINO v2 used 140M high-resolution images, and the two sides have no aligned evidence.
• @SimonGoodman reported a fair comparison after re-tuning: LeJEPA still loses to JEM, and its compute-equivalence curves show it has clearly hit a plateau—hard to improve without changing the recipe; Berton believes LeJEPA plateaued earlier than JEM.
• Berton stressed that JEPA is a very broad term, and there is currently no evidence the method scales well to large datasets.
• A cited French PhD student speculated that his advisor used the word JEPA in the paper hoping to get a retweet from LeCun.

Why It Matters

• The debate strikes at the actual evidence for scaling in the JEPA line (one of the world-model directions LeCun champions): if it can't reliably beat baselines like DINO and JEM even on small-scale ImageNet, its prospects on large datasets look worrying.
• Fair-comparison methodology (compute equivalence, consistent tuning) was also emphasized in the discussion, reminding the community to check whether experimental setups are truly comparable when interpreting papers on new methods

• Researcher: JEPA-titled models are usually bad in practice, V-JEPA 2.1 the only exception — serrjoa · 2026-11-07
• Researchers question whether LeJEPA's method scales beyond ImageNet-sized data — gabriberton · 2026-11-07
• giffmana questions LeJEPA: no scaling law evidence vs DINO — giffmana · 2026-11-07
• Researcher claims JEPA-titled models underperform DINO, mocks LeCun shoutout chasing — gabriberton · 2026-11-07
• Retuned LeJEPA Still Loses to JEM; Scaling Comparison Draws LeCun Into SSL Debate — gabriberton · 2026-11-08
• JEPA debate: Berton says LeJEPA plateaus sooner than JEM, demands scaling laws — gabriberton · 2026-11-08
• LeCun defends JEPA: gap from transformer to ChatGPT was longer than JEPA's age — ylecun · 2026-11-08
• LeCun defends JEPA: only 3 years since 2022 paper, less than Transformer-to-ChatGPT gap — gabriberton · 2026-11-08

Episode 2 · LeCun Clarifies JEPA-DINO Relationship in Long Post (2026-11-08, 2 posts)

Yann LeCun posted a long clarification on JEPA, arguing that DINO is essentially a JEPA and that most JEPA papers are minor variants, stressing that FAIR's foundational research predates its scaling.

• LeCun clarifies: DINO is a JEPA variant and most JEPA papers are just fine-tunes — ylecun · 2026-11-08
• LeCun breaks down JEPA vs DINO technical differences, defends foundational research — May_F1_ · 2026-11-08