Non-record: JEPA-LM — When Synthetic Success Doesn't Transfer to Real…#1012
Open
himanshudongre wants to merge 2 commits intoopenai:mainfrom
Open
Non-record: JEPA-LM — When Synthetic Success Doesn't Transfer to Real…#1012himanshudongre wants to merge 2 commits intoopenai:mainfrom
himanshudongre wants to merge 2 commits intoopenai:mainfrom
Conversation
… Language Implements JEPA (Joint Embedding Predictive Architecture) as training-time auxiliary loss for language modeling. On synthetic Markov chain data, JEPA showed -19.5% cross-entropy improvement. On real English text, the improvement collapsed to -0.24% with +40% throughput overhead. Key finding: Markov chains have exploitable repetitive statistical structure that JEPA excels at, but natural language doesn't. This is a cautionary tale about synthetic benchmark validation. Checks off "JEPA" from Requests for PRs. See companion PR (S4D-Lin SSM Hybrid) for where this research led next.
himanshudongre
added a commit
to himanshudongre/parameter-golf
that referenced
this pull request
Mar 28, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Why This Matters
This is a negative result, but an informative one. Understanding why it doesn't work at this scale is as valuable as showing something that does.
The Trap: Synthetic vs Real Text
The synthetic result was completely misleading. Markov chains have fixed transition probabilities that JEPA's representation prediction can exploit. Natural language has semantic ambiguity, long-range dependencies, and non-stationary statistics — JEPA's predictions become nearly meaningless.
How JEPA-LM Works
Standard LM:
tokens → encoder → LM head → CE lossJEPA-LM adds:
tokens → target encoder (EMA) → target representationsPlus:
tokens → online encoder → predictor → predicted representations → JEPA lossThe target encoder is training-only (discarded at export). Zero eval-time overhead. But the training overhead (+40%) costs ~1500 training steps in 600s, which a -0.24% quality improvement cannot justify.
Companion PR
See S4D-Lin SSM Hybrid PR for where this research led next. The key lesson from JEPA (always validate on real text) informed the SSM approach — though even real-text validation at small scale turned out to be unreliable.
See full README