ER-JEPA: Experience Replay Improves Joint-Embedding Predictive Learning in Language Models

Jingnan Pu; Zi-En Fan; Feng Lian

cs.CLSep 29, 2026

ER-JEPA: Experience Replay Improves Joint-Embedding Predictive Learning in Language Models

Authors: Jingnan Pu, Zi-En Fan, Feng Lian

Organizations: School of Automation Science and Technology, Xi’an Jiaotong University No. 28, West Xianning Road, Xi’an, Shaanxi 710049, China

Abstract

Large language models (LLMs) excel at token-level generation but may learn undesirable abstract semantics and lack comprehensive perception. LLM-JEPA mitigates this by aligning different views of the same underlying knowledge via a joint-embedding predictive architecture (JEPA). However, strong alignment does not necessarily lead to accurate, stable predictions. To address this, we propose ER-JEPA, which adds an episodic replay path to LLM-JEPA. ER-JEPA stores training pairs in a memory. At each step, it stores and retrieves relevant data to provide additional supervision. This enables learning from both the current batch and stored training pairs, providing additional supervision for token prediction and representation alignment. Experiments across multiple datasets (NL-RX, GSM8K, Spider, and NQ-Open) demonstrate that ER-JEPA consistently outperforms LLM-JEPA.

Figures & tables

Figure 1 : ER-JEPA outperforms LLM-JEPA across tasks.

Figure 2 : Comparison of baseline training, LLM-JEPA, and ER-JEPA. (a) The baseline learns through next token prediction. (b) LLM-JEPA adds prediction in the embedding space to learn abstract relations between inputs and targets. (c) ER-JEPA adds a replay path to LLM-JEPA. It stores past training examples and replays their tokens, providing additional supervision for both token prediction and representation alignment. Past experience thus continues to shape learning as the model changes, helping it correct errors and reinforce learned knowledge for more reliable predictions.

Figure 3 : LLM-JEPA can achieve strong alignment yet still make errors and lose previously correct predictions. For LLM-JEPA, the JEPA loss converges in (a). Yet in (b) and (c), errors remain even when cosine similarities between inputs and targets are close to one. Some correct predictions also become incorrect as training continues. For ER-JEPA, the JEPA loss also converges in (d). In (e) and (f), more predictions are correct despite slightly lower cosine similarities. At each checkpoint, ER-JEPA consistently has more correct answers than LLM-JEPA. These results suggest that replay supervision helps resolve errors that alignment alone leaves unresolved. This figure uses Llama-3.2-1B on 2000 SYNTH test examples with seed 82. Heatmap rows are sorted independently within each method by final-checkpoint correctness and prediction history.

Panel of Figure 4 — Figure 4 : ER-JEPA improves accuracy at matched training compute and reduces overfitting. (a) All three replay policies outperform LLM-JEPA at every evaluated budget. These gains support the benefit of introducing the replay path. (b) Baseline loses accuracy after an early increase. LLM-JEPA delays this decline but still loses accuracy later. Content and uniform replay maintain their gains, while hard replay shows only a small final decline. These trends suggest that replay mitigates overfitting. They complement the mean curves in Figure 1 (right). Results use Llama-3.2-1B on SYNTH at six compute budgets from $40.05$ to $240.31$ PFLOPs.

Baseline error category	Baseline errors	Method	Corrected	Correction rate (%)
Over-generation	$718.4\pm 112.5$	LLM-JEPA	$378.8\pm 64.6$	$53.07\pm 7.92$
Over-generation	$718.4\pm 112.5$	ER-JEPA	$610.8\pm 126.1$	$84.51\pm 4.70$
Under-generation	$2.6\pm 1.5$	LLM-JEPA	$0.6\pm 0.5$	$24.00\pm 25.10$
Under-generation	$2.6\pm 1.5$	ER-JEPA	$1.2\pm 0.8$	$54.67\pm 44.07$
Same-length mismatch	$168.4\pm 10.9$	LLM-JEPA	$34.8\pm 5.8$	$20.58\pm 2.35$
Same-length mismatch	$168.4\pm 10.9$	ER-JEPA	$42.4\pm 7.0$	$25.07\pm 2.62$

Table 1 : ER-JEPA achieves higher mean correction rates across all three Baseline error categories. Results use Llama-3.2-1B on SYNTH at 240.309 PFLOPs. Values are means

\pm

one sample standard deviation over five seeds. Correction rates are computed separately for each seed before averaging.

Figure 7 : ER-JEPA achieves higher accuracy and corrects more errors during training. (a) Test accuracy at comparable training compute for SFT, LLM-JEPA, and ER-JEPA. (b) The fraction of wrong predictions that become correct at the next checkpoint. (c) The fraction of correct predictions that become wrong at the next checkpoint. Compared with LLM-JEPA, ER-JEPA improves error correction across all five seeds, while the reduction in lost correct predictions is smaller and varies across seeds. Results use Llama-3.2-1B on the same 2,000 SYNTH test examples.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 10 : ER-JEPA predictions on examples that LLM-JEPA predicts incorrectly. Rows indicate LLM-JEPA error categories, and columns indicate ER-JEPA prediction categories for the same input and seed. The row total $n$ counts LLM-JEPA errors pooled across five seeds. Each cell shows its count divided by $n$ , expressed as a percentage to one decimal place. Results use Llama-3.2-1B on 2,000 SYNTH test examples per seed at a matched training compute budget of 240.309 PFLOPs.

Subset	Classification rule	Meaning
Complement	Contains the operator ~ .	Excludes strings matched by a subexpression.
Intersection	Contains the operator & .	Requires multiple patterns to hold simultaneously.
Alternation	Contains the operator \| .	Allows a choice between alternative patterns.
Quantifier	Contains * , + , ? , or a brace quantifier such as {m,n} .	Specifies repetition counts or optionality.
Boundary / anchor	Contains \b , \B , ^ , or $ .	Constrains matching positions relative to word boundaries or string endpoints.
Character class	Contains a bracketed character class, such as [a-z] .	Specifies a set of allowed characters.

Appendix

Table 2 : Definitions of the six regex structural subsets. Each subset contains examples whose target regular expressions include the specified feature.

Seed	Baseline errors	LLM-JEPA corrected	Correction rate (%)	ER-JEPA corrected	Correction rate (%)
(a) Over-generation
4	901	448	49.72	807	89.57
23	697	435	62.41	583	83.64
37	705	375	53.19	608	86.24
82	591	344	58.21	455	76.99
84	698	292	41.83	601	86.10

Appendix

Table 3 : Correction counts and rates for each seed on the same Baseline errors. Both methods use the same inputs and seed within each row. Corrected subsets may overlap between methods. The counts and rates in Table 1 are summarized across these five seeds.

Figure 11 : Example predictions from Baseline, LLM-JEPA, and ER-JEPA on the SYNTH dataset.

Figure 12 : Example predictions from Baseline, LLM-JEPA, and ER-JEPA on the NQ-OPEN dataset. Correct and incorrect answers are shown in dark green and dark red, respectively. For each prediction, only the first semicolon-separated answer is displayed.

Panel of Figure 13 — Figure 13 : Raw accuracy at six shared target PFLOPs for seeds 4, 23, 37, and 82. Each panel shows one seed. Seed 84 is shown in Figure 4 (b). The panels do not show means or error bars.

Model / target	Regular expression
Input: lines not having the string “dog” followed by a number, 3 or more times
Ground truth	~((dog.[0-9].){3,})
$\mathcal{L}_{\rm LLM}$	~((dog.[0-9].){3,})
$\mathcal{L}_{\rm LLM-JEPA}$	~((dog.[0-9].){3,})
$\mathcal{L}_{\rm ER-JEPA}$	~((dog.[0-9].){3,})
Input: lines containing ending with a vowel, zero or more times

Appendix

Table 4 : Some regular expressions generated by Llama-3.2-1B-Instruct after fine-tuning with

\mathcal{L}_{\rm LLM}

\mathcal{L}_{\rm LLM-JEPA}

, and

\mathcal{L}_{\rm ER-JEPA}

losses. Color code: wrong , extra , missing .

Method	Accuracy (%) $\uparrow$	Min	Max
ER-JEPA (Sparse Retrieval)	$\mathbf{86.37\pm 0.35}$	$85.75$	$86.60$
Dense Retrieval	$86.21\pm 0.37$	$85.85$	$\mathbf{86.70}$

Appendix

Table 5 : Sparse and dense memory addressing for content replay. All other training settings are fixed. Results report mean accuracy, standard deviation, minimum, and maximum over five seeds.

Memory capacity $\mathcal{M}$	Accuracy (%) $\uparrow$	PFLOPs	Time (min)	Tokens (M)
$10$	$84.22\pm 0.75$	$181.536\pm 0.211$	$10.48\pm 0.07$	$8.581\pm 0.166$
$10^{2}$	$84.32\pm 0.65$	$219.605\pm 0.533$	$11.62\pm 0.22$	$9.427\pm 0.175$
$10^{3}$	$84.76\pm 0.52$	$219.766\pm 0.439$	$16.74\pm 1.33$	$10.969\pm 0.097$
$10^{4}$	$\mathbf{84.98\pm 0.35}$	$219.667\pm 0.393$	$22.23\pm 0.17$	$11.252\pm 0.010$

Appendix

Table 6 : Ablation on episodic memory capacity

\mathcal{M}

for Meta-Llama-3.2-1B-Instruct on NL-RX-SYNTH. Values at epoch 4 are mean

\pm

standard deviation over five seeds. The observed training costs are not compute matched.

Figure 14 : Accuracy across four training epochs for four episodic memory capacities. Points show means over five seeds. Error bars show one sample standard deviation. These runs are not compute matched.

Figure 15 : Ablation on the JEPA hyperparameters

\lambda

and

k

. Each entry reports accuracy (%), with the best result in bold.

Explore similar work

Apr 16, 2026cs.CL

Predict and Reconstruct: Joint Objectives for Self-Supervised Language Representation Learning

Masked language modelling (MLM) has been the dominant pre-training objective for text encoders since BERT, yet it encourages representations that are strongly anchored to surface-form token identity rather than deeper semantic structure. Inspired by the success of Joint Embedding Predictive Architectures (JEPA) (LeCun, 2022) in vision and audio, we propose a hybrid pre-training objective that combines a JEPA-style latent-space prediction loss with a standard MLM objective over a single shared encoder. A learnable scalar parameter continuously balances the two objectives during training. We pre-train both a hybrid model and a pure-MLM baseline on English Wikipedia using identical architectures and compute budgets (NVIDIA H100). Extensive representation analysis across five GLUE benchmarks (SST-2, MRPC, MNLI, CoLA, STS-B) using four pooling strategies reveals that the hybrid encoder produces significantly more uniform embeddings (uniformity less than -0.16 vs -0.05 for MLM), exhibits richer spectral geometry under max pooling, encodes less surface-level lexical information, and achieves a better semantic-to-lexical balance. Despite similar linear-probe downstream accuracy, the geometric differences are consistent and significant, suggesting that the JEPA predictive objective reshapes the latent space in ways that standard accuracy metrics alone cannot capture.

Aimen Boukhari

´Ecole Nationale Sup´erieure d’Informatique (ESI), Algiers , Algeria

Joint-Embedding Predictive Architectures Latent Prediction Self-Supervised Representations Objectives

Ask PDF

May 14, 2026cs.LG

Representation Without Reward: A JEPA Audit for LLM Fine-Tuning

Joint-embedding predictive architectures (JEPAs) propose that a model should learn more useful abstractions when trained to predict latent representations rather than observed outputs. For autoregressive language-model fine-tuning the principle entails a stricter requirement: the induced hidden-state geometry must reach the language-model head \emph{and} improve the decoded task metric. We test that requirement under a fixed Llama-3.2-1B-Instruct LoRA harness on natural-language-to-regex generation, comparing twenty-two training-time auxiliaries across trajectory-shape regularisation, distributional constraints, predictor/target asymmetry, Fisher-metric Jacobi residuals, and a decoder-visible JEPA objective constructed to lie in cross-entropy's positive cone. The empirical answer is a structured null: several auxiliaries clear single-cell paired

α= 0.10

without correction (T3-Local at

Δ= +2.53

~pp,

p = 0.003

being the strongest), but none survives Bonferroni or Holm--Bonferroni at the relevant family-wise threshold, even though many change curvature, anisotropy, variance, and gradient direction. Decoder-visible JEPA yields the first positive auxiliary--cross-entropy gradient cosine in the study, yet exact match remains inside seed noise; a full-fine-tuning replication of the same auxiliary at

n = 5

seeds reproduces the null on both benchmarks (TURK:

Δ= +0.04

~pp,

p_{\text{paired}} = 0.96

; SYNTH:

Δ= +0.52

~pp,

p_{\text{paired}} = 0.28

), so the null is robust across LoRA and full fine-tuning for the decoder-visible construction. Hidden-state representation work and decoded-task accuracy are therefore weakly coupled in this regime; we accordingly reframe LLM-domain JEPA evaluation as a coupling problem, in which the operative question is under which metrics useful hidden geometry becomes decoder-visible task signal.

Biswa Sengupta

LLM Suite Team, JPMorgan Chase & Co.

Large Language Model Fine-Tuning Joint-Embedding Predictive Architectures Hidden States Autoregressive Language Models

Ask PDF

Jul 26, 2026cs.CL

The JEPA Paradox in Language: The Geometry of Linguistic Alternatives

Joint-Embedding Predictive Architectures (JEPAs) are effective for images, video, and audio, yet deterministic JEPA-style latent prediction has not become a standard objective for text encoders. We argue that this gap reflects a mismatch between squared-error latent prediction and the conditional structure of language. The key requirement is conditional concentration: given a context and target location, the target representation should lie near a single meaningful point. Local image prediction often satisfies this through spatial continuity, whereas masked text can admit multiple valid token or span completions whose representations need not share a coherent center. We formalize this mismatch through three conditions---predictability, non-collapse, and low conditional variance---and show how their failure creates centroid degeneracy and collapse pressure in text. Matched I-JEPA and T-JEPA experiments reveal the predicted sequence: mutual-information saturation and elevated target variance precede train--validation instability, effective-rank degeneration, cosine collapse, and poor downstream transfer. The same pattern appears across five independent data seeds, indicating that it is not a sampling artifact. These results do not rule out predictive learning for language; they show that text-compatible JEPA objectives must preserve multiple plausible completions rather than compress them into a single latent point.

Anh Trac Duc Dinh, Khang Nhat Hoang Vo

Center for AI Research (CAIR), VinUniversity, Hanoi, Vietnam · Faculty of Computer Science and Engineering Ho Chi Minh City University of Technology (HCMUT), VNU-HCM Ho Chi Minh City, Vietnam · Mohamed bin Zayed University of Artificial Intelligence Abu Dhabi, United Arab Emirates

Joint-Embedding Predictive Architectures Latent Prediction Sentence Encoders Linguistics Encoders Code Completion+1

Ask PDF