PaperlayerView exact revision on arXiv

Source-linked AI summary

TwinJEPA: Action-Preferred Predictive Representations for Goal-Conditioned Control

Feiran You, Hongyang Du

arXiv:2610.02922v1eess.SYcs.SY

TL;DR

Transition-wise predictive objectives supervise observed actions in isolation, providing limited information about alternatives in similar states with different goal-conditioned outcomes. TwinJEPA adds offline-mined local action comparisons with reward-gap and preference supervision. It yields positive benchmark-level mean differences across matched state-based evaluations, with effects varying across tasks and domains.

  • Problem

    Transition-wise prediction supervises observed actions individually, leaving their distinctions from plausible alternatives in similar states only indirectly constrained.

    Source: §1 · pp. 1–2

  • Method

    TwinJEPA adds offline-mined comparisons between action-diverse transitions from similar states, using goal-conditioned relabeling for reward-gap and preference supervision.

    Source: §1 · p. 2

  • Results

    TwinJEPA yields positive mean differences across all matched state-based evaluations, with substantial variation in magnitude and seed-level consistency across benchmarks.

    Source: §1 · p. 3, §4.2 · p. 7

  • Takeaways & Limitations

    Local action-comparison supervision complements temporal prediction by preserving both long-horizon structure and distinctions among alternative actions in similar states.

    Source: §1 · pp. 2–3

Abstract

from arXiv · show

Joint-Embedding Predictive Architectures (JEPAs) are a promising paradigm for representation learning by predicting future latent states without reconstructing observations. Recent work has adapted JEPA-style latent prediction to offline zero-shot control through action-conditioned temporal prediction, enabling representations to capture long-horizon dynamics from fixed behavioral data. However, transition-level predictive objectives supervise actions in isolation and provide limited information about which actions are preferable when similar states admit different goal-conditioned outcomes. We introduce TwinJEPA, a framework for learning action-preferred predictive representations by augmenting JEPA-based control with offline-mined action-preference supervision. TwinJEPA identifies approximately matched states across offline trajectories and constructs preference pairs via goal-conditioned reward relabeling. It then learns two complementary objectives: reward-gap regression, which preserves the magnitude of outcome differences, and preference classification, which captures the relative ordering of alternative actions. Both objectives are used only during training and incur no additional inference-time cost. We evaluate TwinJEPA on long-horizon navigation and continuous-control benchmarks with state- and pixel-based observations. TwinJEPA yields positive benchmark-level mean differences across all matched state-based evaluations, while analyses across domains and observation modalities indicate that larger gains tend to arise when local action alternatives provide more informative outcome contrasts. The results suggest that local action-comparison supervision can complement temporal prediction by encouraging JEPA-based representations to retain both long-horizon temporal structure and fine-grained distinctions among locally observed actions for offline zero-shot control.

1 INTRODUCTION

TwinJEPA addresses the limited information transition-wise prediction provides about alternative actions in similar states by adding offline-mined, goal-conditioned action comparisons. Evaluations show positive benchmark-level mean differences across matched state-based benchmarks, with gains varying across tasks and domains.

1 INTRODUCTION

  • Transition-wise prediction supervises observed actions independently, leaving distinctions between plausible actions in similar states only indirectly constrained.Control also depends on preserving information that separates actions associated with different downstream outcomes.Source: §1 · pp. 1–2
  • TwinJEPA mines action-diverse transitions from similar local states and uses goal-conditioned reward relabeling to supervise outcome-gap magnitude and action-preference ordering.The auxiliary heads are used only during training, leaving downstream control and inference unchanged.Source: §1 · p. 2, Figure 1 · p. 2
  • Unlike preference-based RL, TwinJEPA derives comparisons automatically from offline data and uses them for representation learning rather than reward modeling or policy optimization.It also differs from transition-wise JEPA methods by explicitly comparing actions in locally similar states.Source: §1 · p. 2
  • TwinJEPA yields positive mean differences across all matched state-based evaluations, while task-level and component analyses show substantial variation across domains.The evaluations also include pixel-based ExORL experiments.Source: §1 · p. 3

2 PRELIMINARIES

The preliminaries frame the task as offline goal-conditioned control, with a learned representation supporting zero-shot deployment. They describe TD-JEPA's temporal-prediction backbone and its limitation in directly specifying relationships among local action alternatives.

2.1 OFFLINE GOAL-CONDITIONED AND ZERO-SHOT CONTROL

  • Offline zero-shot control learns reusable behaviors from a fixed dataset, with task latents sampled during training and inferred from task-specific information at deployment.The downstream task is unspecified during representation learning, and deployment requires no additional policy optimization or online fine-tuning.Source: §2.1 · p. 3

2.2 JEPA-BASED TEMPORAL PREDICTIVE CONTROL

  • TD-JEPA uses latent prediction rather than observation reconstruction, with action- and task-conditioned predictors estimating future representations from two encoder spaces.Momentum networks provide bootstrapped targets.Source: §2.2 · pp. 3–4
  • The unchanged representation-learning backbone combines TD-JEPA temporal prediction with representation regularization, while TwinJEPA adds comparative supervision.The temporal objective trains each observed action against its future target without directly comparing outcomes for locally similar states and different actions.Source: §2.2 · p. 4

3 METHOD

TwinJEPA adds offline-mined local action comparisons to the unchanged temporal-prediction backbone. Its reward-gap and preference objectives supervise complementary properties of those comparisons, while downstream control and inference remain unchanged.

3 METHOD

  • TwinJEPA augments the unchanged temporal-prediction and regularization objectives with offline-mined action comparisons and lightweight auxiliary heads used only during training.The representation-learning objective combines the baseline loss with a weighted pairwise loss; the actor-learning procedure remains the same as the baseline.Source: §3 · p. 4, §3.3 · pp. 5–6

3.1 OFFLINE LOCAL ACTION-COMPARISON MINING

  • The method forms action-diverse pairs from locally similar states, then selects an offline-relabeled goal that maximizes their reward contrast and removes negligible contrasts.Candidate goals come from the fixed training dataset, not held-out evaluation goals; the same comparative targets also extend to pixel-based ExORL experiments.Source: §3.1 · pp. 4–5

3.2 PAIRWISE COMPARATIVE SUPERVISION

  • A state-only head predicts local outcome-contrast magnitude, while a second head uses the anchor representation and both candidate actions to predict their preference ordering.The two objectives provide magnitude-sensitive and pair-specific ordering supervision, and neither is assumed to dominate across environments.Source: §3.2 · p. 5

3.3 JOINT REPRESENTATION LEARNING

  • The shared encoder receives gradients from both the baseline representation objective and the pairwise objective, while auxiliary heads train against their corresponding targets.Setting the pairwise-loss weight to zero recovers the matched temporal-prediction baseline.Source: §3.3 · pp. 5–6

3.4 INFERENCE AND COMPUTATIONAL OVERHEAD

  • At test time, TwinJEPA retains the baseline task-inference and control procedures; mined pairs and auxiliary heads are not used for inference.Pair construction and training add offline cost, but inference-time computational cost remains the same as the baseline.Source: §3.4 · p. 6

4 EXPERIMENTAL RESULTS

Matched-protocol evaluations compare TwinJEPA with TD-JEPA across navigation and continuous-control benchmarks. Benchmark means are positive, but the task-level effects vary across goals and seeds.

4.1 EXPERIMENTAL SETUP AND STATISTICAL PROTOCOL

  • Evaluations matched architecture, offline data, training budget, seeds, task inference, and protocol; TwinJEPA added only the action-comparison objective.The reported analysis includes paired differences, bootstrap confidence intervals, sign-flip permutation tests, and positive-seed counts.Source: §4.1 · p. 6

4.2 MATCHED STATE-BASED EVALUATION (RQ1)

  • TwinJEPA's matched mean differences were AntMaze-4 +2.3 pp, Maze2D +5.2 pp, Walker +8.4, Cheetah +33.7, and Quadruped +40.4.AntMaze-4 increased from 50.1% to 52.4%, and Maze2D from 9.2% to 14.4%.Source: §4.2 · pp. 6–7, Figure 2 · p. 7
  • Quadruped had five positive paired differences and a 95% paired-bootstrap confidence interval of [+20.6, +60.3]; AntMaze and Walker effects were smaller and more variable.Cheetah also showed a substantial positive mean difference.Source: §4.2 · p. 7

4.3 EFFECT STRUCTURE ACROSS DOMAINS AND TASKS (RQ2)

  • Task-level changes are not uniform: navigation gains depend on goals, and Maze2D's aggregate gain comes from goals where the baseline already succeeds.Other goals are unchanged in Maze2D; AntMaze includes both positive and negative task-level differences.Source: §4.3 · p. 7
  • Walker and Cheetah gains concentrate on selected tasks, while Quadruped improves across all four tasks; headroom alone does not explain the pattern.The results describe the effect of comparative supervision as task- and data-dependent.Source: §4.3 · p. 7

B. ExORL / DMC

Across ExORL/DMC evaluations, task and seed effects vary, and the contributions of reward-gap regression and preference classification differ by domain.

B. ExORL / DMC

  • Walker and Cheetah concentrate larger gains on selected locomotion tasks, often where the matched baseline retains performance headroom; Quadruped improves across all four tasks.Quadruped's gains occur despite variation in baseline return, so headroom alone does not explain the pattern.Source: B. ExORL / DMC · p. 8, Figure 4 · p. 9
  • Effects vary by task and seed: Cheetah is more consistently positive, Maze2D is positive on most matched seeds, and Quadruped is positive on all five runs.AntMaze and Walker show substantial run-to-run variation.Source: B. ExORL / DMC · p. 8

4.4 COMPONENT CONTRIBUTIONS ACROSS DOMAINS (RQ3)

  • The ablation ordering depends on the setting: preference-only leads on AntMaze medium-navigate, gap-only on medium-stitch, while Cheetah's full objective has the highest evaluated mean.On the matched five-seed Cheetah ablation subset, the full objective reaches 671.4, versus 654.5 for gap-only and 629.8 for preference-only.Source: Figure 3 · p. 8, §4.4 · p. 9
  • Quadruped reaches 599.8 with the full objective, versus 554.4 gap-only and 548.4 preference-only; both isolated objectives fall below TD-JEPA's 559.4.The component results show no fixed ordering across domains; the shared weights are (λ, αgap, αpref) = (0.1, 1.0, 0.5).Source: §4.4 · p. 9

5 CONCLUSION

TwinJEPA adds offline-mined local action comparisons to temporal prediction for zero-shot goal-conditioned control. The paper reports positive average effects in several matched settings, but results and auxiliary-signal contributions vary across tasks and domains.

5 CONCLUSION

  • TwinJEPA combines transition-level prediction with locally matched, action-diverse pairs and goal-conditioned reward relabeling, producing reward-gap and preference-ordering targets.These auxiliary objectives are used during representation learning and add no inference-time computation.Source: §5 · pp. 9–10

A EXTENDED RELATED WORK

  • TwinJEPA complements predictive-representation work by comparing different actions in locally similar states, rather than replacing the temporal representation objective.Its pairwise labels are generated from offline data without human preferences or extra environment interaction.Source: A EXTENDED RELATED WORK · pp. 12–13

B ADDITIONAL EXPERIMENTAL DETAILS

  • The appendix reports implementation details, extended task-level results, ablations, sensitivity studies, and benchmark-specific evaluations under matched TD-JEPA/TwinJEPA protocols.The methods share backbone architectures, optimization settings, training budgets, seeds, and evaluation protocols within each comparison.Source: REPRODUCIBILITY STATEMENT · p. 10, B ADDITIONAL EXPERIMENTAL DETAILS · p. 13

EVALUATION ENVIRONMENTS

  • The benchmark suite spans sparse-reward navigation, dense locomotion, and high-dimensional posture control; results are intended to examine when local comparisons provide useful supervision.The listed environments include AntMaze, Maze2D, Walker, Cheetah, and Quadruped.Source: EVALUATION ENVIRONMENTS · p. 15

B.2 CROSS-DOMAIN AUXILIARY-OBJECTIVE SUMMARY

  • Ablations show domain-dependent component rankings: preference-only leads on medium-navigate, gap-only on medium-stitch, and the full objective exceeds both isolated components on Quadruped.The cross-domain summary does not support a universal ordering or a claim that combining the objectives always dominates.Source: §B.2 · p. 15, Table 3 · p. 15

B.3 SENSITIVITY TO LOCAL MATCHING TOLERANCE

  • AntMaze success stays within 76.0–78.0% as matching tolerance varies from 0.5× to 2× default; Maze2D's non-monotonic probe has only two seeds per setting.The Maze2D probe is exploratory, not evidence for cross-domain robustness; these sensitivity runs do not determine primary configurations.Source: §B.3 · p. 16

C.1 TASK-LEVEL ANTMAZE STATISTICS

  • AntMaze task-level changes include both gains and declines, even though all four domain means are positive.Within medium-navigate, the largest positive difference is on Task 3; other domains also contain sizable positive and negative changes.Source: §C.1 · pp. 17–18, Figure 6 · p. 18

C.2 OFFICIAL TD-JEPA REPRODUCTION CHECK

  • The reproduced Official TD-JEPA medium-navigate result is 68.8±4.0%, close to its reported 70.4%; the matched reproduction is the reference for within-paper effect estimates.TwinJEPA and the reproduced baseline share the training and evaluation setup, differing systematically in the added action-comparison supervision.Source: §C.2 · p. 18

C.3 ADDITIONAL ABLATION AND SENSITIVITY RESULTS

AntMaze ablations show that the two comparative objectives have domain-dependent effects, while medium-navigate learning curves show how performance differences emerge over recorded checkpoints. The Walker extension examines sensitivity across optimization configurations.

C.3 ADDITIONAL ABLATION AND SENSITIVITY RESULTS

  • All three evaluated auxiliary weights improve mean medium-navigate success over matched Official TD-JEPA, but the shared λ = 0.1 setting is not the best in that domain.The shared weight is retained for cross-domain consistency, not selected as a medium-navigate optimum.Source: §C.3 · p. 19, Table 8 · p. 20
  • Preference-only performs best on medium-navigate, whereas gap-only performs best on medium-stitch; the full objective does not dominate both isolated components in either domain.The results indicate distinct, domain-dependent contributions rather than a universally stronger combination.Source: §C.3 · p. 19, Table 8 · p. 20

C.4 LEARNING DYNAMICS ON MEDIUM-NAVIGATE

  • At 1M steps, medium-navigate mean success is 73.2% for TwinJEPA versus 68.8% for Official TD-JEPA (∆+4.4), after a non-monotonic training trajectory.Official TD-JEPA leads at 250k steps; TwinJEPA leads at 500k, and the difference narrows at 750k.Source: Figure 8 · p. 20, §C.4 · pp. 20–21
  • On Task 2, both methods reach high success early and finish at 80% for Official TD-JEPA and 86% for TwinJEPA.On Task 3, TwinJEPA first reaches at least 50% at 500k steps versus 750k for Official TD-JEPA; the final difference is 17 points, partly reflecting a late baseline decline.Source: §C.4 · p. 21

D DMC WALKER EXTENDED EVALUATION

  • The Walker extension tests whether the observed behavior is specific to a single hyperparameter configuration by examining a broader optimization grid.The supplied passage describes the evaluation scope, not its results.Source: D DMC WALKER EXTENDED EVALUATION · p. 21

D.1 HYPERPARAMETER-GRID EVALUATION

The Walker hyperparameter grid shows higher TwinJEPA mean return in most configurations, with effect size depending on the optimization setting. The Cheetah subsection describes a dense-reward locomotion evaluation and its matched zero-shot comparison setup.

D.1 HYPERPARAMETER-GRID EVALUATION

  • TwinJEPA has higher mean return in five of six Walker settings, with differences from −2.5 to +12.6; across all six settings, the aggregate changes from 765.5 to 771.3 (∆+5.8).At the shared Fixed-HP setting, the four-task mean changes from 774.3 to 782.7 (∆+8.4).Source: §D.1 · p. 21, Table 10 · p. 22
  • The Walker grid is a robustness diagnostic, not model selection: the largest difference occurs at (0.1, 10−5), and effect magnitude depends on configuration.The shared Fixed-HP setting is one grid point, not a universally preferred configuration.Source: §D.1 · p. 22
  • The Walker task-level analysis reports gains concentrated on stand and run, while walk and spin operate closer to the reward ceiling.The supplied passage refers to the main-paper task-level analysis.Source: §D.1 · p. 22

E CHEETAH (EXORL) OFFLINE CONTROL

  • Cheetah is evaluated as a dense-reward locomotion setting whose performance depends on coordinated joint motion and sustained velocity rather than explicit goal geometry.The zero-shot comparison covers walk, walk backward, run, and run backward, using 10 matched seeds.Source: E CHEETAH (EXORL) OFFLINE CONTROL · p. 23

E.1 COMPLETE TASK-LEVEL RESULTS

Cheetah's four-task mean return is higher for TwinJEPA under the 10-seed Fixed-HP comparison, with the largest task differences on forward locomotion. Five-seed ablations favor the full objective on average, though seed-level outcomes vary.

E.1 COMPLETE TASK-LEVEL RESULTS

  • The largest Cheetah mean differences occur on walk and run, while walk backward is close to the apparent return ceiling for both methods.This task-level breakdown complements the main-paper visualization.Source: §E.1 · pp. 23–24
  • Cheetah's four-task mean increases from 638.0 to 671.7 (∆+33.7), with positive paired differences on 7 of 10 matched seeds.Across-seed standard deviation is 24.3 for TwinJEPA versus 44.2 for Official TD-JEPA; the passage treats this descriptively.Source: §E.1 · p. 23, Table 11 · p. 24

E.2 AUXILIARY-OBJECTIVE AND WEIGHT SENSITIVITY

  • Within the five-seed ablation subset, the full objective has the highest mean return; gap-only retains most of the baseline improvement, while preference-only yields a smaller positive difference.Both λ = 0.05 and λ = 0.5 yield lower mean returns than the shared λ = 0.1 setting, without establishing a globally optimal weight.Source: §E.2 · p. 24, Table 12 · p. 24
  • Full TwinJEPA exceeds the matched TD-JEPA baseline on four of five ablation seeds, so the aggregate ordering does not hold on every run.The per-seed table and figure provide a finer-grained view of run-to-run variation.Source: §E.2 · pp. 24–25, Table 13 · p. 25, Figure 11 · p. 25

F QUADRUPED EXTENDED EVALUATION

Quadruped shows a positive four-task mean difference for TwinJEPA, with positive paired differences across all five seeds and all four tasks. The full objective leads the isolated components in this domain, but the paper cautions against generalizing that pattern across domains.

F QUADRUPED EXTENDED EVALUATION

  • Quadruped's four-task mean increases from 559.4 for TD-JEPA to 599.8 for TwinJEPA (∆+40.4), with a paired bootstrap 95% confidence interval of [+20.6, +60.3].All five matched seeds are positive, but the exact two-sided sign-flip test gives p = 0.0625.Source: F QUADRUPED EXTENDED EVALUATION · p. 26

F.1 SEED-LEVEL RESULTS

  • TwinJEPA exceeds TD-JEPA on all five matched seeds, with paired differences ranging from +12.6 to +71.6.The uniformly positive direction is consistent across available runs, while the effect magnitude remains heterogeneous.Source: §F.1 · p. 26, Table 14 · p. 27

F.2 TASK-LEVEL NUMERICAL BREAKDOWN

  • All four Quadruped tasks have positive mean differences: walk +72.1, stand +49.6, run +24.8, and jump +15.2.The gain spans locomotion, posture, and jumping, though its magnitude varies by task.Source: Table 15 · p. 27, §F.2 · p. 27

F.3 ADDITIONAL AUXILIARY-OBJECTIVE SENSITIVITY

  • In the five-seed comparison, Full TwinJEPA reaches 599.8 mean return versus 554.4 for Gap-only and 548.4 for Pref-only; both isolated components fall below TD-JEPA.The λ = 0.05 variant reaches 569.2, +9.8 versus TD-JEPA but 30.6 below the shared λ = 0.1 configuration.Source: §F.3 · pp. 27–28, Table 16 · p. 28
  • The stronger Quadruped full-objective result does not establish universal synergy: component behavior changes across AntMaze, Cheetah, and Quadruped.The λ comparison is local sensitivity analysis, not evidence that λ = 0.1 is globally optimal.Source: §F.3 · p. 28

G D4RL MAZE2D EXTENDED EVALUATION

In the simplified planar Maze2D navigation evaluation, TwinJEPA improves mean zero-shot success under the shared Fixed-HP configuration, but gains are limited to three of five tasks. The result is a boundary-case evaluation, not evidence of uniform benefits across navigation settings.

Maze2D setup

  • Maze2D removes AntMaze’s high-dimensional locomotion dynamics while retaining goal reaching in a simpler planar point-mass domain.The evaluation tests the comparative supervision in geometric navigation.Source: G D4RL MAZE2D EXTENDED EVALUATION · p. 28

Task-level results

  • The five-task mean success rate rises from 9.2% to 14.4% (∆+5.2 percentage points) under the shared Fixed-HP configuration.Figure 14 compares Official TD-JEPA and TwinJEPA after 1M training steps using five matched seeds.Source: G D4RL MAZE2D EXTENDED EVALUATION · p. 28
  • T1 and T2 remain unsolved by both methods, while T3, T4, and T5 improve by 12, 6, and 8 percentage points, respectively.The aggregate gain is mainly contributed by tasks where the baseline achieves nonzero performance.Source: G D4RL MAZE2D EXTENDED EVALUATION · p. 29

Interpretation

  • Because Maze2D has simpler dynamics and a lower-dimensional state space, its results do not imply identical benefits across goal-conditioned navigation settings.The evaluation shows positive differences only on the three tasks with observed nonzero performance.Source: G D4RL MAZE2D EXTENDED EVALUATION · p. 29

H PIXEL-BASED EXORL EVALUATION

The pixel experiments apply TwinJEPA’s comparative objective to cached RGB frame stacks in ExORL Walker, Cheetah, and Quadruped. Results vary by domain and seed composition, supporting modality applicability but not uniformly positive or directly comparable state-versus-pixel performance.

Pixel ExORL setup

  • The same reward-gap and preference objectives apply to pixel observations, and their auxiliary heads add no evaluation-time computation.The protocol uses a DrQ encoder with 64×64 RGB observations and a three-frame stack.Source: H PIXEL-BASED EXORL EVALUATION · p. 30

Cheetah seed-set sensitivity

  • Cheetah’s complete ten-seed comparison favors TD-JEPA by −8.5 points, while a diagnostic five-seed subset favors TwinJEPA by +4.3.The authors do not treat the post-inspection subset as a replacement for the ten-seed estimate or as evidence of reliable positive effect.Source: H PIXEL-BASED EXORL EVALUATION · pp. 30–32

Task-level behavior

Overall interpretation

  • The unweighted domain mean changes from 685.5 to 693.0, a descriptive difference of +7.5 points.The average is descriptive because domains use different numbers of matched seeds.Source: Table 17 · p. 31, H PIXEL-BASED EXORL EVALUATION · p. 32
  • The pixel protocol uses cached RGB frame stacks rather than proprioceptive states, with state and pixel returns kept separate because encoder architecture, batch size, and training horizon differ.Figure 15 illustrates the three stacked frames and corresponding state vector for each domain.Source: H PIXEL-BASED EXORL EVALUATION · pp. 30, 32, Figure 15 · p. 31

I THEORETICAL PROPERTIES OF TWINJEPA

TwinJEPA’s theoretical analysis characterizes the auxiliary supervision on a fixed distribution of mined local action-comparison pairs. Reward-gap regression promotes decodability of contrast magnitudes, while preference classification constrains reward-induced local rankings; the results do not establish universal downstream gains.

Reward-gap decodability

  • For a fixed representation, minimum squared reward-gap risk is the expected conditional variance of gap magnitude given Z, so reducing it requires Z to predict local contrast magnitudes.The affine-head specialization has zero risk exactly when the gap is an affine function of Z, under the stated invertibility condition.Source: I THEORETICAL PROPERTIES OF TWINJEPA · pp. 33, 35

Local ranking control

  • On the mined pair distribution, preference cross-entropy upper-bounds the probability that the learned scorer reverses the reward-induced ordering of a retained action pair, up to 1/log 2.The result does not establish correct ordering for arbitrary states, actions, or goals outside that distribution.Source: I THEORETICAL PROPERTIES OF TWINJEPA · pp. 33, 35, 36

Interpretation and scope

  • The propositions characterize supervision for a fixed pair distribution; changing matching or relabeling rules changes that distribution, and the analysis does not establish an optimal mining rule.The theory does not imply universal task improvement, universal complementarity of the auxiliary terms, or unchanged TD-JEPA guarantees at nonzero auxiliary weights.Source: I THEORETICAL PROPERTIES OF TWINJEPA · pp. 33, 36

Target consistency

  • Reward-gap regression and preference classification derive their targets from the same selected signed reward contrast, encoding its magnitude and its ordering, respectively.The selected goal maximizes the absolute one-step reward contrast over the candidate reward family.Source: I THEORETICAL PROPERTIES OF TWINJEPA · pp. 33–34

Exact reduction to TD-JEPA

  • When λ = 0 or both auxiliary weights are zero, TwinJEPA reduces exactly to the original TD-JEPA objective.The analysis transfers TD-JEPA theoretical properties only in this exact-reduction case.Source: I THEORETICAL PROPERTIES OF TWINJEPA · p. 34

J SUMMARY OF ADDITIONAL RESULTS

The additional evaluations show that TwinJEPA’s comparative supervision is applicable across tasks and domains, while downstream effects remain heterogeneous. Task, domain, optimization configuration, and the mined comparison distribution qualify the observed gains.

Overall interpretation

  • The additional results support reward-gap decodability and reward-consistent local ranking on mined pairs, while downstream effects vary with task, domain, optimization, and comparison distribution.The extended evaluation supports a narrower conclusion than uniform superiority across benchmarks.Source: J SUMMARY OF ADDITIONAL RESULTS · pp. 36–37

Task-level heterogeneity

  • AntMaze-4 mean success changes from 50.1% to 52.4% (∆+2.3), with non-negative mean differences across four domains but mixed task-level changes.Medium-navigate Task 3 improves by 17 percentage points, while several other goals change little or decrease.Source: J SUMMARY OF ADDITIONAL RESULTS · p. 36

Auxiliary-objective contributions

  • Component rankings vary by domain: preference-only leads on AntMaze medium-navigate, gap-only on medium-stitch, while Full TwinJEPA leads the listed Cheetah and Quadruped ablations.Cheetah returns are 671.4 for Full, 654.5 for Gap-only, and 629.8 for Pref-only; Quadruped returns are 599.8, 554.4, and 548.4, respectively.Source: J SUMMARY OF ADDITIONAL RESULTS · pp. 36–37

Optimization sensitivity

  • Changing λ can alter downstream effect magnitude, so λ = 0.1 is retained as a shared evaluation configuration rather than treated as a per-domain optimum..Source: J SUMMARY OF ADDITIONAL RESULTS · p. 37

Cross-domain transfer

  • Under the shared Fixed-HP protocol, Walker changes from 774.3 to 782.7 (∆+8.4), Cheetah from 638.0 to 671.7 (∆+33.7), and Quadruped from 559.4 to 599.8 (∆+40.4).Matched-seed uncertainty is substantial for Walker and Cheetah; Quadruped is positive on all five matched seeds, but that seed count limits exact-test resolution.Source: J SUMMARY OF ADDITIONAL RESULTS · p. 37

Maze2D evaluation

  • Maze2D’s five-task mean success changes from 9.2% to 14.4% (∆+5.2), but T1 and T2 remain unsuccessful for both methods.The result is a simplified-navigation boundary case, not evidence that all goals benefit or that the auxiliary losses resolve shared failures.Source: J SUMMARY OF ADDITIONAL RESULTS · p. 37

Pixel evaluation

  • Pixel results show differences of +25.3 on Walker, −8.5 on complete ten-seed Cheetah, and +5.6 on Quadruped; the unweighted mean is descriptive across unequal seed counts.Cheetah’s fixed random five-seed diagnostic reverses direction, so the pixel experiments do not establish modality-independent improvement.Source: J SUMMARY OF ADDITIONAL RESULTS · p. 37

Limitations and future directions

  • TwinJEPA’s effectiveness depends on the quality and coverage of mined comparisons; current fixed, benchmark-specific pair rules may miss informative comparisons across domains.Future directions include adaptive pair construction, automatic selection of informative neighborhoods, and broader offline-control and embodied-learning evaluation.Source: J SUMMARY OF ADDITIONAL RESULTS · p. 38

Related papers

Loading 2610.02922v1…