Determine whether JEPA alignment improves decisions

Determine whether the JEPA alignment terms improve decision performance, rather than merely reshaping the learned representation, using a full-length control with the alignment weight set to zero and a readout that uses the reshaped representation.

Background

The report states that the independent control isolating the JEPA alignment terms was short and demonstrated a change in representation retrieval, but did not establish an improvement in decisions. It therefore identifies the causal effect of the alignment objectives on judgment quality as unresolved and proposes a longer zero-alignment control with an appropriate readout as the relevant evaluation.

References

The alignment terms are isolated only by a short independent control, which shows a reshaped representation but not yet a decision gain. The head's subspace also captures less of $g$ than before, so a full-length control with $\lambda = 0$ and a readout that uses the reshaped representation come next.

— Bongard: Training Machine Intuition  (2609.39111 - Ding et al., 30 Sep 2026) in Section 7, Limitations

Why the success rate of the frozen V-JEPA 2 falls while its readout improves is unresolved.

— Supervise What Decides Success: Criterion-Aligned Auxiliary Losses for Latent World-Model Planning  (2610.01224 - Hara et al., 1 Oct 2026) in Appendix, Section “Frozen-encoder systems: scope of replication and an exception,” paragraph “Interventions on the cost side”