- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper states that whether a Titans-style test-time memory, a 2-layer MLP updated online by gradient steps on an associative loss over frozen DINOv2 features, actually stores stream content. There are two main empirical findings: 1. IID ImageNet features with a recall objective whose key and value come from the same feature. 2. Egocentric video (Nymeria) with a past/future windowed objective. The paper ends with a minimum set of controls for storage claims about test-time memories.
Critical strengths
The paper names concrete shortcut paths: the key's receptive field, the static scene component, and the write gate collapsing while the loss still falls. It then shows that commonly reported diagnostics can be fully explained by these paths. The diagnostics in question are inner loss, gate magnitude, past/future asymmetry, and seen/unseen AUC.
The state-swap control (rather than only disabling writes, which moves the readout off-distribution), the receptive-field copy, the leave-one-video-out stateless filter, and the rotation over truth videos are good.
Critical weaknesses
The IID recall objective derives key and value from the same feature. Proposition 1 then predicts a write-free optimum by design, and associative-recall benchmarks are built to avoid exactly this; Titans itself is not trained this way.
This paper never shows a configuration where the trained memory passes the proposed audit, e.g. a decoupled-key/value recall objective (MQAR-style) or a synthetic stream with planted long-lag dependencies. Without that, readers cannot tell whether the audit discriminates or whether this training pipeline simply never learns to write.
Each configuration is trained once, with one seed. The probe fixture has five videos, from one dataset, with one encoder and a 4.2M-parameter memory.
The only stateless baseline is a 5-coefficient linear filter over the key window. A natural stateful but non-learned baseline is missing: an exponential moving average or running per-video mean of past features with a fitted decay. Figure 2 shows that much long-range similarity is a static per-video component, and the τ = 10 model "helps the readout without storing items". That pattern is exactly what a running-mean state would produce, so the paper's one positive result cannot yet be attributed to anything beyond scene averaging.
Questions for authors
Can you provide a positive control: a training objective or synthetic stream (decoupled keys/values, or planted dependencies beyond the key window) under which the same architecture and training pipeline passes your audit, with item-level traces and a beneficial state swap? This would show that the audit can support your claims.
For the τ = 10 models, how does a non-learned stateful baseline (EMA or running mean of past features with a decay fit on four videos, read out by the same per-offset linear heads) compare with the memory at j = −7 … −10?
Are the five fixture videos excluded from the 1065 training sequences? If not, how much does training-set overlap affect the seen/unseen and write-effect measurements?
How sensitive are the conclusions to the seed?