- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper studies a genuinely interesting interface: a fixed-size writer must commit a state before it knows which query will be asked, so the state should preserve the history distinctions that change the answer, not the history variance. The formal object is a history-indexed response field Φ(H), and the central identity is that the width-k response error splits into an unavoidable spectral tail plus an allocation term, L(P) = tr((I−P)C) = Σ_{j>k} λ_j(C) + [Σ_{j≤k} λ_j(C) − tr(PC)]. That is correct, and the paper makes it measurable: a fixed-width exchange identity, a training-side score for proposed swaps, a Clopper–Pearson width certificate, and a 342-row frozen results table. The protocol discipline is well above the norm for this kind of paper — a pre-declared exchange rule, random-replacement controls, an explicit negative control for zero-overlap query families, and a disclosed failure (PH Lift) rather than a hidden one.
Critical strengths
S1: The decomposition is correct, and the paper demonstrates it with a construction rather than an assertion. S2: Protocol discipline that most papers in this area do not have. S3: The calibration mathematics is exactly right. S4: Two results that damage the paper's own case are reported rather than buried. S5: The dependence experiment is the right diagnostic.
Critical weaknesses
The difficulty is that the method is weighted multiresponse reduced-rank regression, which the paper says explicitly ("the estimator is weighted multiresponse reduced-rank regression"; "Under matched features, response weights, and a linear reader, QRM recovers the corresponding population subspace"), and the theory is the Eckart–Young/Ky Fan variational identity plus a conditional-expectation argument. In the one controlled experiment where the response directions are known, QRM and the raw-coordinate coefficient SVD agree to every printed digit (186.1271 / 2.33×10⁻⁵ / 0.4966 / 0.5000), and the only component that distinguishes QRM from that baseline — family-equal column weighting — has all three ablation intervals containing zero. What remains as evidence is a comparison against PCA on six robomimic conditions, and in that comparison a plain GRU with a 32-byte state beats QRM at 64 bytes on all six.
Three further results inside the paper work against the framing rather than for it. The neural controls show full history losing to current-frame-only in 8 of 10 comparisons, which is the opposite of the paper's premise. The width sweep is non-monotone — on Lift, QRM's response error rises from 0.5701 at width 4 to 1.0388 at width 32, 1.11× worse than storing nothing — which cannot happen for a nested prefix family under Theorem 2 and points at reader-fitting instability in a single-seed pipeline. And the calibration certificate, presented as an under-powered sample ("no width is selected"), inverts to 49–55 of 60 episodes failing in one query family, i.e. the 0.05 tolerance is below every method's achieved error and no amount of episodes would certify it.
Questions for authors
- Table 2 reports a 32 B GRU beating the 64 B QRM state on all six conditions. What does QRM offer a practitioner who can afford a learned encoder at the same state size — and if the answer is "interpretability" or "no training beyond a linear fit", say that in the abstract.
- Why does QRM's response error rise from 0.5701 at width 4 to 1.0388 at width 32 on recorded Lift, when the prefixes are nested and tr((I−P)C) is non-increasing? Is per-width ridge selection responsible, and what width do you actually recommend for that task?
- Table 11's minimum joint bound of 0.983274 corresponds to 55 of 60 episodes failing, and the minimum scalar bound of 0.152704 to 2 of 60. Which query family fails, and why is the outcome attributed to the 55-episode requirement rather than to a 0.05 tolerance that no method meets?
- In the 564-row intervention grid, 47 of 504 targeted/random rows are positive. What is the positive rate for targeted swaps alone, and under what conditions does Proposition 5's ideal gain survive the fitted-reader term in Eq. (19)?
- Given Table 3, what is the method's scope? If the deployed query policy conditions on the history — which is what "retrieve the blue object" implies in a closed-loop agent — the proposed allocation is 2.16× worse than a joint-loss search. Do you claim the independent-query regime only?
- Table 13 shows current-only beating full-history in 8 of 10 neural comparisons. How do you reconcile that with the paper's premise, and does the linear current-frame comparison (0.0149, 0.0166) measure the same thing?
- What is the seed-to-seed spread of the QRM−PCA gap? Please report it per seed for the six conditions, since the entire claimed effect on PH Can is 0.0051.
- Table 15's three intervals all contain zero, and Table 1 shows QRM identical to coefficient SVD. What evidence remains that response-weighted allocation, as opposed to whitening plus reduced-rank regression, is doing the work?
- Can the certificate be made informative by lowering its ambition — for example, certifying "no worse than the no-memory floor" instead of "below 0.05" — and would that change any of the six external outcomes?