- Rating
- 3(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
Focusing on enhancing the expressiveness of state-space models (SSMs), this paper proposes quantum-channel-inspired transition matrices for SSMs whose structures admit both conservative and dissipative transitions with stability. Experimental results show superior state tracking across synthetic and real-world benchmarks, especially under noisy measurements/observations, and competitive time-series classification performance.
Critical strengths
-
The paper clearly provides a novel lens on improving recurrent models by observing the fact that the permutation and reset characterization of finite-state computation suggests two different roles for a recurrent transition. In effect, it motivates input-selective transitions in which a single update can preserve some state distinctions while contracting others.
-
Providing new experimental setups and results demonstrating improved state tracking capability under noisy observations/measurements in synthetic and real-world benchmarks.
Critical weaknesses
-
The paper has not provided any computational complexity analysis of the proposed CPTP-SSM (and a comparison with other SSMs). What are the memory and computational complexities with respect to state size and input length?
-
Although it was not motivated by use cases in language modeling, it would be useful to elaborate/comment the potential advantages of the CPTP-SSM for LLMs, if any.
Questions for authors
-
Noisy measurement experiments are very interesting. Could they be systematically expanded with more synthetic tasks or, better, with more real-world benchmarks? This is a very nice feature to be thoroughly evaluated, which can unlock truly new use cases. Also, PD-SSM is missing from Tables 6 and 3; it would be good to add it for the sake of a complete comparison.
-
For time-series classification, it would be great to provide a more comprehensive comparison with other time-series/long-sequence tasks that LTI systems (like LinOSS) are normally benchmarked.
2.1) In Table 5, it would be great to cross-compare with Flash PD-SSM and benchmark with, e.g., time-series forecasting tasks reported in Flash PD-SSM paper.
2 comments
Response to Reviewer z33K, part 2 of 2. Noisy measurements and long sequences
We agree that tracking under noisy measurements deserves a test on real data, and the pointer to Flash PD-SSM was useful. Both led to new results in the revision.
Q1. More noisy-measurement experiments, and PD-SSM in Tables 3 and 6
Two new real-world benchmarks in Section 4.2. We add orientation tracking from a real gyroscope on the EuRoC drone flights and on the hand-held TUM-VI sequences, with the split and the open-loop protocol of Brossard et al. 2020. The input is a rotation increment per 50 ms, as in Table 3, but measured by a real sensor. Models are trained on windows of at most 2 s. Each whole test sequence, 98 to 150 s long on EuRoC and 2 to 2.5 min on TUM-VI, is then run once from start to end without reset. The table gives the absolute orientation error in degrees, mean ± SD over three seeds.
| Model | EuRoC | TUM-VI |
|---|---|---|
| LSTM | 33.34 ± 8.39 | 10.37 ± 0.36 |
| PD-SSM | 65.99 ± 10.72 | 104.09 ± 10.87 |
| Gated DeltaProduct | 61.50 ± 4.98 | 99.55 ± 6.19 |
| Mamba-3 | 109.82 ± 6.49 | 124.46 ± 4.60 |
| CPTP-SSM, | 9.82 ± 0.71 | 6.03 ± 5.79 |
| CPTP-SSM, | 6.35 ± 0.24 | 1.98 ± 0.27 |
| Gyroscope integration with calibrated bias | 5.75 | 1.76 |
First, CPTP-SSM has by far the lowest error among the learned models, and it is on par with the classical solution of this problem, gyroscope integration with a calibrated bias, although the bias is never given to it. Second, length extrapolation carries over from discrete tasks to continuous real data. Every model is within 1.4° at the training length, and beyond it only CPTP-SSM keeps tracking. Trained on windows of 76.8 s instead, CPTP-SSM with stays the most accurate on EuRoC, with 5.21 ± 0.03° against 14.49 ± 1.16° for the LSTM and 20.75 ± 7.56° for Mamba-3, so the result is not an effect of short training windows. The higher rank helps on real sensors as it does in Table 3.
PD-SSM. It is a baseline in the table above. For Table 6 its released implementation is too slow, as App. C.2 explains. CSP-Bench sequences have up to 12,000 steps, and one training step on JoystickControl took 93 s. Flash PD-SSM removes this cost, but its repository had no model code when we checked on 11 November.
Q2 and Q2.1. Long-sequence benchmarks and Flash PD-SSM
The revision adds Flash PD-SSM to Tables 4 and 5. It has the highest average on the six UEA datasets, 69.1 against 67.6 for CPTP-SSM, which keeps the highest mean on EigenWorms with 96.1 and on Heartbeat with 81.3.
PPG-DaLiA. The table gives the test MSE in units of on sequences of 49,920 steps. Our row is the mean over five seeds, with 27 of the 162 grid configurations searched. The other rows are from Rusch & Rus 2025 and Terzić et al. 2026.
| Model | MSE |
|---|---|
| LinOSS-IM | 6.4 ± 0.23 |
| Flash PD-SSM | 6.5 ± 0.35 |
| LinOSS-IMEX | 7.5 ± 0.46 |
| CPTP-SSM | 9.22 ± 0.80 |
| Log-NCDE | 9.56 ± 0.59 |
| NRDE | 9.90 ± 0.97 |
| Mamba | 10.65 ± 2.20 |
| LRU | 12.17 ± 0.49 |
| S5 | 12.63 ± 1.25 |
| S6 | 12.88 ± 2.05 |
| NCDE | 13.54 ± 0.69 |
We hope this addresses the weaknesses and questions, and we would be grateful if you would reconsider the rating.
Response to Reviewer z33K, part 1 of 2. Complexity and language models
Thank you for the encouraging assessment and the concrete requests. Section, table and appendix numbers below refer to the revised PDF.
Computational and memory complexity
Section 3.2 now states the complexity, and Table 8 in App. B.4 gives the per-token costs. The table below summarizes it for sequence length L, width D, and nq heads of dimension d and Kraus rank r, with state size N=nq(d2−1).
Adding heads is linear in N, while raising d costs d6 per head. The backward pass is the same scan run in reverse on transposed transitions, as App. B.3 describes, and costs 1.81 times the forward pass according to Table 7.
Comparison with other SSMs, Table 9 in App. B.4. One layer of width 1,024 with about 6.7M parameters, batch 4, one A100 80GB, three seeds. PD-SSM runs in FP32 and the others in BF16.
At L=16,384 our layer takes 1.5 times the time of Mamba-3 and 1.8 times that of Gated DeltaProduct, and at L=512 it is the fastest. PD-SSM runs out of the 80 GB of memory at L=16,384. Flash PD-SSM is not included because its repository had no model code when we checked on 11 November.
Potential advantages for language models
We expect the advantage where language meets physical data, as in vision-language-action models. The layer tracks both finite automata and continuous sensor states, and such models need both at once. Section 5 now says so. Whether qubit heads have enough capacity at the scale of a language model and kernels that make them cheap enough there are future work.