acceptodds
Under review as a conference paper at ICLR 2027

LUNA: Generalizable and Trustworthy State Estimation with Large Language Models

Abstract

State estimation infers latent states from noisy and partial observations to provide both state estimate and uncertainty quantification (UQ). Classical Kalman-family filters (KFs) heavily depend upon the adopted system transition and observation dynamic models, suffering from low generalization. Pretraning-model based learnable state estimators aim for high generalization, yet with untrustworthy state estimates due to missed UQ. Particularly, offering UQ in LLM-based state estimators is hard due to missing UQ sources. To this end, in this paper, we propose an LLM-based UNcertainty-aware stAte estimation framework, namely \methodname, to offer trustworthy estimates with high generalization. Moreover, \methodname develops three stages to obtain the principled UQ as classic estimators from both observed systems and LLMs. Experiments on six synthetic and real-world benchmarks show that \methodname achieves the lowest RMSE on five benchmarks and the lowest NLL on four under in-domain evaluation, reducing RMSE by 71.8% on average relative to LLMFilter. Under cross-system evaluation, \methodname reduces RMSE by 41.3% on average relative to the best-performing external baseline. The code is available at an anonymous repositoryhttps://anonymous.4open.science/r/LUNA-V1AC.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Bayesian Learning, Uncertainty Quantification, State Estimation, Large Language Model

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_bp7hOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors provide LUNA, a state-estimation framework that combines a frozen language model with classical Kalman filtering. It propagates sigma points through a learned transition model, predicts additional uncertainty using an evidential regression head, and corrects the resulting state estimate using new observations through a UKF update. Experiments on six synthetic and real-world benchmarks show promising improvements in estimation accuracy and predictive likelihood, alongside cross-system transfer results.

Critical strengths

1- they are addressing an important problem by combining learned state prediction, uncertainty propagation, and classical measurement correction. The predicted covariance directly affect how the filter balances model predictions against new observations.

2- LUNA achieves the good reported RMSE on five of six benchmarks, including a tie, and the best NLL on four.

3- The experiments examine cross-system transfer without parameter updates, using target-system context, and robustness to changing dynamics and noise which is insightful.

4- The paper evaluates multiple backbones, uncertainty components, and transition-network choices. These comparisons help identify which design choices contribute to performance and shows that the framework can also support simpler prediction models.

Critical weaknesses

I list the weaknesses as follow:

1- At each sigma point, the aleatoric and epistemic covariance terms share the same matrix structure and differ only by a scalar factor. Please explain how they enter the predictive likelihood and what allows their separate identification (to me they look identical and referring your figure 3, the epistemic vs full covariance differences is marginal). Controlled experiments would also make your interpretation stronger: with the underlying noise fixed, increasing relevant training data should generally reduce epistemic uncertainty (that I cannot see in the results), while aleatoric uncertainty should approach the underlying process-noise level (how to manipulate it as it stems from your LLM function approximator?).

  1. The method retains the classical filtering (the whole storyline), while introducing a learned prediction model (by your proposed LLM). Please explicitly present the transition likelihood, observation likelihood, and recursive posterior, and identify where the learned transition, sigma-point approximation, and Gaussian approximation enter. This would clarify what is exact under the assumed model and what is approximated. If you discuss it appropriately, then the previous point regarding covariances would clarified as well.

  2. Related to the previous point, please clarify whether the NIW distribution represents a learned conditional distribution over transition means and covariances, or a posterior obtained through an explicit Bayesian update (after 3 times reading, still unclear to me). This distinction matters for interpreting its components as process noise and uncertainty about the transition model. A clearer derivation would support these interpretations.

  3. Although the experiments include Lorenz96 and VL20, etc., the method assumes a known, fixed-dimensional state representation. Its behavior as state dimension increases, or when the number of state variables changes, remains unclear, specially when you use strong function approximator like LLM, such extension worth a check. Please discuss this limitation and the associated computational scaling. A simple example would be inputs as images containing a varying number of moving objects that additionally require perception and state discovery (with reconsideration of your query perhaps as images rather than exact states).

  4. Please report the number of independent training runs and provide means with standard deviations or confidence intervals for the main comparisons. Without these statistics, it is difficult to judge whether the reported improvements are stable, particularly when numerical differences are small (I checked your appendix, but could not find it, otherwise please point me to the correct location in appendix).

  5. The statement that FM-UKF’s prediction step provides no uncertainty quantification should be revised because FM-UKF does sigma-point trajectories through its foundation model and reconstructs predictive covariance using the unscented transform. The statements regarding the baselines need to be fair and accurate.

Questions for authors

My questions are mainly embedded in the weaknesses please see them above.

Besides them, there is one more question:

You are modeling the prediction step with strong function approximator which is LLM. But you switch to conventional classic kalman formulation for update, after observing your observation. Why not apply your context aware LLM for update formulation as well?

6 comments
AuthorsOpenReview ↗

Summary of Paper Revision & Detailed Responses to Reviewer bp7h(1/4)

Summary of Paper Revision

We have revised the paper to address your concerns as follows:

  1. For W1, we have clarified the additive covariance reconstruction in Appendix A and added fixed-noise, varying-data experiments in Section 4.3 (Table 3).

  2. For W2 & W3, we have reorganized Appendix A.1 to present the likelihoods, recursive posterior, and approximation steps, and clarified the learned conditional NIW distribution in Appendix A.3.

  3. For W4, we have added state-dimension experiments in Appendix E.3 (Table 15), runtime and memory measurements in Appendix E.4 (Table 18), and a discussion of dimensional limits and visual state discovery in Section 5.

  4. For W5, we have added five independent training runs with mean and standard deviation in Appendix E.3 (Tables 12–14).

  5. For W6, we have revised the Introduction and Related Work to distinguish FM-UKF's propagated uncertainty from uncertainty in the pretrained model's predictions.

  6. For Q1, we have added Average fusion and Learned-gate fusion experiments in Appendix E.3 (Table 16) and discussed observation-side uncertainty in Section 5.


Detailed Responses

We thank the reviewer for the constructive feedback. We address each concern below with additional experiments, analysis, and clarifications.

W1W1. NIW component identification and data dependence

Predictive likelihood. LUNA incorporates the aleatoric and epistemic terms into the predictive likelihood through their summed covariance, estimated in one NIW forward pass. This additive reconstruction follows the law of total covariance Appendix A\.1, Eqs\. 13–14, with the derivation in Appendix C\.2.

Separate identification. Their shared matrix structure follows from NIW, and the predictive likelihood alone does not uniquely identify their split [1, 2]. NIW defines them as the expected transition covariance and variation in the transition mean, respectively, with $\boldsymbol{\Sigma}^{epi} = \frac{ \boldsymbol{\Sigma}^{ale}}{κ} $ AppendixA3˙Appendix A\.3. Figure 3's small gap indicates a modest additional contribution from the epistemic term in that setting.

Fixed-noise, varying-data experiment. In Table 3, we follow the work [3] and measure the test-set mean tr⁡(Σepi)d\frac{\operatorname{tr}(\boldsymbol{\Sigma}^{\mathrm{epi}})}{d} and tr⁡(Σale)tr⁡(Q)\frac{\operatorname{tr}(\boldsymbol{\Sigma}^{\mathrm{ale}})}{\operatorname{tr}(\mathbf{Q})}, where tr⁡(Σepi)\operatorname{tr}(\boldsymbol{\Sigma}^{\mathrm{epi}}) and tr⁡(Σale)\operatorname{tr}(\boldsymbol{\Sigma}^{\mathrm{ale}}) indicate the total epistemic and aleatoric variances, respectively, d is the state dimension, and tr⁡(Q)\operatorname{tr}(\mathbf{Q}) is the total variance of the true process noise. By varying the training samples eg˙,˙5e\.g\., 5% to 100% on four datasets, we clearly find that the 100% training samples lead to tr⁡(Σepi)d\frac{\operatorname{tr}(\boldsymbol{\Sigma}^{\mathrm{epi}})}{d} close to zero, indicating the least epistemic uncertainty, and tr⁡(Σale)tr⁡(Q)\frac{\operatorname{tr}(\boldsymbol{\Sigma}^{\mathrm{ale}})}{\operatorname{tr}(\mathbf{Q})} close to the true process noise level.

Table 3 in Section 4.3: NIW uncertainty versus training data.

Epistemic: tr⁡(Σepi)d\frac{\operatorname{tr}(\boldsymbol{\Sigma}^{\mathrm{epi}})}{d}.

MethodOscillatorHopfLorenz96VL20
LUNA(Ours), 5% training data0.0790.0050.0740.152
LUNA(Ours), 25% training data0.0030.0030.1170.071
LUNA(Ours), 50% training data0.0040.0040.0550.045
LUNA(Ours), 75% training data0.0040.0020.0380.032
LUNA(Ours), 100% training data0.0030.0030.0240.025

Aleatoric: tr⁡(Σale)tr⁡(Q)\frac{\operatorname{tr}(\boldsymbol{\Sigma}^{\mathrm{ale}})}{\operatorname{tr}(\mathbf{Q})}.

MethodOscillatorHopfLorenz96VL20
LUNA(Ours), 5% training data5.9521.1672.0702.367
LUNA(Ours), 25% training data1.2601.1491.5720.867
LUNA(Ours), 50% training data1.1691.1411.4230.946
LUNA(Ours), 75% training data0.9950.9971.2140.968
LUNA(Ours), 100% training data0.9980.9981.0090.998

[1] Nis Meinert and Alexander Lavin. Multivariate Deep Evidential Regression. arXiv:2104.06135, 2021.

[2] Nis Meinert, Jakob Gawlikowski, and Alexander Lavin. The Unreasonable Effectiveness of Deep Evidential Regression. AAAI, 2023, pp. 9134–9142. DOI: 10.1609/aaai.v37i8.26096.

[3] Luís Marques and Dmitry Berenson. Quantifying Aleatoric and Epistemic Dynamics Uncertainty via Local Conformal Calibration. arXiv:2409.08249, 2024.

AuthorsOpenReview ↗

Summary of Paper Revision & Detailed Responses to Reviewer bp7h(2/4)

W2W2. Likelihoods, recursive posterior, and approximation steps

Probability model and approximations. LUNA learns the transition likelihood p(xt∣xt−1)p(x_t | x_{t−1}), uses the known observation likelihood p(yt∣xt)p(y_t | x_t), and updates the recursive posterior p(xt∣y1:t)p(x_{t} | y_{1:t}) through Bayes' rule. Appendix A.1, Eq. 10 presents these objects together and labels their corresponding modules. The LLM–NIW model specifies the transition likelihood, sigma points approximate predictive moments over the previous recursive posterior, and Gaussian approximations represent both the prediction distribution and recursive posterior Eq\. 11. The Bayesian relations and total-covariance identity Eq\. 13 are exact under the learned model, while the learned transition and moment/distribution approximations make the implemented filter approximate.

W3W3. Meaning of the learned NIW distribution

Learned conditional distribution. NIW represents a learned conditional distribution over transition means and covariances, rather than a posterior obtained through an explicit Bayesian update. Its parameters are predicted from the input state and system context by modules trained offline. During filtering, the UKF updates the posterior over the state, without performing an explicit Bayesian update of the NIW distribution. Appendix A.1 presents the filtering recursion, and A.3 defines the input-dependent NIW distribution.

W4W4. Variable state dimensions and computational scaling

Changing state dimension. OOn four datasets provided by LLMFilter, we evaluate each trained model at its original and reduced state dimensions without retraining. Full uses the original state dimension. Half and Quarter retain the first half and first quarter of the state dimensions in their original order, respectively. For Lorenz96, these correspond to the first 36 and first 18 dimensions of the original 72-dimensional state. Table 15 in Appendix E.3 reports RMSE and joint NLL, with lower values indicating better performance at the same state dimension. LUNA maintains similar estimation accuracy across the tested dimensions, with RMSE differing from the full-state result by less than 0.9%.

Table 15 in Appendix E.3: State dimension sensitivity.

DatasetFull, RMSEFull, NLLHalf, RMSEHalf, NLLQuarter, RMSEQuarter, NLL
Oscillator0.8032.3900.7961.187N/AN/A
Hopf0.7692.3150.7641.151N/AN/A
Lorenz964.017202.7523.987101.1763.98350.579
VL200.88193.1370.87946.5130.88023.274

N/A indicates that the quarter-state condition was not evaluated for the two-dimensional systems.

Computational scaling. A d-dimensional state requires 2d+1 sigma-point transition evaluations and d² entries per full covariance. The separate timing study uses six backbones on 2D Oscillator and 72D Lorenz96, with synchronized complete-step timing and peak allocated memory AppendixE4˙,Table18Appendix E\.4, Table 18. It shows higher latency and nearly constant backbone-dominated memory at these operating points. Because system identity also changes, this comparison is not an isolated dimension-scaling experiment.

Table 18 in Appendix E.4: Runtime and memory for six backbones.

RTX 4090.

BackboneOscillator time ms/stepms/stepOscillator memory GBGBLorenz96 time ms/stepms/stepLorenz96 memory GBGB
MLP3.5130.0139.6270.053
RNN3.8310.02111.7360.378
Transformer4.0540.03818.4430.530
TimesFM-2.518.1040.94124.3980.987
TimesFM-3.020.2811.33827.6311.374
LLaMA-7B (LUNA)20.48813.25135.72313.316

H800.

BackboneOscillator time ms/stepms/stepOscillator memory GBGBLorenz96 time ms/stepms/stepLorenz96 memory GBGB
MLP5.2940.03813.5120.078
RNN5.6380.06315.8680.420
Transformer6.1730.06316.9900.555
TimesFM-2.525.8400.96634.5661.012
TimesFM-3.029.3051.36340.3161.399
LLaMA-7B (LUNA)22.50513.27630.60313.341

Images and changing object counts. The coordinate-subset tests do not evaluate images with varying object counts or online state discovery. Those settings require perception and object identity tracking beyond the predefined state representation. Section 5 discusses this distinction, the restriction to reduced dimensions, and the need for more scalable uncertainty propagation.

AuthorsOpenReview ↗

Summary of Paper Revision & Detailed Responses to Reviewer bp7h(3/4)

W5W5. Stability across independent training runs

Training stability. Using the datasets provided by LLMFilter, we initialize and train each learned method's trainable modules with five seeds. Tables 13–14 of Appendix E.3 report mean ± standard deviation of RMSE and NLL, with lower means better. LUNA maintains stable estimation accuracy, with RMSE standard deviations of 0.4–2.3% of the corresponding means, and achieves the lowest mean RMSE and NLL among learned methods on all five datasets.

Table 13 in Appendix E.3: Five-seed RMSE comparison on simulated datasets.

MethodOscillatorHopfTrackingLorenz96VL20
EKF†0.8010.7681.2993.0960.785
UKF†0.7980.7681.2993.0810.785
FM-UKF0.978±0.061‾\underline{0.978 \pm 0.061}0.837±0.033‾\underline{0.837 \pm 0.033}11.127 ± 2.53012.984 ± 0.3985.738 ± 0.363
LLMFilter3.573 ± 0.0532.009 ± 0.03714.827 ± 0.7016.792 ± 0.0086.589 ± 0.014
LLMFilter+UKF1.955 ± 0.0281.133 ± 0.01813.324 ± 0.7806.722 ± 0.0063.321 ± 0.004
MOP0.983 ± 0.0040.893 ± 0.0082.393±0.035‾\underline{2.393 \pm 0.035}5.667±0.012‾\underline{5.667 \pm 0.012}2.665±0.045‾\underline{2.665 \pm 0.045}
LUNA (Ours)0.805 ± 0.0050.773 ± 0.0031.947 ± 0.0454.045 ± 0.0360.882 ± 0.012

Table 13 in Appendix E.3: Five-seed NLL comparison on simulated datasets.

MethodOscillatorHopfTrackingLorenz96VL20
EKF†2.392.316.35180.5784.73
UKF†2.382.316.35180.2884.73
FM-UKF2.92±0.08‾\underline{2.92 \pm 0.08}2.72±0.03‾\underline{2.72 \pm 0.03}548.97±424.00‾\underline{548.97 \pm 424.00}4371.01 ± 296.791102.54 ± 135.52
LLMFilterN/AN/AN/AN/AN/A
LLMFilter+UKF8.68 ± 0.213.69 ± 0.081578.96 ± 376.401675.28±23.54‾\underline{1675.28 \pm 23.54}832.63±1.58‾\underline{832.63 \pm 1.58}
MOPN/AN/AN/AN/AN/A
LUNA (Ours)2.40 ± 0.012.33 ± 0.018.37 ± 1.27203.25 ± 0.9093.23 ± 0.99

N/A indicates that no covariance-based NLL is reported.

Real-world variability. In Table 12, we evaluate RMSE and NLL on UTIAS and KITTI tworeal−worlddatasetstwo real-world datasets over five independent training runs. Entries are mean ± standard deviation, with lower values better for both metrics. The UTIAS results below show larger variability than the datasets provided by LLMFilter: LUNA's RMSE is 0.711 ± 0.063, with a standard deviation of approximately 8.9% of its mean, and NLL also varies more across runs. LUNA still achieves the lowest mean RMSE and NLL among learned methods.

Table 12 in Appendix E.3: RMSEs on two real datasets UTIAS and KITTI results.

MethodUTIAS<br>RMSE mmUTIAS<br>NLLKITTI<br>RMSE mmKITTI<br>NLL
EKF†0.2462.1402.9806.920
UKF†0.2351.8002.9806.920
FM-UKF0.862 ± 0.05920.780 ± 10.8933.260 ± 0.8079.550 ± 3.099
LLMFilter1.270 ± 0.015N/A4.310 ± 0.002N/A
LLMFilter+UKF1.250 ± 0.013209.000 ± 3.7104.050 ± 0.01711.400 ± 0.082
MOP2.110 ± 0.009N/A9.840 ± 3.340N/A
LUNA(Ours)0.711 ± 0.06318.900 ± 8.9002.790 ± 0.6035.420 ± 2.770

W6W6. Fair characterization of FM-UKF

FM-UKF uncertainty quantification. The Introduction and Related Work now acknowledge that FM-UKF reconstructs predictive covariance from sigma-point trajectories using the unscented transform. The intended distinction is that LUNA additionally estimates input-dependent uncertainty in the learned transition through NIW, rather than suggesting that FM-UKF provides no prediction-side UQ. Appendix A.1 makes these two covariance sources explicit.

Reviewer_ZjznOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LUNA, an LLM-based state estimation framework that explicitly models uncertainty within a prediction–update pipeline. It combines sigma-point propagation with evidential regression to capture propagated state uncertainty and model-induced aleatoric and epistemic uncertainty, followed by a UKF measurement update. Experiments on six synthetic and real-world benchmarks demonstrate improved estimation accuracy and uncertainty quantification in most settings, alongside cross-system transfer results and ablations supporting the proposed components.

Critical strengths
  1. Task-appropriate uncertainty modeling. LUNA meaningfully extends existing LLM-based state estimators by explicitly modeling uncertainty within a prediction–update framework. This design aligns well with the probabilistic nature of state estimation, allowing uncertainty to guide measurement correction and improve estimation accuracy.

  2. Clear presentation and convincing empirical support. The paper presents a coherent connection between its motivation, method, and results. Comparisons across six benchmarks, cross-system evaluations, and component-wise ablations provide convincing evidence for the effectiveness of the proposed framework.

Critical weaknesses
  1. Limited real-world evaluation. Only one of the six benchmarks uses real-world data, and cross-system transfer is evaluated entirely on simulated systems. Additional real-world datasets would strengthen the evidence for the framework’s generalization and uncertainty calibration in practical settings.

  2. Insufficient evidence for deployment efficiency. The 7B backbone raises concerns about computational and memory requirements. Although the paper reports an average latency of 25.23 ms per step, further evaluation across hardware platforms and state dimensions, including latency and memory comparisons with smaller backbones, would better establish its suitability for real-time applications.

Questions for authors
  1. Could the authors provide additional real-world evaluations, particularly under changing dynamics or noise conditions, to further validate generalization and uncertainty calibration?
  2. Could the authors clarify the hardware and timing protocol behind the reported 25.23 ms per step? Does this include the complete filtering pipeline, and how do latency and memory consumption scale with state dimensionality?
  3. Could the authors include MLP- or RNN-based variants in the cross-system evaluations to better isolate the contribution of the pretrained LLM beyond the proposed filtering and uncertainty modeling framework?

Overall, I find the paper well written, with clear motivation and a coherent methodological presentation. I would consider raising my score if the authors satisfactorily address these questions.

2 comments
AuthorsOpenReview ↗

Summary of Paper Revision & Detailed Responses to Reviewer Zjzn(1/2)

Summary of Paper Revision

We have revised the paper to address your concerns as follows:

  1. For W1 & Q1, we have added two real-world datasets UTIAS and KITTI results in Appendix E.3 (Table 12), included prediction-coverage evaluation in Section 4.3 (Table 2), and extended Figure 5 with combined dynamics and noise changes in simulated Tracking.

  2. For W2 & Q2, we have added backbone runtime and memory comparisons on RTX 4090 and H800 in Appendix E.4 (Table 18), covering low- and high-dimensional systems.

  3. For Q3, we have expanded the backbone comparison in Appendix E.1 (Table 8) with MLP and RNN variants while retaining the filtering and uncertainty modules.


Detailed Responses

We thank the reviewer for the constructive feedback. We address each concern below with additional experiments, analysis, and clarifications.

W1&Q1W1 \& Q1. Real-world evaluation and changing conditions

To perform the evaluation in challenging settings, we consider three motion manners: straight motion AA, curved motion BB, and the curved motion with higher process noise CC, and evaluate LUNA with three stages: A, A+B, A+B+C, where the first stage involves the straight motion alone and the third one involves all of three manners. Table 12 of Appendix E.3 gives the RMSEs and NLL uncertainty of 7 approaches on two newly added real datasets (UTIAS and KITTI). As shown in this table, LUNA leads to the lowest RMSEs and NLL uncertainties on both datasets.

Table 12 in Appendix E.3: RMSEs on two real datasets UTIAS and KITTI results.

MethodUTIAS<br>RMSE mmUTIAS<br>NLLKITTI<br>RMSE mmKITTI<br>NLL
EKF†0.2462.1402.9806.920
UKF†0.2351.8002.9806.920
FM-UKF0.862 ± 0.05920.780 ± 10.8933.260 ± 0.8079.550 ± 3.099
LLMFilter1.270 ± 0.015N/A4.310 ± 0.002N/A
LLMFilter+UKF1.250 ± 0.013209.000 ± 3.7104.050 ± 0.01711.400 ± 0.082
MOP2.110 ± 0.009N/A9.840 ± 3.340N/A
LUNA(Ours)0.711 ± 0.06318.900 ± 8.9002.790 ± 0.6035.420 ± 2.770

† EKF/UKF are deterministic physical-model references. N/A indicates that no covariance-based NLL is reported.

W2&Q2W2 \& Q2. Timing protocol, memory, and dimensional scaling

In Table 18 of Appendix E.4, we compare runtime and memory across six backbones on RTX 4090 48GB48 GB and H800 80GB80 GB. Time measures a complete state-estimation step, including prediction and UKF update. Memory is peak allocated GPU memory in GB, with lower values better for both. Smaller models require substantially less time and memory, while the LLM offers the transfer advantage above.

Table 18 in Appendix E.4: Runtime and memory for six backbones.

RTX 4090.

BackboneOscillator, time ms/stepms/stepOscillator, memory GBGBLorenz96, time ms/stepms/stepLorenz96, memory GBGB
MLP3.5130.0139.6270.053
RNN3.8310.02111.7360.378
Transformer4.0540.03818.4430.530
TimesFM-2.518.1040.94124.3980.987
TimesFM-3.020.2811.33827.6311.374
LLaMA-7B (LUNA)20.48813.25135.72313.316

H800.

BackboneOscillator, time ms/stepms/stepOscillator, memory GBGBLorenz96, time ms/stepms/stepLorenz96, memory GBGB
MLP5.2940.03813.5120.078
RNN5.6380.06315.8680.420
Transformer6.1730.06316.9900.555
TimesFM-2.525.8400.96634.5661.012
TimesFM-3.029.3051.36340.3161.399
LLaMA-7B (LUNA)22.50513.27630.60313.341
AuthorsOpenReview ↗

Summary of Paper Revision & Detailed Responses to Reviewer Zjzn(2/2)

Q3Q3. Controlled MLP and RNN cross-system comparisons

In the following table, we add the new results of four variants LUNA−FM,LUNA−Transformer,LUNA−RNN,LUNA−MLPLUNA-FM, LUNA-Transformer, LUNA-RNN, LUNA-MLP to perform the in-domain and cross-domain evaluation. In this table, our full LUNA model achieves the lowest RMSE among all LUNA variants. In the high-dimensional setting eg˙,˙Lorenz96andVL20e\.g\., Lorenz96 and VL20, the backone MLP suffers from the highest errors, and the low one eg˙,˙HopfandOscillatore\.g\., Hopf and Oscillator, either Transformer or RNN is with the worse performance.

Table 8 in Appendix E.1: In-domain / cross-domain state-estimation RMSE for different backbones of LUNA.(left of →: training system, right of →: test system).

In-domain state-estimation RMSE

MethodOscillator→ OscillatorHopf → HopfLorenz96 → Lorenz96VL20 → VL20
LUNA-FM0.3850.2080.5170.276
LUNA-Transformer0.4020.2130.5080.262
LUNA-RNN0.4080.2110.5220.287
LUNA-MLP0.4010.2100.5490.291
LUNA-O0.3950.2110.4400.264
LUNA (ours)0.3780.2070.4310.252

Cross-domain state-estimation RMSE

MethodHopf → OscillatorOscillator → HopfVL20 → Lorenz96Lorenz96 → VL20
LUNA-FM1.8360.9671.2740.538
LUNA-Transformer1.8420.9621.7780.634
LUNA-RNN1.8480.9481.8030.607
LUNA-MLP1.8430.9531.6960.641
LUNA-O1.1500.8721.2170.553
LUNA (ours)0.9120.7040.9820.387
Reviewer_YLiXOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LUNA, an LLM-based state estimation framework that combines pretrained language models with Kalman filtering to provide state estimates and uncertainty quantification. The method propagates sigma points through an LLM-based dynamics predictor and estimates uncertainty. Experiments across six benchmarks demonstrate improvements in estimation accuracy, uncertainty quantification, and cross-system generalization.

Critical strengths
  1. Reasonable methodological approach: The paper presents a sensible integration of pretrained models with classical filtering and uncertainty estimation techniques.

  2. Efficient uncertainty estimation: The proposed framework provides uncertainty quantification with relatively low computational overhead compared with iterative alternatives, while achieving competitive performance against existing neural and LLM-based filters.

Critical weaknesses
  1. Limited experimental complexity: The evaluation focuses on relatively simple dynamical systems, with limited evidence of significant improvements over existing methods in challenging settings. More complex scenarios where conventional approaches struggle would better demonstrate the practical significance of LUNA.

  2. Limited generalization claims: Despite emphasizing generalization across systems, the method still requires system-specific training data for its learned adapters. Cross-system evaluation also relies on target-system examples, limiting the strength of the generalization claims.

  3. Insufficient baseline comparisons: The paper lacks comparisons with approaches that train smaller transformers directly for state estimation, as explored in [1]. Such baselines would help establish whether pretrained LLMs provide meaningful advantages over simpler architectures.

  4. Limited methodological novelty: Learning state transitions using pretrained LLMs and task-specific adapters has been explored previously. Given such a learned dynamics model, the remaining framework largely combines standard sigma-point filtering with established evidential uncertainty estimation techniques, leaving the methodological novelty unclear.

[1] Du et. al. "Can transformers learn optimal filtering for unknown systems?"

Questions for authors
  1. Unified model training: What happens if a single dynamics model is trained jointly on data from all available systems? Would such a model achieve comparable generalization without requiring pretrained LLMs and system-specific adapters?

  2. Necessity of LLMs: What advantages do pretrained LLMs provide over conventional learned dynamics models or world models trained directly on state-transition data? Could a smaller model achieve similar estimation accuracy and generalization at substantially lower computational cost?

4 comments
AuthorsOpenReview ↗

Summary of Paper Revision & Detailed Responses to Reviewer YLiX(1/3)

Summary of Paper Revision

We have revised the paper to address your concerns as follows:

  1. For W1, we have extended Figure 5 with combined dynamics and noise changes and added real-world datasets UTIAS and KITTI results in Appendix E.3 (Table 12).

  2. For W2, we have added target-context and source-training-data sensitivity experiments in Appendix E.1 (Tables 9–10).

  3. For W3, we have added MOP to the baseline descriptions and five-seed comparisons in Appendices D.4 and E.3 (Tables 13–14).

  4. For W4, we have clarified LUNA’s framework contribution and how sigma-point propagation, NIW uncertainty estimation, state-change prediction, and residual correction address the challenges of integrating LLMs into traditional filtering, supported by the design and covariance ablations in Table 4 and Figure 4.

  5. For Q1, we have expanded the backbone comparison to include TimesFM as a pretrained time-series reference in Appendix E.1 (Figure 8).

  6. For Q2, we have added backbone runtime and memory comparisons on RTX 4090 and H800 in Appendix E.4 (Table 18).


Detailed Responses

We thank the reviewer for the constructive feedback. We address each concern below with additional experiments, analysis, and clarifications.

W1W1. Evaluation in challenging settings

To perform the evaluation in challenging settings, we consider three motion manners: straight motion AA, curved motion BB, and the curved motion with higher process noise CC, and evaluate LUNA with three stages: A, A+B, A+B+C, where the first stage involves the straight motion alone, and the third one involves all of three manners. Table 12 of Appendix E.3 gives the RMSEs and NLL uncertainty of 7 approaches on two newly added real datasets (UTIAS and KITTI). As shown in this table, LUNA leads to the lowest RMSEs and NLL uncertainties on both datasets.

Table 12 in Appendix E.3: RMSEs on two real datasets UTIAS and KITTI results.

MethodUTIAS<br>RMSE mmUTIAS<br>NLLKITTI<br>RMSE mmKITTI<br>NLL
EKF†0.2462.1402.9806.920
UKF†0.2351.8002.9806.920
FM-UKF0.862 ± 0.05920.780 ± 10.8933.260 ± 0.8079.550 ± 3.099
LLMFilter1.270 ± 0.015N/A4.310 ± 0.002N/A
LLMFilter+UKF1.250 ± 0.013209.000 ± 3.7104.050 ± 0.01711.400 ± 0.082
MOP2.110 ± 0.009N/A9.840 ± 3.340N/A
LUNA(Ours)0.711 ± 0.06318.900 ± 8.9002.790 ± 0.6035.420 ± 2.770

† EKF/UKF are deterministic physical-model references. N/A indicates that no covariance-based NLL is reported.

W2W2. Source-training and target-context dependence

Source-system training. In Table 10 of Appendix E.1, we vary the source training data samples from 5% to 100%, and give the cross-domain evaluation results. We find that increasing source-system training data consistently improves cross-system transfer across all four cross-domain tests.

Table 10 in Appendix E.1: RMSEs of Cross-Domain Evaluation with Various Source Training Data Samples.

Source → Target5%25%50%75%100%
Hopf → Oscillator1.0570.9790.9460.9260.912
Oscillator → Hopf0.9460.8160.7600.7270.704
VL20 → Lorenz961.2331.0981.0401.0060.982
Lorenz96 → VL200.4590.4200.4040.3940.387

Target-state examples. In Table 9, we study the sensitity of LUNA to various state target context with 0, 3, 5, and 8 target-state examples and measure the RMSE of LUNA in the cross-domain settings. As shown in this table, the 8 target-state examples can achieve the least RMSEs for all four settings.

Table 9 in Appendix E.1: Cross-domain RMSE versus the number of target-state examples.

Source → Target0 states3 states5 states (default)8 states
Hopf → Oscillator1.1500.9100.9120.910
Oscillator → Hopf0.8720.7000.7040.700
VL20 → Lorenz961.2171.0070.9820.962
Lorenz96 → VL200.5530.4270.3870.371
AuthorsOpenReview ↗

Summary of Paper Revision & Detailed Responses to Reviewer YLiX(2/3)

W3W3. Comparison with smaller Transformers

In Table 12-14 of Appendix E.3, we add the suggested MOP baseline [1] and evaluate it on five simulated datasets provided by LLMFilter and two real-world datasets. Entries report mean ± standard deviation of state-estimation RMSE over five independent training runs, with lower values better. LUNA achieves the lowest learned-method mean on all five datasets, reducing RMSE relative to MOP by 13.4–66.9%. LUNA benefits from the known observation model and observation-noise covariance for UKF update steps, whereas MOP predicts outputs directly.

Table 12 in Appendix E.3: RMSEs on two real datasets UTIAS and KITTI results.

MethodUTIAS,<br>RMSE mmUTIAS,<br>NLLKITTI,<br>RMSE mmKITTI,<br>NLL
EKF†0.2462.1402.9806.920
UKF†0.2351.8002.9806.920
FM-UKF0.862 ± 0.05920.780 ± 10.8933.260 ± 0.8079.550 ± 3.099
LLMFilter1.270 ± 0.015N/A4.310 ± 0.002N/A
LLMFilter+UKF1.250 ± 0.013209.000 ± 3.7104.050 ± 0.01711.400 ± 0.082
MOP2.110 ± 0.009N/A9.840 ± 3.340N/A
LUNA(Ours)0.711 ± 0.06318.900 ± 8.9002.790 ± 0.6035.420 ± 2.770

Table 13 in Appendix E.3: Five-seed RMSE comparison on simulated dataset.

MethodOscillatorHopfTrackingLorenz96VL20
EKF†0.8010.7681.2993.0960.785
UKF†0.7980.7681.2993.0810.785
FM-UKF0.978±0.061‾\underline{0.978 \pm 0.061}0.837±0.033‾\underline{0.837 \pm 0.033}11.127 ± 2.53012.984 ± 0.3985.738 ± 0.363
LLMFilter3.573 ± 0.0532.009 ± 0.03714.827 ± 0.7016.792 ± 0.0086.589 ± 0.014
LLMFilter+UKF1.955 ± 0.0281.133 ± 0.01813.324 ± 0.7806.722 ± 0.0063.321 ± 0.004
MOP0.983 ± 0.0040.893 ± 0.0082.393±0.035‾\underline{2.393 \pm 0.035}5.667±0.012‾\underline{5.667 \pm 0.012}2.665±0.045‾\underline{2.665 \pm 0.045}
LUNA (Ours)0.805 ± 0.0050.773 ± 0.0031.947 ± 0.0454.045 ± 0.0360.882 ± 0.012

Table 14 in Appendix E.3: Five-seed NLL comparison on simulated dataset.

MethodOscillatorHopfTrackingLorenz96VL20
EKF†2.392.316.35180.5784.73
UKF†2.382.316.35180.2884.73
FM-UKF2.92±0.08‾\underline{2.92 \pm 0.08}2.72±0.03‾\underline{2.72 \pm 0.03}548.97±424.00‾\underline{548.97 \pm 424.00}4371.01 ± 296.791102.54 ± 135.52
LLMFilterN/AN/AN/AN/AN/A
LLMFilter+UKF8.68 ± 0.213.69 ± 0.081578.96 ± 376.401675.28±23.54‾\underline{1675.28 \pm 23.54}832.63±1.58‾\underline{832.63 \pm 1.58}
MOPN/AN/AN/AN/AN/A
LUNA (Ours)2.40 ± 0.012.33 ± 0.018.37 ± 1.27203.25 ± 0.9093.23 ± 0.99

N/A indicates that no covariance-based NLL is reported.

[1] Zhe Du, Haldun Balim, Samet Oymak, and Necmiye Ozay. Can Transformers Learn Optimal Filtering for Unknown Systems? IEEE Control Systems Letters, 7:3525–3530, 2023. DOI: 10.1109/LCSYS.2023.3335318.

W4W4. Methodological contribution beyond existing components

Bayesian prediction construction. Our contribution is an LLM-based uncertainty-aware state estimation framework and the designs that address the challenges of integrating LLMs into traditional filtering. The framework combines the generalization capabilities of LLMs with the prediction-update structure of traditional filters to support accurate and generalizable state estimation. This integration requires uncertainty quantification and an effective mapping between token embeddings and continuous states. We combine sigma-point propagation with NIW evidential regression to account for uncertainty from both the previous state distribution and the learned transition, reusing the transition embeddings without additional repeated LLM sampling for uncertainty estimation. We further use one-step state-change prediction and residual correction to improve the accuracy of continuous-state predictions decoded from LLM embeddings. These techniques serve specific requirements of the proposed framework, enabling LLM-based predictions and their uncertainty to be used consistently in recursive filtering.

Q1Q1. Unified model training

We use TimesFM, pretrained on a broad time-series corpus, as related evidence for a unified pretrained dynamics model. Table 8 in Appendix E.1 compares it with LUNA on four source → target transfers, retaining the filtering and uncertainty modules. Lower RMSE is better. LUNA reduces cross-system RMSE by 32.13% on average.

AuthorsOpenReview ↗

Summary of Paper Revision & Detailed Responses to Reviewer YLiX(3/3)

Q2Q2. Advantages and cost relative to smaller learned models

In the following table, we add the new RMSE results of four variants LUNA−FM,LUNA−Transformer,LUNA−RNN,LUNA−MLPLUNA-FM, LUNA-Transformer, LUNA-RNN, LUNA-MLP to perform the in-domain and cross-domain evaluation. In this table, our full LUNA model achieves the least RMSE among all LUNA variants. In the high-dimensional setting eg˙,˙Lorenz96andVL20e\.g\., Lorenz96 and VL20, the backone MLP suffers from the highest errors, and the low one eg˙,˙HopfandOscillatore\.g\., Hopf and Oscillator, either Transformer or RNN is with the worse performance.

Table 8 in Appendix E.1: In-domain / cross-domain state-estimation RMSE for different backbones of LUNA.(left of →: training system, right of →: test system).

In-domain state-estimation RMSE

MethodOscillator→ OscillatorHopf → HopfLorenz96 → Lorenz96VL20 → VL20
LUNA-FM0.3850.2080.5170.276
LUNA-Transformer0.4020.2130.5080.262
LUNA-RNN0.4080.2110.5220.287
LUNA-MLP0.4010.2100.5490.291
LUNA-O0.3950.2110.4400.264
LUNA (ours)0.3780.2070.4310.252

Cross-domain state-estimation RMSE

MethodHopf → OscillatorOscillator → HopfVL20 → Lorenz96Lorenz96 → VL20
LUNA-FM1.8360.9671.2740.538
LUNA-Transformer1.8420.9621.7780.634
LUNA-RNN1.8480.9481.8030.607
LUNA-MLP1.8430.9531.6960.641
LUNA-O1.1500.8721.2170.553
LUNA (ours)0.9120.7040.9820.387

Runtime and memory. In Table 18 of Appendix E.4, we compare runtime and memory across six backbones on RTX 4090 48GB48 GB and H800 80GB80 GB. Time measures a complete state-estimation step, including prediction and UKF update. Memory is peak allocated GPU memory in GB, with lower values better for both. Smaller models require substantially less time and memory, while the LLM offers the transfer advantage above.

Table 18 in Appendix E.4: Runtime and memory for six backbones.

RTX 4090.

BackboneOscillator, time ms/stepms/stepOscillator, memory GBGBLorenz96, time ms/stepms/stepLorenz96, memory GBGB
MLP3.5130.0139.6270.053
RNN3.8310.02111.7360.378
Transformer4.0540.03818.4430.530
TimesFM-2.518.1040.94124.3980.987
TimesFM-3.020.2811.33827.6311.374
LLaMA-7B (LUNA)20.48813.25135.72313.316

H800.

BackboneOscillator, time ms/stepms/stepOscillator, memory GBGBLorenz96, time ms/stepms/stepLorenz96, memory GBGB
MLP5.2940.03813.5120.078
RNN5.6380.06315.8680.420
Transformer6.1730.06316.9900.555
TimesFM-2.525.8400.96634.5661.012
TimesFM-3.029.3051.36340.3161.399
LLaMA-7B (LUNA)22.50513.27630.60313.341

I thank the authors for their efforts and have reconsidered my initial assessment. In particular, I find the demonstrated improvements over MOP convincing.

Reviewer_kGH1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes LUNA, an LLM-based state estimation framework that combines pretrained language models with a recursive prediction-update filtering procedure. Its prediction step propagates the previous state distribution through deterministic sigma points, estimates LLM-induced aleatoric and epistemic uncertainties using Normal-Inverse-Wishart evidential regression, and combines the uncertainty components through the law of total covariance. A measurement update then produces the final state estimate. Experiments on six synthetic and real-world benchmarks evaluate estimation accuracy, probabilistic quality, cross-system generalization, and uncertainty ablations. LUNA achieves the best RMSE on five of six benchmarks and the best NLL on four, and reports a 41.3% average RMSE reduction over the strongest external baseline across four cross-system transfer settings.

Critical strengths

1. A coherent integration of state estimation and uncertainty modeling. LUNA combines sigma-point propagation, evidential regression, and a recursive measurement update to explicitly account for multiple uncertainty sources. This is a meaningful design direction for improving the reliability of pretrained-model-based state estimation, rather than relying solely on point predictions.

2. Broad empirical evaluation with encouraging transfer results. The evaluation covers six benchmarks spanning linear, nonlinear, chaotic, and real-world localization settings. The cross-system experiments and component ablations provide useful evidence that uncertainty modeling contributes to robust filtering, with particularly promising improvements over existing pretrained-model-based baselines.

Critical weaknesses

1. Uncertainty quantification is insufficiently validated. The calibration analysis relies mainly on an aggregate error-to-covariance ratio on one benchmark, which does not establish reliable calibration across systems or validate the individual uncertainty components.

2. The source of generalization gains is unclear. The framework combines several components, making it difficult to isolate the contributions of pretrained LLM knowledge and uncertainty modeling. The strong in-domain performance of the MLP variant further motivates more controlled comparisons to substantiate the claimed benefits of LLMs.

Questions for authors
  1. Uncertainty calibration: Can you provide calibration results across benchmarks to validate the reliability of individual uncertainty components?

  2. Generalization gains: How much does the pretrained LLM contribute beyond the filtering architecture and uncertainty modeling?

  3. Transfer robustness: How sensitive are the cross-system results to the availability of the five-state target context?

2 comments
AuthorsOpenReview ↗

Summary of Paper Revision & Detailed Responses to Reviewer kGH1(1/2)

Summary of Paper Revision

We have revised the paper to address reviewer kGH1's concerns and questions as follows:

  1. For W1 & Q1, we have expanded the calibration evaluation in Section 4.3 with cross-benchmark coverage and component-removal results (Table 2), together with fixed-noise, varying-data diagnostics (Table 3), to assess overall calibration and the behavior of individual uncertainty components.

  2. For W2 & Q2, we have expanded the prediction-backbone comparison in Appendix E.1 (Figure 8), keeping the filtering and uncertainty modules fixed. Together with the covariance ablations in Figure 4, this distinguishes the tested backbone benefit from the contribution of uncertainty reconstruction.

  3. For Q3, we have added target-context sensitivity experiments to Appendix E.1 (Table 9), varying the number of target-state examples while keeping the source-trained parameters fixed, to quantify transfer dependence on the default context.


Detailed Responses

We thank the reviewer for the constructive feedback. We address each concern below with additional experiments, analysis, and clarifications.

W1&Q1W1 \& Q1. Calibration across systems and individual components

Cross-benchmark calibration. To sufficiently validate uncertainty quantification UQUQ, we give the UQ result on all fix datasets in Table 2 of Section 4.3, where each cell value indicates the Cov90 value. The Cov90 value, if closer to 90%, indicates a better result. In this experiment, for each data set, we remove one UQ compoenent eg˙,˙epistemicoraleatoricorLLMtransitione\.g\., epistemic or aleatoric or LLM transition from the full LUNA model, and then measure the Cov90 value. This table clearly demonstrates that the full LUNA model is with the Cov90 value nearest to 90%.

Table 2 in Section 4.3: Cov90 % under covariance-only ablations on six main-paper benchmarks.

DatasetFull LUNAw/o Epistemicw/o Aleatoricw/o LLM transition
Oscillator88.487.473.936.8
Hopf90.087.168.552.1
Lorenz9689.387.057.729.1
VL2090.881.716.10.8
Tracking90.589.481.513.1
PosNCLT89.989.689.011.8

Individual components. In Table 3, we follow the work [1] and measure the test-set mean tr⁡(Σepi)d\frac{\operatorname{tr}(\boldsymbol{\Sigma}^{\mathrm{epi}})}{d} and tr⁡(Σale)tr⁡(Q)\frac{\operatorname{tr}(\boldsymbol{\Sigma}^{\mathrm{ale}})}{\operatorname{tr}(\mathbf{Q})}, where tr⁡(Σepi)\operatorname{tr}(\boldsymbol{\Sigma}^{\mathrm{epi}}) and tr⁡(Σale)\operatorname{tr}(\boldsymbol{\Sigma}^{\mathrm{ale}}) indicate the total epistemic and aleatoric variances, respectively, d is the state dimension, and tr⁡(Q)\operatorname{tr}(\mathbf{Q}) is the total variance of the true process noise. By varying the training samples eg˙,˙5e\.g\., 5% to 100% on four datasets, we clearly find that the 100% training samples lead to tr⁡(Σepi)d\frac{\operatorname{tr}(\boldsymbol{\Sigma}^{\mathrm{epi}})}{d} close to zero, indicating the least epistemic uncertainty, and tr⁡(Σale)tr⁡(Q)\frac{\operatorname{tr}(\boldsymbol{\Sigma}^{\mathrm{ale}})}{\operatorname{tr}(\mathbf{Q})} close to the true process noise level.

Table 3 in Section 4.3: NIW uncertainty versus training data.

Epistemic: tr⁡(Σepi)d\frac{\operatorname{tr}(\boldsymbol{\Sigma}^{\mathrm{epi}})}{d}.

MethodOscillatorHopfLorenz96VL20
LUNA(Ours), 5% training data0.0790.0050.0740.152
LUNA(Ours), 25% training data0.0030.0030.1170.071
LUNA(Ours), 50% training data0.0040.0040.0550.045
LUNA(Ours), 75% training data0.0040.0020.0380.032
LUNA(Ours), 100% training data0.0030.0030.0240.025

Aleatoric: tr⁡(Σale)tr⁡(Q)\frac{\operatorname{tr}(\boldsymbol{\Sigma}^{\mathrm{ale}})}{\operatorname{tr}(\mathbf{Q})}.

MethodOscillatorHopfLorenz96VL20
LUNA(Ours), 5% training data5.9521.1672.0702.367
LUNA(Ours), 25% training data1.2601.1491.5720.867
LUNA(Ours), 50% training data1.1691.1411.4230.946
LUNA(Ours), 75% training data0.9950.9971.2140.968
LUNA(Ours), 100% training data0.9980.9981.0090.998
AuthorsOpenReview ↗

Summary of Paper Revision & Detailed Responses to Reviewer kGH1(2/2)

W2&Q2W2 \& Q2. Backbone contribution beyond filtering and uncertainty modeling

In the following table, we add the new results of four variants LUNA−FM,LUNA−Transformer,LUNA−RNN,LUNA−MLPLUNA-FM, LUNA-Transformer, LUNA-RNN, LUNA-MLP to perform the in-domain and cross-domain evaluation. In this table, our full LUNA model achieves the least RMSE among all LUNA variants. In the high-dimensional setting eg˙,˙Lorenz96andVL20e\.g\., Lorenz96 and VL20, the backone MLP suffers from the highest errors, and the low one eg˙,˙HopfandOscillatore\.g\., Hopf and Oscillator, either Transformer or RNN is with the worse performance.

Table 8 in Appendix E.1: In-domain / cross-domain state-estimation RMSE for different backbones of LUNA.(left of →: training system, right of →: test system).

In-domain state-estimation RMSE

MethodOscillator→ OscillatorHopf → HopfLorenz96 → Lorenz96VL20 → VL20
LUNA-FM0.3850.2080.5170.276
LUNA-Transformer0.4020.2130.5080.262
LUNA-RNN0.4080.2110.5220.287
LUNA-MLP0.4010.2100.5490.291
LUNA-O0.3950.2110.4400.264
LUNA (ours)0.3780.2070.4310.252

Cross-domain state-estimation RMSE

MethodHopf → OscillatorOscillator → HopfVL20 → Lorenz96Lorenz96 → VL20
LUNA-FM1.8360.9671.2740.538
LUNA-Transformer1.8420.9621.7780.634
LUNA-RNN1.8480.9481.8030.607
LUNA-MLP1.8430.9531.6960.641
LUNA-O1.1500.8721.2170.553
LUNA (ours)0.9120.7040.9820.387

Q3Q3. Sensitivity to the five-state target context

In Table 9, we study the sensitity of LUNA to various state target context with 0, 3, 5, and 8 target-state examples and measure the RMSE of LUNA in the cross-domain settings. As shown in this table, the 8 target-state examples can achieve the least RMSEs for all four settings.

Table 9 in Appendix E.1: Cross-Domain RMSE versus the number of target-state examples.

Source → Target0 states3 states5 states (default)8 states
Hopf → Oscillator1.1500.9100.9120.910
Oscillator → Hopf0.8720.7000.7040.700
VL20 → Lorenz961.2171.0070.9820.962
Lorenz96 → VL200.5530.4270.3870.371

[1] Luís Marques and Dmitry Berenson. Quantifying Aleatoric and Epistemic Dynamics Uncertainty via Local Conformal Calibration. arXiv:2409.08249, 2024.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Kind Request for Review Status Update

Dear Reviewers of Submission 26155,

We greatly appreciate the time and effort you have invested in reviewing our submission and would like to confirm that we have carefully considered and addressed the feedback provided. We are writing to kindly inquire about the current status of the review process for our paper.

Please let us know if you require any further information or clarification regarding our response. We would appreciate an update on the review status at your earliest convenience.

Thank you once again for your valuable time and contributions to our research.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.