acceptodds
Under review as a conference paper at ICLR 2027

Inter-Trajectory Importance Sampling: A Principled Mitigation of Mode Collapse

Abstract

Sampling from unnormalized distributions is a fundamental challenge in computational science. Diffusion samplers typically approach it by minimizing the reverse Kullback–Leibler divergence between path measures but remain susceptible to mode collapse. Conventional single-trajectory estimators evaluate trajectories independently, without directly assessing sample diversity. We introduce Inter-Trajectory Importance Sampling (ITIS), a multi-sample estimator that couples trajectory losses through path-measure importance weighting. ITIS remains unbiased, preserves the expected gradient and critical points, requires no additional forward passes or heuristic inter-particle forces, and yields i.i.d. samples at inference. At the finite-batch level, it assigns lower realized losses to diverse trajectory configurations than to collapsed ones. We characterize theoretically how this method mitigates mode collapse and we demonstrate its efficacy experimentally. As a drop-in loss replacement, ITIS improves established diffusion samplers on standard benchmarks and interacting-particle systems.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Amortized sampling, Diffusion samplers, Mode collapse, Variational inference, Generative modeling, Probabilistic machine learning, Importance sampling

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.60 over 5 reviews, above 70% of the others.
Reviewer_yHRSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Inter-Trajectory Importance Sampling (ITIS), a multi-sample loss estimator designed to mitigate mode collapse in diffusion-based samplers for unnormalized target distributions. Rather than modifying the underlying reverse KL objective or introducing explicit particle repulsion, ITIS couples the loss contributions of independently generated trajectories through importance weights constructed from their transition kernels. The authors show that the resulting estimator preserves the expected objective and its gradient under appropriate conditions, while typically assigning smaller realized losses to batches with separated transition kernels.

Critical strengths

ITIS provides a flexible training modification that can be integrated into existing diffusion samplers without requiring additional sampling trajectories or introducing interactions at inference time. Its effectiveness is supported by improvements across DDS, DIS, and PIS on several benchmarks, suggesting practical applicability beyond a single sampling architecture.

Critical weaknesses
  1. Although ITIS employs analytically tractable importance weights, their variance can grow exponentially with the squared standardized separation between Gaussian transition kernels, potentially causing severe instability in high-dimensional settings or at small diffusion noise levels. Importantly, this is precisely the regime in which Proposition 1 establishes that the realized loss ratio approaches (1/K) in probability, meaning that the claimed preference for diverse trajectories is accompanied by increasingly rare but potentially enormous importance weights and gradient contributions. Since ITIS preserves the expected gradient of the original reverse KL objective, the manuscript does not sufficiently explain why these high-variance stochastic updates should systematically improve mode coverage rather than merely alter the optimizer's behavior. Moreover, the paper overlooks [1], which addresses a closely related problem through importance weighting and faces analogous concerns about importance-sampling variance and high-dimensional scalability, despite employing a different estimator. Without a more careful analysis of gradient variance, optimization stability, and the relationship to this prior work, the theoretical justification for the claimed principled mitigation of mode collapse remains insufficient.
  2. A potentially important issue is that the finite-batch ITIS gradient is not invariant to arbitrary multiplicative rescaling of the unnormalized target density. Specifically, replacing π~T(x)\widetilde{\pi}T(x) with eaπ~T(x)e^a\widetilde{\pi}T(x), which represents exactly the same target distribution, shifts the terminal cost by −a-a but changes the ITIS stochastic gradient by −a∇θAθ,T-a\nabla\theta A{\theta,T}, where Aθ,T=K−2∑m,kwmk,T(XT(m))A_{\theta,T}=K^{-2}\sum_{m,k}w_{mk,T}(X_T^{(m)}) is generally not constant for a realized batch. Although this additional term has zero expectation under the assumptions of Proposition 1, it can alter the stochastic-gradient variance and the resulting optimization dynamics, potentially affecting the claimed mode-coverage mechanism. This is particularly concerning because the method targets distributions with intractable normalization constants, for which the energy function is defined only up to an arbitrary additive constant. The authors should clarify how this invariance is handled in their implementation and demonstrate that the reported improvements are robust to such rescaling.

[1] Wang C, Zhang X, Cui K, et al. Importance weighted score matching for diffusion samplers with enhanced mode coverage[J]. arXiv preprint arXiv:2505.19431, 2025.

Questions for authors
  1. For Gaussian transition kernels with equal covariance s2Is^2I, the importance-weight variance satisfies Var⁡(wmk)=exp⁡(∣μm−μk∣2/s2)−1\operatorname{Var}(w_{mk})=\exp(|\mu_m-\mu_k|^2/s^2)-1. How does the proposed method control this exponential variance growth, particularly in high-dimensional settings or at diffusion steps with small noise levels?
  2. Proposition 1 establishes that the realized loss ratio approaches 1/K1/K as the transition kernels become increasingly separated, while the corresponding importance-weight variance diverges exponentially. How can the authors justify interpreting this typical loss reduction as a beneficial mode-coverage mechanism when it is accompanied by increasingly heavy-tailed stochastic estimates?
  3. Since ITIS and the conventional estimator have identical expected gradients, their different optimization outcomes must arise from differences in stochastic-gradient distributions and their interaction with the optimizer. Can the authors characterize the gradient variance or provide empirical evidence showing that the additional stochasticity systematically facilitates mode discovery rather than destabilizing training?
  4. The terminal cost is defined as CT(x)=log⁡[N(x;0,σ2I)/πT(x)]C_T(x)=\log[\mathcal N(x;0,\sigma^2I)/\pi_T(x)], and the proposed ITIS estimator multiplies this cost by the inter-trajectory importance weights. Since the target density is only available up to an unknown normalization constant, does the implementation replace πT\pi_T with its unnormalized counterpart π~T\widetilde{\pi}_T? If so, how is the omitted log⁡Z\log Z term handled after importance weighting?
  5. Consider replacing π~T(x)\widetilde{\pi}T(x) with eaπ~T(x)e^a\widetilde{\pi}T(x), which leaves the normalized target distribution unchanged. The terminal ITIS loss then changes by −aAθ,T-aA_{\theta,T}, where Aθ,T=K−2∑m,kwmk,T(XT(m))A_{\theta,T}=K^{-2}\sum_{m,k}w_{mk,T}(X_T^{(m)}), implying a change of −a∇θAθ,T-a\nabla_\theta A_{\theta,T} in the stochastic gradient. Can the authors explain whether this dependence is present in their implementation and, if not, identify the mechanism that eliminates it?
  6. Could the authors provide empirical evidence that ITIS is robust to additive shifts of the target energy function, which leave the underlying target distribution unchanged? In particular, do such shifts affect the stochastic-gradient behavior, training stability, or reported mode-coverage improvements relative to the conventional single-trajectory estimator?
2 comments

Thank you for your detailed feedback and thoughtful questions. Please find our response below

W1:
We agree that ITIS is a higher-variance estimator of the same expected loss as the single-sample estimator. However, equality of expected losses does not imply equivalent optimization dynamics. The relevant question is how the structure of the estimator’s noise interacts with the optimizer. The analogy with stochastic versus full-batch gradient descent illustrates this distinction: stochasticity frequently is the essential ingredient in non-convex optimization. We use the standard DDS optimizer, unchanged from the original implementation (Vargas et al., 2023) and the benchmark of Blessing et al. (2024), and identical for the single-sample baseline. In our response to Q3 below we show that ITIS indeed has lower SNR in the raw gradient but that this variance is significantly reduced by the standard optimizers. For ITIS, the key feature is the skewed importance-weight distribution: typical samples receive small weights, whereas rare samples receive large weights. A typical sample lies on its own kernel’s side of the other kernel’s mean, so its weight decreases as the kernels separate. Consequently, typical batches yield r:=LMulti/LSingle<1r := \mathcal{L}_{\text{Multi}}/\mathcal{L}_{\text{Single}} < 1, with gradient contributions that favor kernel separation and thus diversity. Equation 12 and the subsequent discussion identify the two contributions responsible for this behavior: the second term in the rhs of Eq. 12 is the gradient through the weights which pushes the trajectory into regions of diversity. The first terms of Eq. 12 is for typical batches just a scaled version of the single loss gradient but scaled down in weight space regions of high diversity. Table 10 tests these terms separately. Removing the gradient through the weights, so that each trajectory's cost is only reweighted by its own weight, loses the coverage gain. Shuffling these constant weights among the trajectories of a chunk, which keeps their distribution but removes their link to the trajectories' positions, performs worse still. Training with the gradient through the weights alone, without the reweighted single-sample term, diverges. The gain therefore requires the inter-trajectory term of Eq. (12) together with the reweighted single-sample gradient, and the reweighting alone, with own or permuted weights, does not produce the coverage gain.

IWSM/[1] is now discussed in Sec. 2.

Q1. We agree with the reviewer's expression for the weight variance. ITIS does not bound the variance of the weights themselves. What matters for training is how far a single batch can move the parameters, and three properties of the setup limit this. First, the weights are ratios of one-step transition kernels, not of path densities, so their variance does not accumulate over the diffusion steps. Second, we train with the standard DDS optimizer, unchanged from the original implementation which is based on Adam. App. C.8.5 shows that the updates exhibit a significantly higher SNR than the plain gradients. Third, small noise levels are not forced on the sampler, because the noise scale is tuned separately for each estimator; the advantage of ITIS persists when the number of diffusion steps is increased (App. C.8.7, Fig. 4). Explicit control of the weights is not needed in our experiments: weight clipping did not improve results on most benchmarks and the unclipped run performs best on GMM40 50D (Table 13). Regarding high dimensions, ITIS improves over the single-sample loss on GMM40 in d = 50 (Fig. 1) and d = 100 (App. C.8.3, Table 11), and on LJ-55 with d = 165 (Tab. 1).

Q2. We do not rely on the smaller loss value itself. A smaller loss alone would act like a smaller learning rate. Table 9 confirms that the gain is not a smaller effective step, since no scaling of the learning rate with the chunk size reproduces it. The gain instead requires the gradient through the importance weights. Without it, the coverage gain disappears (Table 10), as expected from our discussion of Eq. (12). That this mechanism yields genuine mode coverage is visible in the samples since ITIS places substantially more mass on the modes far from the prior, and far fewer modes remain underpopulated (Fig. 2, Fig. 3).

Q3. We characterize the gradient variance in App. C.8.5 for the runs of Fig. 1. Here g is the raw gradient and u the applied Adam update divided by the learning rate. SNR is the norm of the mean divided by the square root of the total variance over 1000–2000 batches at a checkpoint, with Adam's state held fixed for u. Each SNR is the median over 16 checkpoints per seed, reported as mean ± SE over 3 seeds; norms are medians over all batches, checkpoints and seeds.

Training lossmedian ‖g‖median ‖u‖SNR of gSNR of u
d = 2L_Single, K = 110.855.10.144 ± 0.0152.238 ± 0.015
L_Multi, K = 40030.335.70.040 ± 0.0041.47 ± 0.14
d = 50L_Single, K = 16,5601825.33 ± 0.3413.50 ± 0.29
L_Multi, K = 40037.749.20.0342 ± 0.00082.19 ± 0.24

The ITIS gradient has a lower SNR in both dimensions. The optimizer substantially narrows this gap in the applied update, although it remains larger in d = 50. Random reweighting does not produce the gain: shuffling the weights among the trajectories of a chunk, without gradient through them, keeps their distribution but removes their link to the trajectories' positions, and the coverage gain is lost (Table 10). The gain requires the gradient through the weights, whose direction depends on where the trajectories lie relative to each other (Table 10).

W2/Q4–Q6 Energy Offset. We thank the reviewer for this precise observation, which is correct. We studied it in detail and answer these points jointly. The full study is in the new App. C.8.8. Q4/Q5. Yes, we use the energy functions exactly as provided by the benchmarks, without any added constant, so the term −a ∇θAθ,T-a\,\nabla_\theta A_{\theta,T} is present in our implementation.

Aθ,TA_{\theta,T} is the total importance weight at the final step, averaged over a chunk. It lies between 1/K1/K, when the final transition kernels of the trajectories do not overlap, and 11, when they all overlap. The terminal ITIS term therefore contains the mean terminal cost as a coefficient on Aθ,TA_{\theta,T}. If this coefficient is positive, minimizing the loss pushes the final transition kernels apart, and the offset shifts the coefficient by −a-a. As the reviewer notes, the term has zero expectation. In practice, however, typical batches have Aθ,TA_{\theta,T} well below one, and the rare batches with very large weights that restore the zero mean are relatively suppressed by the optimizer (Adam). The standard remedy from RL, subtracting a running mean of the terminal cost, makes ITIS exactly invariant. However, it also sets the coefficient on Aθ,TA_{\theta,T} to zero and thereby removes the associated push-apart gradient term, and ITIS then no longer learns with the original hyperparameters of Fig. 1 at K=400 (row κ=0\kappa = 0 below).

We therefore let each terminal cost cmc_m enter the ITIS loss as cm−b+κc_m - b + \kappa. Here bb is a moving average of the mean terminal cost of previous batches, without gradient, and κ≥0\kappa \geq 0 is a fixed margin. An offset shifts every cmc_m and bb by the same amount, so the entire training run is independent of aa up to floating-point rounding. Since bb depends only on previous batches and κ\kappa is a constant, the expected gradient of Proposition 1 is unchanged. The margin κ\kappa sets the coefficient on Aθ,TA_{\theta,T} explicitly. b+κ thus replaces the implicit dependence on the unknown normalization constant with an explicit hyperparameter that does not depend on the offset.

Q6. We test this on 2D GMM40 (log⁡Z=0\log Z = 0) with K=400K = 400, batch size 2000 and the hyperparameters of Fig. 1, kept fixed without any retuning. Sinkhorn distance, mean over 3 seeds.

a = −5a = 0a = +5
plain ITIS62.5116.4295.2
b+κ, κ = 0 (plain centring)370.8370.8370.7
b+κ, κ = 1369.2370.0369.1
b+κ, κ = 5259.1261.9262.6
b+κ, κ = 10123.8119.1141.1
b+κ, κ = 1590.190.487.0
b+κ, κ = 2061.074.465.5

Fig. 5 in App. C.8.8 shows that plain ITIS is sensitive to the offset, while b+κ, with bb given by a running average, removes this dependence for every κ\kappa. The parameter κ\kappa is an additional, optional hyperparameter. We add a detailed discussion of this study to App. C.8.8.

Reviewer_3cfxOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces a reformulation of reverse KL loss for training diffusion samplers: inter-trajectory importance sampling (ITIS). The new loss is based on the observation that the naive reverse KL treats each trajectory (or subtrajectory) independently, meaning that the batch loss is just the sum of the losses from each trajectory, denoted as LSingle(Xt0:K−1)≈1K∑k=0K−1Ct(Xtk)\mathcal{L}_{\text{Single}}(X_{t}^{0:K-1}) \approx \frac{1}{K}\sum_{k=0}^{K-1} C_t(X_t^k) (tt is the time index of diffusion, kk is the sample index in a batch, and CtC_t is a cost function for time tt). The proposed method introduces dependencies among samples in a batch via importance weighting, mitigating the mode-collapse issues of reverse KL. More specifically, for each mm-th trajectory, the authors use importance sampling within a batch and rewrite the loss as L(Xt0:K−1)m≈1K∑k=0K−1pθ(Xtm∣Xt−1k)pθ(Xtm∣Xt−1m)C(Xtm)\mathcal{L}(X_{t}^{0:K-1})_m \approx \frac{1}{K} \sum_{k=0}^{K-1} \frac{p_{\theta}(X_t^{m} \mid X_{t-1}^{k})}{p_{\theta}(X_t^{m} \mid X_{t-1}^{m})} C(X_t^{m}). By averaging over all difference choices of mm, LMulti(Xt0:K−1)≈1K2∑m,k=0K−1pθ(Xtm∣Xt−1k)pθ(Xtm∣Xt−1m)C(Xtm)\mathcal{L}_{\text{Multi}}(X_{t}^{0:K-1}) \approx \frac{1}{K^2} \sum_{m,k=0}^{K-1} \frac{p_{\theta}(X_t^{m} \mid X_{t-1}^{k})}{p_{\theta}(X_t^{m} \mid X_{t-1}^{m})} C(X_t^m). The authors show that the expected gradient of LMulti\mathcal{L}_{\text{Multi}} (expectation over trajectories from pθp_{\theta}) is equivalent to LSingle\mathcal{L}_{\text{Single}} (i.e., unbiased), and also provide an intuitive explanation of why it leads to mode coverage behaviour. The method is evaluated on a standard synthetic benchmark suite and particle systems.

Critical strengths
  1. This paper is clearly written. The authors kindly explain each point in detail. I can tell the manuscript is written by humans, not LLMs.
  2. The method is surprisingly simple, versatile, and effective. Reverse KL is the standard training loss for amortised sampling and has many interesting connections with, e.g., stochastic optimal control and the Schrödinger bridge (as noted in the paper), as well as entropy-regularised reinforcement learning (see, e.g., [1]). The proposed method could potentially be extended to these adjacent areas of study. The results show consistent gains over naive reverse KL.

[1] Choi, Sanghyeok, et al. "Reinforced sequential Monte Carlo for amortised sampling." ICML 2026

Critical weaknesses

My only concern is scalability. The authors didn't provide any discussion or analysis of the variance of LMulti\mathcal{L}_{\text{Multi}}. Also, I assume they use a smaller number of steps (TT) than the original single-sample baselines, but they don't explain why. I suspect that TT is tuned only for ITIS, and it would be concerning if the method doesn't scale with TT. Adding an ablation study for TT and a brief discussion of scalability (in both TT and dimensionalities) and variance would strengthen the paper.

Questions for authors
  1. Why does ITIS use a different TT than DDS? What happens if you increase ITIS's TT to match DDS?
  2. Does the proposed method increase the variance of losses or gradients? This may be difficult to answer theoretically; if so, I’d like to know how the variance changes empirically.
4 comments

Thank you for your detailed and positive feedback on our work and for your thoughtful follow-up questions. Please find our responses below.

The additional experiment in Section C.8.7 of our revised manuscript shows that while a small number of steps TT is sufficient to achieve strong performance, increasing TT further yields improvements and eventual saturation.

Although our loss estimator can have higher variance, our experiments indicate that this does not negate its benefits for mode discovery. The key distinction is between the magnitude of the variance and its structure. As an analogy, stochastic gradients introduce noise that can facilitate exploration during optimization. For ITIS, the relevant mechanism is more specific: the importance weights have a skewed distribution, with small weights for typical samples and large weights for rare samples. A typical sample lies on its own kernel’s side of the other kernel’s mean, and its weight wmkw_{mk} decreases as the kernels separate. Consequently, on typical batches, the realized loss decreases with separation, yielding r:=LMulti/LSingle<1r := \mathcal{L}{\text{Multi}}/\mathcal{L}{\text{Single}} < 1, and the corresponding gradients favor kernel separation and diversity. In the zero-weight-variance limit, by contrast, r=1r = 1, recovering the single-sample loss, which is susceptible to mode collapse. Rare, large-weight batches counteract this tendency and cancel it in expectation, as established in Proposition 1. However, cancellation in the expected gradient need not imply cancellation in the optimizer’s updates. Adaptive optimizers such as Adam, which we use throughout, damp the influence of large gradients through their second-moment normalization, please see the new experiments in Section C.8.5 of our updated manuscript which confirm this behavior. This provides a plausible mechanism by which typical batches can have a greater influence on the optimization trajectory, favoring diversity and improved mode coverage. Finally, the additional variance does not destabilize training in our experiments.

Thank you for bringing the connection to [1] to our attention, we now discuss it in the related work of our updated manuscript.

Thank you for your response.

The paragraph "Although our loss ..." in the rebuttal is a bit hard to follow. If this matters, I would appreciate it if the authors included it with clearer wording in the paper.

I checked the additional experiments, especially C.8.5, C.8.7, and C.8.8. Please make sure to point to each of them in the main section for future readers. Also, the appendices are slightly wordy in my opinion, so please improve readability.

Also, a discussion of high variance (and why an Adam-like optimiser is necessary) should be included in the main section as well. It might be helpful to discuss when ITIS with Adam can be much worse because of the higher variance, if you know such cases.

I still think the paper is interesting. I'd like to maintain my positive rating.

Thank you for your continued engagement and positive assessment.

The paragraph "Although our loss estimator ..." makes three points. First, a higher-variance gradient estimator does not imply worse learning, as the example of SGD shows. Second, the gradients of typical ITIS batches reduce the overlap of the transition kernels and thereby favor mode coverage. Third, under standard optimizers such as Adam these typical gradients dominate the learning dynamics, which is consistent with our experiments (App. C.8.5, Table 10).

We will include these considerations with clearer wording in App. C.8.5, point to App. C.8.5, C.8.7 and C.8.8 from the main text, add a discussion of the variance to the main text, and improve the readability of the appendices.

We have now implemented these changes in the updated manuscript. Sec. 4.2 contains a new paragraph on the variance of ITIS and the role of the optimizer, which points to App. C.8.5 and notes that ITIS also trains successfully with SGD with momentum (Table 10). The discussion of Fig. 1 now points to App. C.8.5 and C.8.7, the limitations point to App. C.8.8, and App. C.8.5 opens with a clearer version of the paragraph on variance. We will further streamline the appendices for the final version. Thank you again for these suggestions, which improved the paper.

Reviewer_8wKYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents Inter-Trajectory Importance Sampling (ITIS): an unbiased multi-sample estimator of the rKL path objective that mitigates mode collapse in diffusion samplers. The key insight is to use cross-trajectory importance weights derived from the transition kernels of other trajectories to assign lower loss to diverse configurations than to collapsed ones. Experiments across a variety of benchmarks demonstrate consistent gains over single trajectory training while requiring minimal implementation changes.

Critical strengths

The approach is conceptually intuitive with a simple yet novel idea that seems natural in hindsight. It is also easy to incorporate into existing diffusion samplers without architectural changes or additional sampling. The authors give strong theoretical justification by characterizing ITIS relative to the standard single-trajectory loss, establishing that both estimators have the same expected gradient and that their loss ratio approaches K−1K^-1 as the overlap between the transition kernels of different trajectories vanishes.

Critical weaknesses

The expected gradient equality for LSingle\mathcal L_{\text{Single}} and LMulti\mathcal L_\text{{Multi}} is well motivated, but the paper does not characterize gradient’s variance behavior. Paper lacks comparisons with recent approaches specifically designed to improve mode coverage [a, b]. There’s limited ablations provided, particularly to: isolate contribution of the two gradient terms Eqn 12, sensitivity to optimizer choice, gradient clipping, chunk partitioning etc. While no additional sampling is required, given that computing pair-wise importance weights is O(K2)O(K^2), the increased runtime requirements warrant systematic testing.

[a] He, et al. “Training Neural Samplers with Reverse Diffusive KL Divergence.” AISTATS 2025. https://arxiv.org/abs/2410.12456 [b] Wang, et al. “Importance Weighted Score Matching for Diffusion Samplers with Enhanced Mode Coverage.” https://arxiv.org/pdf/2505.19431

Minor notes: πT\pi_T is referred to as unnormalized in section 2, though it is defined as normalized. Eqn 1 defined the distribution, not sampling as is denoted by the surrounding text. SDE should be introduced before acronym is used other typographical errors and notational errors.Overall, clarity would benefit from a final editing pass

Questions for authors

Please see weaknesses.

2 comments

Thank you for your constructive review, please find our response below.

Although our loss estimator can exhibit higher variance, our experiments suggest that this does not undermine its benefits for mode discovery. The relevant distinction is between the magnitude of the variance and its structure. Just as stochastic gradient noise can facilitate exploration during optimization, variability in our estimator may influence the optimization trajectory. For ITIS, we propose a more specific mechanism arising from the skewed distribution of importance weights: typical samples receive small weights, whereas rare samples receive large weights.

A typical sample lies on the same side of the other kernel’s mean as its own kernel’s mean, and its importance weight, wmkw_{mk}, decreases as the kernels separate. On typical batches, the realized loss therefore decreases with separation, yielding r:=LMulti/LSingle<1r := \mathcal{L}_{\text{Multi}}/\mathcal{L}_{\text{Single}} < 1, with corresponding gradients that favor kernel separation and diversity. In the limit of zero importance-weight variance, by contrast, r=1r = 1, recovering the single-sample loss, which is susceptible to mode collapse. Rare batches with large importance weights counterbalance this tendency in expectation, as established in Proposition 1. However, cancellation in the expected gradient does not necessarily imply cancellation in the optimizer’s updates. Adam, which we use throughout, rescales gradients and can therefore dampen the influence of large gradients which is confirmed by our new experiments in Section C.8.5 of the updated version of our manuscript. This offers a plausible mechanism through which the separation-favoring gradients of typical batches can exert greater influence on the optimization trajectory, promoting diversity and mode coverage. Finally, we observe no training instability in our experiments.

Thank you for these constructive suggestions. In the revised manuscript, we have added experiments examining the contributions of the two gradient terms in Eq. 12 (Section C.8.2) which confirm that the full ITIS gradient is needed to mitigate mode collapse. Our new ablation in Section C.8.2 shows, when switching from Adam to SGD, our approach remains robust with respect to optimizer choice. The new weight clipping ablation in Section C.8.6 confirms that ITIS does not depend on a specific weight clipping parameter. The original manuscript already addresses chunk partitioning and computational requirements in Figure 1 and Section C.1 where it can be seen that the runtime in our experimental setups is mainly driven by the number of diffusion steps TT instead of KK and ITIS requires significantly less diffusion steps than single-sample losses. A crossover study is also included in Table 4. Thank you also for highlighting [a, b]. We selected our experimental baselines primarily because they share most of the benchmarks in our evaluation pipeline, facilitating consistent comparisons. We recognize that [a, b] are relevant to the question of mode coverage and have added a discussion comparing their methodology with ours.

Finally, we have corrected the normalization terminology in Section 2, introduced “stochastic differential equation” before using “SDE,” and corrected other typographical and notational issues.

Dear Reviewer 8wKY, as the discussion closes soon, we briefly summarize how the revision addresses your points.

The gradient variance is characterized in App. C.8.5 (Table 12). The two gradient terms of Eq. (12) are isolated in Table 10, which also shows that ITIS works with SGD with momentum instead of Adam, and weight clipping is examined in Table 13.

The works of He et al. (2025a) and Wang et al. (2025) are now discussed in Sec. 2, and the runtime is reported in App. C.1.

If you feel these updates address your concerns, we would be grateful if you could take them into account in your final assessment, and we would of course be happy to clarify anything that remains open.

Reviewer_DkAmOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Inter-Trajectory Importance Sampling (ITIS), a multi-sample estimator for the reverse KL divergence that aims to mitigate mode collapse in diffusion samplers. ITIS uses importance weighting across trajectories and typically assigns lower realized losses to diverse trajectory configurations than to collapsed ones, while preserving the expected objective and gradients. Experiments on standard sampling benchmarks and particle systems demonstrate improvements over conventional single-trajectory loss estimators.

Critical strengths

A novel and simple modification to the training signal for mitigating mode collapse. ITIS introduces inter-trajectory importance weighting without modifying the underlying diffusion sampler or requiring additional forward passes. The method can be readily applied to existing diffusion samplers.

Critical weaknesses

1. Unclear scalability of the proposed mode-collapse mitigation mechanism in high dimensions.

ITIS relies on importance sampling across trajectories, but importance sampling is known to suffer from weight degeneracy in high-dimensional spaces. In particular, even small differences between transition kernels in individual dimensions can lead to nearly zero cross-trajectory importance weights as dimensionality increases.

As a result, the proposed mechanism may provide little useful signal for distinguishing mode-collapsed configurations from mode-covering ones in high-dimensional systems. It is unclear whether the proposed mechanism remains effective in such settings, and further analysis of the importance weights and their behavior with increasing dimensionality would be valuable.

2. Limited empirical evidence for mode-collapse mitigation and competitiveness against strong baselines.

While the paper reports quantitative improvements in sampling performance, it does not provide qualitative visualizations of the generated samples. For example, sample plots for the 2D GMM40 benchmark and projected samples for the 50D GMM40 benchmark would help illustrate whether ITIS actually mitigates mode collapse.

On the high-dimensional LJ-55 benchmark, DDS with ITIS is outperformed by AS, ASBS, and PDNS. Although ITIS improves over the original DDS, the relatively weaker performance compared to these baselines raises questions about its effectiveness on more challenging systems. Further discussion or analysis of this performance gap would be helpful.

Some recent relevant baselines, such as Bridge Matching Sampler (BMS) and Flow Sampling (ICML 2026), are missing from the experimental comparison. Including these methods would help provide a more comprehensive assessment of the proposed approach.

Questions for authors
  1. Could the authors provide qualitative results demonstrating the mitigation of mode collapse? In particular, sample plots for the 2D GMM40 benchmark and projected samples or per-mode occupancy plots for the 50D GMM40 benchmark would help illustrate the improvements over the single-trajectory baseline.

  2. Could the authors include comparisons with recent methods, such as Bridge Matching Sampler (BMS) and Flow Sampling (ICML 2026), on shared benchmarks?

  3. Could the authors report the total number of target energy evaluations and the overall training cost? In particular, can DDS with ITIS achieve better sampling performance than DDS without ITIS under matched target energy evaluation budgets and training compute budgets?

  4. Can ITIS be extended to other stochastic optimal control-based diffusion samplers, such as Adjoint Sampling (AS) and Adjoint Schrödinger Bridge Sampler (ASBS), which use Adjoint Matching-based objectives?

6 comments

Thank you for your detailed feedback and thoughtful questions. Please find our response below.

Scalability:
We thank the reviewer for raising the question of dimensional scalability. While importance-weight degeneracy is a potential concern, dimensionality alone does not establish that the proposed mechanism loses its effectiveness. The relevant question is whether discrepancies between the trajectory distributions are sufficiently large to make cross-trajectory comparisons uninformative in the settings considered. The overlap of transition kernels can be controlled and scaled with dimensionality by the noise scale σt\sigma_t hyperparameter of the transition kernels which enables to mitigate possibly vanishing importance weights in higher dimensions. Our experiments already investigate this question empirically in 39, 50, and 165 dimensions. These settings allow the accumulation of transition-kernel discrepancies described by the reviewer, yet we do not observe the corresponding deterioration in the effectiveness of our mechanism. Thus, our evidence extends beyond low-dimensional illustrations and supports the practical utility of the proposed approach across the dimensionalities studied. Our experimental design follows the paper’s central contribution: introducing and investigating an algorithmic principle that addresses mode collapse in diffusion samplers. We deliberately use synthetic benchmarks because they enable controlled comparisons and make mode coverage and failure modes directly interpretable. Extending the evaluation to a particular high-dimensional application, such as molecular systems, would require substantial additional methodological and implementation choices, making it harder to isolate the behavior of the proposed mechanism. Such an extension would be valuable, but it constitutes a separate application study rather than a prerequisite for establishing the contribution presented here. Furthermore, the LJ-55 dataset has 165 dimensions, exceeding the dimensionalities of the standard molecular sampling benchmarks alanine dipeptide and alanine tripeptide, which have 66 and 96 dimensions, respectively. We also added a new experiment that confirms the scalability of our mechanism in Table 11 of the updated version of our manuscript.

Evidence for mode collapse mitigation:
We thank the reviewer for this helpful suggestion. The mode discovery and coverage experiments in Figure 1 evaluate mode collapse directly on the 2D and 50D GMM40 benchmarks, providing quantitative evidence that ITIS mitigates this behavior in a controlled setting. We prioritized these metrics because they assess mode coverage explicitly, including in high dimensions where low-dimensional projections may obscure missing modes. Nevertheless, we agree that qualitative visualizations would complement these results. We therefore added sample plots for the 2D benchmark and mode occupancy plots for the 50D benchmark alongside the quantitative evaluation in Section C.8.4 of our updated manuscript which confirm the mode collapse mitigation effect of our method. Both the sample plots and the mode occupancy measurements in Figure 2 show that the single-sample loss misses modes farther from the prior, whereas our multi-sample approach maintains substantially higher occupancy of these modes. A similar pattern emerges in the 50-dimensional case: unlike the single-sample loss, ITIS nearly matches the mode occupancy when sampling from the target distribution, as shown in Figure 3.

Competitiveness against baselines:
We thank the reviewer for the comments. The LJ-55 results should be interpreted in light of the contribution being evaluated: ITIS is a training objective, and our central claim concerns its improvement over the corresponding single-loss baseline. DDS with ITIS improves over the original DDS on LJ-55, supporting this claim even on the high-dimensional benchmark. Although AS, ASBS, and PDNS achieve better results on this system, this comparison evaluates complete sampling methods and does not isolate the effectiveness of the training objective. Following the reviewer’s suggestion, we also assess computational cost using NFEs, which further supports the competitiveness of ITIS as it requires only 80 NFEs, compared with 1,000 for PDNS. Furthermore, DDS with ITIS outperforms these more sophisticated samplers on other particle-based benchmarks. Taken together, these results demonstrate the benefit of ITIS within the evaluated setting without implying that DDS equipped with ITIS will outperform every advanced sampler on every benchmark. Whether incorporating ITIS into more advanced adjoint-based diffusion samplers yields similar improvements remains an open research question. Although we anticipate no immediate obstacles to such an extension, establishing its effectiveness would require a systematic evaluation beyond the scope of the present work.

We appreciate the suggestion to include BMS and Flow Sampling and discuss these relevant methods in the revised manuscript. Comparisons with them would broaden the assessment of overall sampling performance, while the matched comparison against the single-loss baseline remains the direct test of our proposed objective.

NFEs and computational cost:
Given that the PISGradNet architecture takes the gradient of the energy function as an input, we require one NFE per diffusion step. Together with the terminal cost this adds up to T+1T+1 NFEs.The NFEs for our multi-loss objective can be deduced from Section C.7 in the Appendix by adding 1 to the number of diffusion steps. Following the evaluation protocol of Blessing et al. (2024) results in optimized T=128T=128 for each method using the baseline single-loss objective. As the computational cost is mostly driven by NFEs as can be seen in Section C.1, DDS with ITIS achieves better sampling performance than DDS without ITIS under matched compute budgets.

Thank you for the added visualizations. I acknowledge the improvements over DDS, including on 165D LJ-55.

My remaining request is a quantitative analysis of cross-trajectory importance-weight ratios within the same GMM40 family as the dimension increases to 100 and beyond—for example, their distributions and the fraction of near-zero ratios. This would clarify how the cross-trajectory importance-weight ratios change with dimension without requiring a new application. Limited scalability itself would not be a problem if its behavior and limitations are clearly characterized.

Reviewer_qJ1gOpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
2(+1)
Presentation
2
Soundness
1
Read review
Paper summary

The paper provides an unbiased estimate of the gradient of the reverse KL loss using interactions between multiple trajectories. It relies on an importance sampling identity between the conditional one-step transition kernels of the learned dynamic, creating inter-trajectory terms in the gradient estimate. Although both target the same loss, the authors give heuristic intuitions on why the estimate using their importance sampling technique favours mode discovery. Experiments seem to show that larger chunk sizes (i.e. more interaction between trajectories) lead to better mode covering.

Critical strengths
  1. Equation (9) in Proposition 1 is correct.
Critical weaknesses

The proposed estimate has horrible variance, which is admitted by the authors themselves ("this push is balanced in expectation by rare batches with large weights"). It targets exactly the same reverse KL loss and has exactly the same minimizer, so it's difficult to see in principle how this leads to better mode discovery.

In practice, as the authors again admitted, "many pairs of chunk carry a small weight". I suspect what happens is that the magnitude of their new update is more or less 1/K1/K times the magnitude of the old update, so increasing KK simply corresponds to a lower learning rate, which might explain the better results in the experiment.

The current experimental protocol to produce Figure 1 obscures this point, because the authors "tune the learning rate and noise scale separately for each KK". Is the noise scale the σt\sigma_t in the beginning of the Section 3? I think that if the authors use exactly the same setting for small KK and large KK (including schedules etc), but with the learning rate being proportional to KK, then the improvements in Figure 1 would disappear. In other words, good results for large K that are currently being displayed in Figure 1 can be reproduced for smaller KK simply by using a learning rate that is KK times smaller.

Questions for authors

My most critical question is in the Weaknesses section. Here I list minor suggestions.

  1. Near Eq (10), the convergence in probability "as the overlap of pkp_k and pmp_m vanishes" is unclear. Better to write "as δ→∞\delta \to \infty" where δ\delta (defined in Appendix A.2) quantifies the separation of pkp_k and pmp_m.

  2. In Section 3, the loss notation CC actually depends on tt, this should be stated clearly.

UPDATE. I have decided to recommend an acceptance of this paper. While I am still not convinced theoretically, the experimental results do show improvements and I think it is beneficial to accept this paper so that the community can investigate the basis of this improvement.

6 comments

We thank the reviewer for the careful reading. We address the main points below.

Summary: This review primarily rests on a clearly stated hypothesis, namely that ITIS acts effectively as a K times smaller learning rate. We ran precisely the experiment the reviewer specified, with identical settings for all K and the learning rate scaled with K, and small chunk sizes still do not reach the large-K result (first table below / Table 9 in new App.). Our ablations further show that the gains largely disappear once the gradient through the importance weights is removed while the reweighting of the loss is kept, which no learning-rate change can mimic (Table 10). The remaining objection, that an estimator with the same expectation and the same minimizer cannot change the training outcome, would equally deny any effect of minibatch noise or of potential-based reward shaping in reinforcement learning, which leaves the optimal policy unchanged yet changes how quickly it is found (Ng et al., 1999). We therefore believe that this argument does not hold in machine learning in general, and in particular not for ITIS, whose empirical benefits over the single-sample loss are consistent.

  1. Structured variance of ITIS. ITIS indeed is a higher-variance estimator (see Tab. 12 in updated Appendix) of the same loss than the standard single loss. This is a similar trade-off as the one between stochastic and full-batch gradient descent, i.e. the minibatch gradient is an unbiased but noisier estimate of the full gradient, yet SGD typically reaches better minima than GD because its noise has a suitable variance structure and lets it explore rather than exploit in the nearest basin. ITIS adds variance in structured way, i.e. via a skewed distribution, i.e. typical samples receive small weights and rare samples large ones. A typical sample lies on its own kernel's side of the other kernel's mean, so its weight wmkw_{mk} decreases as the kernels separate. Thus on typical batches the realized loss shrinks with separation r:=LMulti/LSingle<1r := \mathcal{L}_{\text{Multi}}/\mathcal{L}_{\text{Single}} < 1, and their gradients push the kernels apart, increasing sample diversity. With zero weight variance we would always have r=1r = 1, the single-sample loss which is prone to mode collapse. The rare large-weight batches cancel this push in expectation (Prop. 1), but their effect on the actual optimization trajectory is systematically suppressed. Adam keeps the step induced by any single batch on the order of the learning rate, independently of the gradient magnitude (Kingma & Ba, 2014), and in the low signal-to-noise regime of ITIS (Tab. 12) it behaves like sign descent, which follows the majority sign of the per-batch gradients rather than their average (Balles & Hennig, 2018, and Bernstein et al., 2018). More generally, an importance-sampling average needs a number of samples that grows exponentially with the KL-divergence between the kernels before it reaches its mean (Chatterjee & Diaconis, 2018), so at any nonvanishing learning rate the parameters move on before the rare restoring batches accumulate. The net effect is a push towards diversity, which results in mode coverage. In practice this variance does not destabilize training and e.g. our mode-discovery runs (Fig. 1) do not even require weight clipping. Our results with the stop gradient on the weights, which are shown in ablation (A) in the second table below (Tab. 10 in new Appendix) clearly show that the benefit of ITIS depends on the gradient contribution of weights (second term Eq 12, ff.) which encourages the diversity as discussed above.

  2. Identical minimizers do not imply identical training outcomes. Here we would like to address the reviewer’s statement: “It targets exactly the same reverse KL loss and has exactly the same minimizer, so it's difficult to see in principle how this leads to better mode discovery.” The single-sample estimator also has the target as its global minimizer, and it still collapses. The central motivation of our work is that mode collapse of rKL-based losses is not a problem of the minimizers of the loss but of the basin optimization ends in, i.e. not of where the global minimum lies. Consequently, our method is constructed to affect the optimization dynamics but not the minimizers of the loss which have to remain the target. With finite batches and a nonlinear optimizer, the expected update differs from the update computed from the expected gradient, so estimators with identical expectation do in general lead to different solutions. Numerous well-known methods are based on this principle, e.g. in ML for variational inference (Roeder et al., 2017; Tucker et al., 2019) and for the implicit regularization of SGD noise (Mandt et al., 2017; Smith et al., 2021).

  1. ITIS does not act as a KK times smaller learning rate.

The reviewer conjectures the following: “good results for large KK that are currently being displayed in Figure 1 can be reproduced for smaller KK simply by using a learning rate that is KK times smaller.”

We do not share this expectation since under Adam, which we use throughout, multiplying the loss by a constant leaves the updates unchanged (up to the stabilizer ϵ\epsilon). A 1/K1/K-scaled loss therefore does not correspond to a KK times smaller learning rate (see new Tab. 12). This observation does of course not rule out the reviewer’s hypothesized mechanism behind ITIS according to which the effect is a result of a reduced effective learning rate. To address this question empirically we ran dedicated experiments in the settings of Fig. 1, i.e. on GMM40 in 2D and 50D. All training runs use the configuration tuned for K=400K=400 (noise scale, learning rate η400\eta_{400}, batch size, constant schedule, 10410^4 iterations, 5 seeds) and only KK and the learning rate change as follows. Arm A probes if scaling the learning rate of K=400 to the other K values according to the 1/K scaling hypothesis reproduces the benefits of ITIS. This corresponds to using the following learning rates ηK=η400K/400\eta_K=\eta_{400} K/400. Arm B keeps ηK=η400\eta_K = \eta_{400} and only varies KK. This tests the scenario that KK has actually no impact, e.g. due to Adam’s normalization, and that only η400\eta_{400} is responsible for the strong performances since it might happen to be the generally better learning rate.

Kd=2, arm Ad=2, arm Bd=50, arm Ad=50, arm B
1371.0 ± 1.1492.2 ± 2.714218 ± 49093 ± 124
5369.8 ± 1.1164.6 ± 8.513176 ± 538283 ± 133
50259.5 ± 8.4114.0 ± 5.53295 ± 2263983 ± 224
200131.9 ± 16.0104.8 ± 14.93141 ± 523141 ± 36
400 (Fig. 1)114.6 ± 17.0114.6 ± 17.03139 ± 1073139 ± 107

We find that neither the results for Arm A nor Arm B indicate that the benefits of ITIS stem from the learning rate but indicate that it is the number of interacting particles KK.

Consequently, we now directly investigate the effect of these couplings in the ITIS loss. By Eq. (12), ∇θLMulti=r,∇θLSingle+LSingle,∇θr\nabla_\theta \mathcal{L}_{\mathrm{Multi}} = r\\,\nabla_\theta \mathcal{L}_{\mathrm{Single}} + \mathcal{L}_{\mathrm{Single}}\\,\nabla_\theta r. Any reduction of the update magnitude, as described in this review, comes from the reweighting in the first term. The second term, the gradient through the importance weights, couples the trajectories. We trained with two modifications of ITIS (K=400K=400):

(A) No gradient through the weights: the wmkw_{mk} are treated as constants. This keeps the reweighting of every trajectory's cost, and with it any modification of the update magnitude, but removes the inter-trajectory term. (B) Permuted weights: at every diffusion step, the total weight ∑kwmk\sum_k w_{mk} of each trajectory is replaced, without gradient, by that of a randomly chosen other trajectory of the same chunk. The weights keep their distribution, and hence the distribution of the update scaling, but lose their link to where a trajectory lies relative to the others.

All runs use 2D GMM40 (DDS), the same initialisation, the noise scale and learning rate tuned for K=400K=400, and 5 seeds; only the loss differs.

Loss used for trainingSD (↓)EMC (↑)
LSingle\mathcal{L}_{\mathrm{Single}}491.6 ± 2.60.629 ± 0.003
LMulti\mathcal{L}_{\mathrm{Multi}} (K=400K=400)144.8 ± 31.90.925 ± 0.006
(A) LMulti\mathcal{L}_{\mathrm{Multi}}, no gradient via wmkw_{mk}367.0 ± 3.10.725 ± 0.002
(B) LMulti\mathcal{L}_{\mathrm{Multi}}, permuted wmkw_{mk}409.2 ± 4.20.701 ± 0.003

Only the full ITIS loss, with the gradient through the importance weights, gains coverage (SD 144.8). ITIS runs show a larger seed-to-seed spread (± 31.9). Each of the five ITIS seeds nevertheless ends below the best seed of every other setting.

Together, these results essentially rule out an effectively smaller learning rate as the explanation of the gains in Fig. 1. The benefit of ITIS requires the gradient through the importance weights, i.e. the inter-trajectory term of Eq. (12). This term depends on where the trajectories lie relative to each other, and no rescaling of the loss or of the learning rate could reproduce it. We added these results to the new section Appendix C.8.

  1. Noise Scale The noise scale is indeed the σt\sigma_t introduced in the beginning of section 3. We clarified this now explicitly.

  2. Minor points. We adopted the proposed formulation concerning the overlap using δ\delta and we clarified the time dependence of cost C explicitly in the updated version.

We are thankful to the reviewer for their well-motivated initial objections. We are convinced that the presented clarification on how ITIS operates did improve our manuscript significantly.

References

Balles, L. and Hennig, P. (2018). Dissecting Adam: The Sign, Magnitude and Variance of Stochastic Gradients. ICML 2018. https://arxiv.org/abs/1705.07774

Bernstein, J., Wang, Y.-X., Azizzadenesheli, K. and Anandkumar, A. (2018). signSGD: Compressed Optimisation for Non-Convex Problems. ICML 2018. https://arxiv.org/abs/1802.04434

Chatterjee, S. and Diaconis, P. (2018). The Sample Size Required in Importance Sampling. Annals of Applied Probability 28(2), 1099-1135. https://arxiv.org/abs/1511.01437

Kingma, D. P. and Ba, J. (2014). Adam: A Method for Stochastic Optimization. ICLR 2015. https://arxiv.org/abs/1412.6980

Mandt, S., Hoffman, M. D. and Blei, D. M. (2017). Stochastic Gradient Descent as Approximate Bayesian Inference. JMLR 18(134), 1-35. https://arxiv.org/abs/1704.04289

Ng, A. Y., Harada, D. and Russell, S. (1999). Policy Invariance under Reward Transformations: Theory and Application to Reward Shaping. ICML 1999, 278-287.

Roeder, G., Wu, Y. and Duvenaud, D. (2017). Sticking the Landing: Simple, Lower-Variance Gradient Estimators for Variational Inference. NeurIPS 2017. https://arxiv.org/abs/1703.09194

Smith, S. L., Dherin, B., Barrett, D. G. T. and De, S. (2021). On the Origin of Implicit Regularization in Stochastic Gradient Descent. ICLR 2021. https://arxiv.org/abs/2101.12176

Tucker, G., Lawson, D., Gu, S. and Maddison, C. J. (2019). Doubly Reparameterized Gradient Estimators for Monte Carlo Objectives. ICLR 2019. https://arxiv.org/abs/1810.04152

Comments on the paper

3 comments
AuthorsOpenReview ↗

General response

We thank the reviewers for their careful evaluation of our manuscript and the constructive feedback. We appreciate their positive assessments of the manuscript’s clarity and the conceptual simplicity, applicability, and effectiveness of our approach. We have addressed all reviewers’ questions and concerns in our individual responses and in Section C.8 of the revised manuscript. These primarily concerned the variance of the proposed loss and requests for additional ablation studies. The variance of the ITIS estimator of the ELBO is indeed larger than for the conventional single trajectory ELBO estimator. However, we would like to stress that ITIS is not proposed as a more efficient estimator of the ELBO but as a better training objective for neural samplers. Similarly, it is important to recognize that two losses with the same minimizers will in general yield different training outcomes. Here we would like to invoke the comparison between SGD (high variance and better) vs GD (low variance and worse) or of other rKL based losses in the field of neural samplers like the log-variance loss. Nevertheless, we agree that a thorough investigation of the role of variance in ITIS and its practical impact was warranted. Accordingly we added these aspects in the updated appendix (e.g. in App. C.8.5) and show that the variance of the gradient is indeed larger for ITIS but we maintain that ITIS yields better samplers regardless.

Reviewer qJ1g made an interesting conjecture that the success of ITIS is a result of the lower effective learning rate and proposed a corresponding experiment to check if this is the case. We performed this experiment and subsequently extended the experimental setup to close further similar potential loopholes that could arise under a flawed hypereparametrization. The results contradict the reviewer's proposed explanation. In addition, we now provided detailed studies on the gradient level to unambiguously show that our results stem from the mechanisms that are discussed in the paper. In particular, ablations at the level of training show that the gain requires the gradient through the importance weights, and that weights with the same distribution but no link to the trajectories' positions do not produce it (Table 10).

We believe that the extended analysis in the new App. C.8, which isolates the mechanism of ITIS at the level of both gradients and full training runs, substantially strengthens the case for its soundness and effectiveness.

We thank the reviewers again for their careful assessment and constructive feedback. We have addressed each concern in our point-by-point responses and revised the manuscript accordingly. We believe these changes resolve the issues underlying the initial assessments and strengthen the clarity and rigor of the paper. We respectfully ask reviewers who initially recommended rejection to reassess whether their concerns have been resolved and, if so, to consider updating their scores and recommendations to reflect the revised manuscript and our responses. If any substantive concern remains, we would appreciate its identification so that we can address it directly. Thank you again for your time and thoughtful consideration.

Area_Chair_tDooOpenReview ↗

Author-Reviewer Discussion ends in 1 day, 12 hours

Dear authors and reviewers,

Thank you to the authors for their detailed responses, the new experiments and the updated manuscript. The discussion period closes in about a day and a half, so this is the last chance for a direct exchange before the reviewer-AC discussion starts on Nov 19.

Reviewers qJ1g, 8wKY and 3cfx: I haven’t yet seen a reply from you. Please read the authors’ responses and the revision, which address several of your points directly. Let us know whether your concerns have been addressed, and please update your recommendation if your assessment has changed. Even a short acknowledgment helps.

Reviewers DkAm and yHRS: thank you for engaging. The authors have replied to your latest comments (Nov 17). Please confirm whether these replies settle your remaining points.

Authors: please answer any new questions in the time left, focusing on the points most likely to affect the decision.

Thank you all for keeping the discussion constructive and timely. Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.