acceptodds
Under review as a conference paper at ICLR 2027

Variational Autoreasoners: Recursive Reasoning by Amortizing the Other Half of EM

Abstract

Recursive reasoning models share a set of design choices-a persistent state, a shared transformation, a prediction at every iteration, and local supervision with truncated gradients-whose computational role remains unclear. We interpret them through variational EM. Existing amortized inference amortizes only the E-step: it refines a posterior that conditions on the target, with no way to carry the result forward or to continue without the target. We introduce an instance-level M-step that adapts a per-problem memory conditioning the prior while shared weights stay fixed; the resulting inference-adaptation process inherits the monotone likelihood ascent of EM and anneals the prior toward the best explanations of the target. Variational Autoreasoners amortize this process with local variational objectives and run without the target through a recursive reasoning prior. The divergence between posterior-driven and prior-driven execution decomposes into the KL terms of the local objectives, and each design choice of recursive reasoners acquires a computational role.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Recursive Reasoning, Amortized Inference, Latent-Variable Models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_h8FMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the relationship between persistent state, shared transformations, stepwise supervision, and gradient truncation in recursive reasoning models. It proposes an interpretation based on variational EM. Memory for each problem instance controls the latent prior. The E-step infers latent explanations from the input, target, and current memory. The instance-level M-step then updates memory while keeping the shared model parameters fixed, bringing the next prior closer to the current posterior. Based on this formulation, the authors introduce Variational Autoreasoners, which use a posterior network and a memory adapter for recursive computation. Training uses local variational objectives. The objective at the next step trains the preceding memory update, while gradients through earlier steps are truncated. The experiments evaluate depth scaling, solution coverage, memory updates, and training GPU memory use on graph coloring, N-Queens, complete Sudoku grid generation, and a Gaussian mixture toy task.

Critical strengths
  1. Separating shared parameters, instance memory, and the current latent variable clarifies how one step affects later generation. Memory enters the model only through the prior, allowing the ideal memory update to be written as a KL projection that fits the next prior to the previous posterior. The paper also explains how the next objective trains the memory adapter and how gradient truncation limits long-term credit assignment.

  2. The paper increases test depth while keeping the trained parameters fixed, separating the effect of additional computation from that of additional training. On NQ10, Table 9 shows that both single-answer accuracy and solution coverage over 20 independent draws improve when the same checkpoints are run beyond the training depth. These results support the benefit of additional iterations in the learned loop.

  3. Both removing memory updates and replacing the sampled latent input to the memory updater with a learned constant token substantially reduce success on GC8. These ablations support the usefulness of memory updates that depend on the sampled latent content. Table 10 also shows that, with the backbone, hardware, and batch size fixed, VAR’s peak GPU memory use remains at about 2.88 GiB across the tested training horizons. Memory use increases with the horizon in its full-BPTT controls, which do not use gradient checkpointing.

Critical weaknesses
  1. The paper describes GRAM’s objective as a truncated surrogate of a trajectory ELBO and VAR’s objective as a complete local ELBO. However, GRAM also uses a posterior conditioned on the target during training, prior sampling at test time, a recurrent state, and gradient truncation. Conditional on the current deterministic state, GRAM’s Equation (14) can also be interpreted as a local ELBO for a single-step model. The distinction between these descriptions does not by itself establish the algorithmic contribution. A clearer comparison of the actual objectives, state updates, and gradient paths is needed.

  2. The ideal M-step holds the previous posterior fixed and updates memory to fit it. In the method, the adapter receives a single sample, the posterior is recomputed under the new memory, and the next objective trains the adapter. This objective does not directly require the new prior to match the previous posterior. It acknowledges this difference and relates the next objective to the target likelihood under exact inference. However, the paper provides limited direct evidence that the learned updates approximate the ideal M-step or follow the power reweighting process. Whether this EM interpretation explains the observed gains from greater depth therefore remains unclear.

  3. The Sudoku task is to generate a complete grid from a blank input, as stated in Appendix E.1. Repeatedly outputting one valid training grid would achieve 100% validity. Near-perfect validity therefore does not by itself show that the model generates diverse solutions or unseen grids. For Sudoku, the paper does not report duplicate rates, the number of unique valid grids, overlap with the training set, or diversity after accounting for grid symmetries. These missing measurements limit what can be concluded about the diversity and generalization of the generated solutions.

Questions for authors
  1. What is the specific algorithmic contribution beyond GRAM? With one stochastic transition per supervision interval and the same supervision frequency, how do the objectives, state updates, and gradient paths differ? Could you compare the two methods using the same backbone and optimization settings and evaluate the contribution of each difference separately?

  2. How closely do the learned memory updates approximate the ideal instance-level M-step? In a controlled setting where the exact posterior and ideal memory update can be computed, could you compare the learned updates with the ideal ones and examine how the differences relate to the gains from increasing test depth?

  3. For Sudoku, could you report the number of distinct valid grids and the fraction of valid outputs that appear in the training set, using the same number of samples at each execution depth? As depth increases, do the valid outputs remain diverse, or do they become concentrated on a small set of solutions?

5 comments

We thank the reviewer for the constructive feedback and thoughtful questions. We address each concern below, including the comparison with GRAM, the relationship to the idealized M-step, and Sudoku diversity.

[W1, Q1] Algorithmic contribution beyond GRAM when GRAM's Eq. 14 can also be read as a local ELBO

Our main contribution is to explain recursive reasoning as variational EM, a view that GRAM does not provide. Under this view, each recursion step can be read as one step of instance-level EM that updates the prior for the next step, which explains how recursion turns additional steps into better predictions. Prior recursive reasoning models, including GRAM, show that this recursion is effective but leave open why it works (lines 33–37).

This view also resolves a critical mismatch in GRAM between its objective and its per-step parameter updates. Like HRM and TRM, GRAM updates its parameters at every recursion step. Its objective, however, is a single ELBO over the whole trajectory, which assumes no update until the trajectory ends, so GRAM treats its per-step updates as a biased approximation. In our framework, each step has its own complete objective, the local ELBO, so per-step updates follow directly from the objective. This explains why Eq. 14, which the reviewer reads as a local ELBO, is a valid objective for per-step updates. GRAM's trajectory ELBO does not account for this.

GRAM's architecture can be seen as one implementation of our framework. In our framework, the parameters at each step are trained with the local ELBO of the model p(y∣x,ht)p(y\mid x,h_t) conditioned on the memory hth_t, which is a complete objective at that step (Algorithm 1). GRAM implements this core with additional structures, such as the H/L hierarchy, the inner low-level recursion, and ACT (adaptive computational time). Because the framework does not require these structures, VAR implements only the core.

We compared the two methods under the reviewer's condition: one stochastic transition per supervision interval and the same supervision frequency. To do so, we reduced GRAM's H cycle to 1 and set the maximum number of supervision intervals to 8, the value used by VAR. On N-Queens (table below, %), the simpler VAR performs comparably to GRAM (H=1) on NQ8 and better on NQ10. In the revision, we will explicitly describe how the EM framework explains the per-step parameter updates that the trajectory ELBO does not.

ModelNQ8 AccuracyNQ8 Coverage@20NQ10 AccuracyNQ10 Coverage@20
VAR98.47 ± 0.6990.44 ± 0.0984.47 ± 0.7055.82 ± 0.14
GRAM (H=1)99.11 ± 0.3290.69 ± 0.2682.63 ± 3.2449.96 ± 1.69

[W2, Q2] How closely the learned memory updates, which do not fit the new prior to the previous posterior, approximate the ideal M-step, and how this relates to the gains from greater depth

The ideal M-step and VAR's learned update are derived separately from the same root objective, the ELBO F(q,h)\mathcal F(q,h) (line 114), rather than VAR's update being derived from the ideal M-step. Section 2 adds idealized assumptions to this objective, namely exact, target-conditioned inference and explicit memory optimization. Here, "idealized" means restrictive and impractical, not a standard that VAR's update should follow. These assumptions only make the role of each step analyzable through EM, as the paper states at the start of Section 3 ("…to specify the roles of the two steps, not to prescribe an update algorithm"). Under these assumptions, the ideal M-step fits the new prior to the previous posterior (line 126), but it needs the target and cannot run at test time (lines 180–181). Section 3 therefore does not implement the ideal M-step. It returns to the root objective without these assumptions and amortizes both steps, so that the memory can be built from prior samples without the target (line 182). Removing these assumptions provides no explicit guarantee but makes the algorithm usable in practice.

Amortizing these explicit optimizations of the memory and the posterior follows common practice. A VAE replaces per-instance posterior optimization with an amortized encoder, and trains the encoder and decoder jointly through the ELBO instead of alternating exact E- and M-steps. Iterative amortized inference (Marino et al., 2018) trains a network that outputs the updated posterior so that the resulting ELBO increases. Like VAEs and iterative amortized inference, VAR does not hold the posterior fixed and trains the adapter to directly raise the ELBO recomputed at the updated memory (line 210). The ideal M-step fits the new prior to the previous posterior only because it holds the posterior fixed (line 126).

We applied each update once from the same initial memory in the Gaussian toy of Figure 3, which is already the controlled setting the reviewer describes: the exact posterior and the ideal memory update can be computed. The two updates raise the evidence almost equally (+3.15 vs. +3.19 nats), although the KL between the new prior and the previous posterior differs (0 vs. 23.1 nats). The learned update thus raises the objective by almost as much as the ideal M-step without fitting the prior to the previous posterior.

The gains from greater depth have no ideal counterpart, because the ideal M-step needs the target yy at every step and cannot run at test time; this is one motivation for amortizing the update. The adapter instead learns a rule that raises the ELBO from the incoming memory hth_t, so in the same toy the mode-region mass keeps rising beyond the 7 updates used in training, from 47.0% after 7 updates to 61.1% after 15 (Figure 3). We will add this controlled comparison to the revision.

[W3, Q3] Whether high Sudoku validity reflects diverse, novel grids rather than reproduced training data, and whether diversity holds as depth increases

VAR generates distinct valid Sudoku grids, even up to Sudoku symmetries, with no training-set overlap or diversity collapse as depth increases. At the same sample count, distinct valid grids increase from 5,377.3 (K=8) to 39,851 (K=32), with zero duplicates.

We evaluated the 40,000-board banks from Appendix E.4 (blank input, three seeds, the same VAR checkpoints trained with N=8 steps and executed for K steps) against 50,000 training grids.

Execution depth KValid (of 40,000)Distinct validIn training set
85,377.35,377.30
1636,46036,4600
3239,85139,8510

Results are averaged over seeds; every seed individually shows zero duplicates and zero training-set overlap. Thus, the fraction of distinct grids among valid outputs remains 100% at every depth, indicating no observed concentration onto repeated solutions, while the absolute number of distinct valid grids increases with validity.

Symmetry-aware diversity. The results remain unchanged when accounting for all Sudoku symmetries (digit relabeling, row/column permutations within bands/stacks, band/stack permutations, and transposition; approximately 1.2 × 10¹² transformations). No two valid outputs are equivalent, and none is equivalent to a training grid.

Why the samples are novel. Learning the training distribution does not mean reproducing training grids. A 9×9 Sudoku has about 6.7 × 10²¹ valid grids, so a model that samples from this distribution would match one of the 50,000 training grids with expected count ~10⁻¹³ among 40,000 outputs (~0.4 even modulo symmetry). Zero overlap is therefore what distribution learning predicts, whereas memorization would produce repeated training grids.

We will include this analysis in Appendix E.

Thanks for the authors' detailed response.

I agree that the mode-region mass keeps rising beyond the training depth, and that "The gains from greater depth have no ideal counterpart". My remaining concern is how much this result explains about why the extra steps help.

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

This paper proposes Variational AutoReasoner, a recursive reasoning model that views model construction through the lens of variational EM. In AutoReasoner, the model introduces an instance-level M-step, which adds a per-problem memory that carries the information the model sees throughout the recursive reasoning steps. The model is trained using local variational objectives.

Critical strengths

Framing the proposed model through the EM view is very interesting. In particular, I feel that the memory adaptation (Sec. 2.2) and the monotonicity of EM are central properties of the method. Given this property, recursive reasoning makes sense under this framework. Although the amortized model does not have an exact guarantee like this, the framework can be a nice starting point for designing a recursive reasoning model.

Critical weaknesses

The amortized method is not explained well enough. Since it carries no formal monotonicity guarantee, the paper should at least emphasize more on the intuition for why it approximately retains the benefit of the exact EM updates. I suggest the authors emphasize this more. The paper should make clear why the amortized method approximates exact EM’s monotonicity, since the justification for recursive reasoning depends on it.

Questions for authors
  1. The memory $h_T$ is really the main “red string” that connects every step together. It is Markovian in w.r.t to the t, but I am not really sure whether going beyond the maximum N can generalize well in terms of performance for the amortized proposed model (see weaknesses). (clarification)
  2. I wonder whether it would make sense, during training, to penalize $L_{t+1} < L_t$ to force more improvement given more steps, if that makes sense at all. In Algorithm 1, I could not see any kind of hierarchical signal (later steps providing a stronger signal than earlier ones) for training the networks. This is an important property for ensuring that the monotonicity of exact inference discussed in Section 2.2 also holds more strongly in the amortized version (Algorithm 1). (clarification/future work)
  3. While the concept is very interesting and the experiments are quite well designed, the problem tackled in the experiments section is rather small and toyish. For example, the prior adaptation examples are toyish. Would it be better, for example, to use some examples, like people did in [1] and [2], where it can go beyond 2D parameters and not be very toyish. (future work)

Minor:

  • Figure clarity (Fig 1). The red dots appear to be samples of z from the prior. It would help to state this explicitly in the caption or legend. The label $a_\psi$ in panel (e) is confusing. It reads as if the red dots are h, but I understand they are samples from the prior condition on h, where h is computed by $a_\psi$. Please clarify this in the figure or caption. The red dots are plotted as if they lie in the observed y space. Please label the axes and make clear which space (z or y) is being shown.
  • Notation. $p_\theta$ is currently used for both the posterior and the decoder, which makes the text hard to follow. I suggest using distinct notation.
3 comments

[General Clarification] Monotonicity Guarantees: From EM to Amortized Learning

We thank the reviewer for the constructive review. Since W1 and Q2 both concern monotonicity, we first clarify the trade-offs from EM to amortized variational learning.

Method(Instance level) Likelihood monotonicity(Instance level) ELBO monotonicityAverage ELBO improvementExact posterior not requiredEfficient & generalizable inference via amortization
EMOOOXX
Variational EMXOOOX
VAEXXOOO

EM monotonically increases the likelihood under exact inference and improving M-steps, with a tight ELBO. Variational EM allows approximate inference for more complex models, retaining monotonic ELBO ascent rather than likelihood ascent under coordinate updates. VAEs further replace per-instance inference optimization with a shared encoder, enabling efficient inference and generalization across instances. All three optimize the dataset-average ELBO, but the stochastic joint training of VAEs does not guarantee monotonic improvement.

Sections 2 and 3 share the same underlying ELBO but serve distinct purposes. Section 2 analyzes instance-level memory adaptation under idealized (but impractical) EM assumptions, where exact inference and improving memory updates guarantee likelihood ascent. Section 3 instead develops a practical amortized algorithm directly from the same ELBO, learning inference and memory adaptation without inheriting this guarantee. As explicitly stated in Section 2.2: "The guarantee holds for the idealized process, not by itself for the learned, target-free procedure of Section 3."

[W1] Intuition for why the amortized method approximately retains the benefit of the exact EM updates without a formal monotonicity guarantee

The key intuition is that VAR learns memory updates that improve the conditions for subsequent inference through the next-step ELBO. As clarified in [General Clarification], VAR does not reproduce the exact instance-level M-step, but learns its computational role: using current inference to improve subsequent inference. Specifically, the adapter writes sampled explanations into memory, and the subsequent ELBO Lt+1L_{t+1} , evaluated with the posterior recomputed at ht+1h_{t+1}, trains the adapter to prepare a better context for the next inference. This encourages iterative improvement without explicitly enforcing monotonic ascent.

Learning operators through variational objectives rather than reproducing explicit optimization steps is a common practice in amortized inference. VAEs replace per-instance posterior optimization with a shared encoder, while iterative amortized inference (IAI; Marino et al., 2018) learns iterative posterior-refinement operators by optimizing the ELBO rather than imitating explicit gradient-ascent E-step updates. Similarly, VAR learns memory-conditioned prior adaptation through the subsequent ELBO rather than reproducing the exact instance-level M-step. Its shared operator can also be applied beyond the training depth (Q1).

This intuition is supported empirically: GC10 exhibits mostly positive posterior-driven next-ELBO gains (Figure 4a), while Sudoku shows increasing cumulative prior-driven write utility (Figure 4b), with validity improving from 13.5% (K=8) to approximately 99.5% (K=32) (Figure 4d).

In the revision, we will clarify how next-ELBO supervision encourages iterative improvement in Section 3, while distinguishing it from the monotonicity guarantee of idealized EM in Section 2.2.

[Q1] Generalization Beyond the Training Depth: Can performance continue to improve when the model is executed beyond the maximum number of iterations NN used during training?

VAR learns a shared recursive inference–adaptation operator rather than iteration-specific updates. Training through the subsequent ELBO (W1) encourages this operator to make memory updates that benefit the next inference. The same learned operator can therefore be repeatedly applied beyond the training depth, with further improvements possible when it generalizes to newly encountered memory states.

We observe such depth extrapolation by running the same frozen checkpoint for additional iterations. Under posterior-driven inference, GC10 exhibits generally positive next-ELBO gains beyond the training depth (Figure 4a). Under prior-driven inference, NQ10 accuracy increases from 84.47% at K=8 to 92.52% at K=32 (Table 9). Similarly, Sudoku validity improves from 13.5% at K=8 to approximately 99.5% at K=32 using the same N=8 checkpoint (Figure 4d, Table 5). However, parameter sharing alone does not guarantee extrapolation; its success depends on whether the learned operator remains effective on newly encountered memory states.

[Q2] Need for Explicit Training Signals to Encourage Monotonicity: penalizing cases where Lt+1<LtL_{t+1} < L_t or assigning greater weights to later iterations through hierarchical supervision?

The reviewer's suggestions of monotonicity regularization and greater weighting of later iterations are promising ways to further encourage iterative improvement. As explained in W1, we believe VAR's next-step ELBO supervision serves a similar purpose by encouraging memory updates that benefit subsequent inference, although it does not explicitly enforce monotonicity. These approaches could further strengthen this learning signal. We appreciate these suggestions and consider them interesting directions for future work.

[Q3] Insufficient Experimental Scale and Complexity: Can the proposed method demonstrate its effectiveness on higher-dimensional and more complex problems?

The 2D Gaussian toy in Figure 3 was designed to visualize prior adaptation and analyze its relationship to exact EM. Our main experiments (GC, NQ, and Sudoku) already involve much higher-dimensional latent spaces (for example, 128-dimensional latents for each of 81 Sudoku cells). However, exact inference is intractable in these tasks, preventing direct comparison with the idealized M-step.

To address this, we extended the Figure 3 experiment to 10D, matching the dimensionality of the Gaussian Linear benchmark in PriorGuide (Ref. [1] in the review), while retaining the same setup and evaluation protocol (5 seeds). Idealized EM performs 16 exact inference–adaptation iterations, while VAR is trained with N=8 and evaluated up to K=16 using the same frozen checkpoint.

Metrich0h1h3h7h15
Exact EM log-likelihood (nats) ↑−27.80−7.14−4.00−3.00−2.60
VAR log-likelihood (posterior-driven) (nats) ↑−21.57−7.50−7.63−7.83−7.98
VAR target-free mode mass (%) ↑0.01.622.346.153.1

Both methods substantially improve likelihood after the initial memory update, although their subsequent dynamics differ. Exact EM continues improving, whereas VAR stabilizes after the initial update. Importantly, VAR's target-free mode-region mass continues increasing beyond the training depth, reaching 53.1% at h15, compared to 0% with fixed memory.

[M1] Figure Caption Revision

Thank you. Fig. 1 is a schematic of the latent space z. The red dots are samples from the current prior pθ(z∣x,ht)p_θ(z | x, h_t), and the aψa_ψ arrow in panel (e) denotes the memory update that shifts this prior, so the dots are not hh. In the revision, we will state this in the caption, label the axes as zz, and describe the answer regions and the × as the explanations zz that decode to each answer and to the observed yy.

[M2] Revision of the pθp_\theta Notation

Thank you for pointing this out. In the revision, we will use distinct symbols for the prior, the decoder, and the exact posterior, which currently share pθp_θ.

Dear Reviewer kAgA,

Thank you again for your thoughtful review. We would appreciate hearing whether our responses have adequately addressed your concerns. We would also be happy to clarify any remaining questions.

Thank you for your time and consideration.

Best regards, The Authors

Reviewer_rL35OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Consider a VAE or conditional AVE model, where the prior distribution for the latent (or explanation as used in this paper) variable zz is assumed to follow from some fixed 'prior' distribution, this paper introduces an additional memory state hh for the prior. The decoder is independent of this memory state. The estimation procedure consists of the 'E' step where the ELBO is maximized over a family of qh(z)q_h(z) for a given hh, and then a 'M" step where the memory hh is updated by say minimizing the KL divergence between the variational posterior and the prior of z. What's interesting is this additional step of memory update thought the procedure which is key to the recursive reasoning as described in this paper.

They then amortized the latent posterior as in the usual VAE as well as the memory update through some mapping α\alpha that's parametrized by another inference network (the amortized network which is a separate networks from the one that amortizes the posterior of zz). They call this new procedure the Variational auto reasoner.

Critical strengths

(1) in my opinion, the idea of introducing a persistent memory as prior adaptation is interesting and novel. The distinction between refining a posterior under a fixed prior and changing the prior context is nice and potentially useful.

Critical weaknesses

I am trying to carefully understand the difference between this ideated M-step and what the learnt adapter is actually doing. The ideal E step computes the posterior of zz given prior of zz conditional on hh, and the ideal M-step just optimizes the hh that minimizes the KL divergence between the posterior q(z)q(z) and the prior pθ(z∣h)p_{\theta}(z|h). Since the decoder does not depend on hh, then the MM step is equivalent maximizing the expectation of log prior with respect to qt(z).q_t(z).

Now here is what the VAR is actually doing: one first samples a ztz_t and then passes through the update from hth_t to ht+1h_{t+1} as in the paper: ht+1=αψ(x,ht,zt)h_{t+1}=\alpha_{\psi}(x, h_t, z_t). Then one recomputes qt+1q_{t+1} based on the new memory ht+1h_{t+1} and the retrain the adapter. So basically both the posterior and the prior change with the adapter’s output. The training asks the adapter to create a memory at which the next inference problem has a high ELBO. But is this what the ideal M step is supposed to do? is this equivalent to the next prior being close to the current posterior?

since the previous variational posterior is not fixed which is recomputed with the new memory, the two are clearly not the same?

Questions for authors

see above

2 comments

[W1] Why can VAR’s learned adapter be interpreted as an amortized version of the idealized M-step, despite optimizing a different objective?

We thank the reviewer for recognizing the value of memory-based prior adaptation and for raising this important question about the relationship between the idealized M-step and the learned adapter.

Sections 2 and 3 serve distinct purposes but are rooted in the same variational objective (ELBO, line 114). Section 2 adopts deliberately idealized but impractical assumptions (exact, target-conditioned inference and explicit per-instance memory optimization) to analyze the theoretical role of memory adaptation, not to prescribe an update algorithm, as stated at the beginning of Section 3. This yields an idealized M-step (line 126) that provides insight into what the instance-level M-step does, rather than a practical algorithm. Since exact posterior inference is intractable and repeated per-instance memory optimization is expensive, Section 3 instead returns to the same ELBO (line 114) to develop a practical amortized algorithm, rather than implementing the idealized M-step derived in Section 2. Specifically, VAR trains a shared adapter through the subsequent ELBO (line 210), with the posterior recomputed at the updated memory, rather than explicitly minimizing the fixed-posterior KL of the idealized M-step. In summary, we believe the difference between the two procedures is reasonable given their distinct assumptions, while both remain rooted in the same variational objective.

The key correspondence lies in their computational role: carrying information from the current inference into the next. With a fixed prior, repeating the exact E-step yields the same posterior. The idealized M-step overcomes this by incorporating inferred information into the next prior, changing the conditions for subsequent inference. VAR preserves this inference–adaptation structure by updating the memory-conditioned prior using sampled explanations and learning updates that benefit subsequent inference. Thus, VAR amortizes the computational role of the instance-level M-step—using current inference to improve subsequent inference—rather than its exact optimization rule.

Our approach follows the established amortization principle of replacing explicit optimization with learned operations trained through variational objectives. VAEs replace per-instance posterior optimization with an amortized encoder, jointly training the encoder and generative model through the ELBO. Similarly, iterative amortized inference (IAI; Marino et al., 2018) learns iterative posterior-refinement operators through ELBO-based training rather than reproducing explicit gradient based E-step. VAR applies the same principle to posterior inference and instance-level memory adaptation.

Figure 4 provides empirical support: GC10 shows generally positive next-ELBO gains even beyond training depth, while Sudoku prior-driven validity increases from 13.5% (K=8) to approximately 99.5% (K=32). We will clarify that Sections 2 and 3 build on the same variational objective for distinct purposes, rather than deriving the learned algorithm directly from the idealized M-step.

Dear Reviewer rL35,

Thank you again for your thoughtful review. We would appreciate hearing whether our responses have adequately addressed your concerns. We would also be happy to clarify any remaining questions.

Thank you for your time and consideration.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.