acceptodds
Under review as a conference paper at ICLR 2027

Recovery-Directed Symbolic Distillation of Neural Likelihoods

Abstract

Amortized neural likelihoods enable computationally expensive inference for models with analytically intractable or unspecified likelihoods, but their black-box nature limits interpretability. We introduce a symbolic distillation pipeline that converts trained neural likelihoods into explicit, interpretable expressions optimized for efficient parameter estimation. Our approach uses a recovery-directed objective to guide symbolic regression toward expressions that preserve parameter-recovery accuracy rather than merely approximating the likelihood function. Candidate expressions are evaluated on held-out datasets and selected using a criterion that jointly accounts for expression complexity, parameter-recovery performance, and distributional distance from the learned likelihood. We evaluate the pipeline on the diffusion decision model, a classical cognitive model, whose analytically tractable likelihood provides ground truth for controlled evaluation. The proposed recovery-directed objective improves parameter recovery over standard symbolic-regression objectives. The resulting symbolic likelihoods enable over 100 times faster parameter evaluation than both neural likelihoods and, when available, the exact likelihood, while maintaining a manageable loss in precision. We further demonstrate these computational benefits in Bayesian hierarchical inference on empirical data. Our pipeline provides a lightweight interface for integrating symbolic distillation with existing neural-likelihood estimation methods and can be adapted to a range of simulation-based inference settings.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
symbolic distillation, neural likelihood estimation, diffusion-decision model (DDM), simulation-based inference

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_qXVpOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how to distill neural likelihood models into compact, interpretable, and computationally efficient symbolic expressions. The proposed method adds a parameter-recovery term to the symbolic regression objective to favor expressions that preserve downstream parameter-estimation performance, then further screens candidates through full posterior inference. The authors evaluate the recovery-directed objective on the diffusion decision model, compare the computational costs of symbolic, neural, and analytical likelihoods, and demonstrate hierarchical modeling on real behavioral data.

Critical strengths

1.Accurately approximating neural likelihood values does not necessarily preserve the structure needed for parameter estimation, and the paper addresses this mismatch directly. Adding the recovery-directed term increases the proportion of expressions passing the recovery criteria for all four parameters from 1.2% to 19.5%. The authors also report results using independent search runs as the unit of replication, avoiding treating correlated expressions from the same run as independent observations. This is the paper’s most convincing contribution.

2.Expression 2 contains only 25 nodes, and its structure is related to the exponent in a known first-passage-time density. The runtime comparisons include compiled implementations of the exact likelihood and LAN, as well as gradient evaluations, rather than merely comparing a simple formula against an unoptimized network. These comparisons provide convincing evidence of practically meaningful computational gains.

Critical weaknesses

My main concern is whether the symbolic expressions preserve reliable parameter inference. High recovery correlations can mask systematic bias, and the SBC results concern normalized expressions rather than the unnormalized functions used in the main experiments. Reporting bias, RMSE, and interval coverage for the actual inference procedure would help establish whether the speed–accuracy trade-off is acceptable.

Questions for authors
  1. Could you report parameter bias, RMSE, and interval coverage for the unnormalized expressions used in the main experiments? High recovery correlations alone may not indicate accurate estimates or uncertainty.

  2. Equation (6) evaluates each parameter while fixing the others at their true values. How well does this proxy predict recovery when all parameters are estimated jointly, especially when parameters interact or compensate for one another?

3.The abstract states that candidate selection jointly considers complexity, parameter recovery, and distributional distance, whereas Section 2.3 describes ranking primarily by the weakest parameter-recovery correlation. Please clarify how these criteria enter the selection process.

3 comments
AuthorsOpenReview ↗

Clarification to pipeline details

We sincerely appreciate your constructive comments, which have helped us substantially improve the manuscript and rethink directions for further improvement. We are grateful for your recognition of the novelty, rigor, and empirical contributions of our pipeline design. Each weakness and question is addressed point by point below, grouped by topic. Revised content from the manuscript is quoted in our responses.

Regarding the one-at-a-time recovery proxy.{\color{#f26921}{\text{Regarding the one-at-a-time recovery proxy.}}}

Question-2:{\color{#f26921}{\text{Question-2:}}} Equation (6) evaluates each parameter while fixing the others at their true values. How well does this proxy predict recovery when all parameters are estimated jointly, especially when parameters interact or compensate for one another?

Response:{\color{#f26921}{\text{Response:}}} We agree. Because R(f)\mathcal{R}(f) perturbs one parameter at a time, it misses trade-offs that arise in joint estimation, and at large λ\lambda the search can exploit this; the proxy is therefore not always aligned with joint recovery under NUTS. We now compute the recovery proxy R(f)R(f) for all 3,371 candidates (achieved before stage 3) across five different λ\lambda values and compare it with NUTS rr for joint sampling on the 60 test subjects (columns 3-4).

λ\lambdamean R(f)R(f)correlation with weakest rrcorrelation with mean rrall r>0.5r>0.5 (selected / pool)
0−0.190.290.3912 / 968 (1.2%)
0.30.080.420.5455 / 560 (9.8%)
10.240.610.5995 / 486 (19.5%)
30.380.110.1199 / 634 (15.6%)
100.47−0.14−0.0715 / 723 (2.1%)

We found that Correlation between R(f)\mathcal{R}(f) and NUTS recovery performance doesn't always positive. We report spearman correlation rate within each λ\lambda, since λ\lambda shifts R(f)R(f) itself (see mean R(f)R(f) column).

  • λ=0\lambda = 0: Only the plain MSE objective is used, yet it still predicts recovery (0.29/0.39). That means the proxy carries some real signal when the expression is close to the learnt likelihood.
  • λ=0.3–3\lambda = 0.3–3: the agreement is high, and the expression NUTS ranks first is in the top 2–6% by R(f)R(f).
  • λ=10\lambda = 10: the search exploits a shortcut in the proxy: the mean R(f)\mathcal{R}(f) across candidates reaches 0.47, yet only 2.1% recover all four parameters under NUTS. R(f)\mathcal{R}(f) no longer predicts NUTS recovery (ρ=−0.14\rho = -0.14 / −0.07-0.07).

We include R(f)\mathcal{R}(f) in the recovery-directed objective because it is cheap enough to evaluate for every candidate during the search, whereas NUTS is not. However, since R(f)\mathcal{R}(f) is only an imperfect proxy, we never select expressions on R(f)\mathcal{R}(f). Instead, we include stage 3 to screen and rank expressions by NUTS recovery. Because λ\lambda balances likelihood approximation against parameter recovery, it should be kept moderate and swept over several values.

We have added this analysis to Appendix A.2 VALIDITY OF THE RECOVERY PROXY in the manuscript.

Regarding the selection criteria.{\color{#f26921}{\text{Regarding the selection criteria.}}}

Question-3:{\color{#f26921}{\text{Question-3:}}} The abstract states that candidate selection jointly considers complexity, parameter recovery, and distributional distance, whereas Section 2.3 describes ranking primarily by the weakest parameter-recovery correlation. Please clarify how these criteria enter the selection process.

Response:{\color{#f26921}{\text{Response:}}} We apologize for the confusion. Section 2.3 is more rigorous, as only the recovery metric is used when ranking expressions in F∗\mathcal{F}^*. However, expression selection runs through the whole pipeline, not only Stage 3. Therefore, the abstract compressed it into one sentence that reads as a single joint score. The three criteria enter in sequence:

  • Stage 2: complexity and distributional distance, through the PySR loss and the accuracy–complexity Pareto front (node count ≤ 50). Expressions that fit the neural likelihood poorly for these two criteria never reach Stage 3.
  • Stages 3–4: parameter recovery. Every expression on the front is fit with NUTS on the 60 test subjects; the survivors of the pass criterion are ranked by their weakest r. This is the only ranking criterion.

Distributional distance and complexity act mainly as filters. They admit qualified expressions into F∗\mathcal{F}^*, which are then ranked and checked for calibration in Stages 3–4. We have updated the corresponding sentence in the abstract:

"...Candidate expressions are searched on an accuracy–complexity Pareto front, validated with a full Bayesian sampler on simulated subjects, and ranked by their weakest parameter-recovery correlation...."

Thank you for your response. It has addressed most of my concerns, and I will maintain my positive rating.

AuthorsOpenReview ↗

Follow up response on parameter inference

We really appreciate your quick response and positive confirmation. Here is the response to your main concern.

Weakness:{\color{#f26921}{\text{Weakness:}}} My main concern is whether the symbolic expressions preserve reliable parameter inference. High recovery correlations can mask systematic bias, and the SBC results concern normalized expressions rather than the unnormalized functions used in the main experiments. Reporting bias, RMSE, and interval coverage for the actual inference procedure would help establish whether the speed–accuracy trade-off is acceptable.

Question-1:{\color{#f26921}{\text{Question-1:}}} Could you report parameter bias, RMSE, and interval coverage for the unnormalized expressions used in the main experiments? High recovery correlations alone may not indicate accurate estimates or uncertainty.

Response:{\color{#f26921}{\text{Response:}}} Your concern about the posterior inference on unnormalized expression alone is fair. The unnormalized expression defines a generalized posterior q(y∣θ)q(y \mid \theta), whose posterior mean/median is a consistent estimate of the true θ0\theta_0 only if Ep(y∣θ0)exp⁡fbest∗(y∣θ)E_{p({y} \mid {\theta_0})} \exp f_{best}^*({y} \mid {\theta}) is maximized at θ=θ0{\theta} = {\theta_0} (Bissiri et al., 2016).

However, due to the missing θ-dependent normalizer in exp⁡fbest∗(y∣θ)\exp f_{best}^*({y} \mid {\theta}), the reported uncertainty metrics obtained directly from MCMC have no calibration guarantee. For uncertainty quantification, we therefore recommend the normalized likelihood exp⁡f(θ∣y)∫exp⁡f(θ∣y′)dy′\frac{\exp f(\theta \mid y)}{\int \exp f(\theta \mid y') \mathrm{d}y'}, whose posterior is a standard Bayesian posterior under the approximate model. That is the reason why the SBC results in Appendix H.2 are based on normalized expression. We now state explicitly in the manuscript when the unnormalized expression can be used directly and when normalization and additional calibration are needed in the METHOD section.

"Figure 1 summarizes the pipeline,...Finally, the top-ranked expression fbest∗(y∣θ)f_{\text{best}}^{*}(y \mid \theta) is used in MCMC sampling. Since q(θ∣y)q(\theta \mid y) is a generalized posterior (Bissiri et al., 2016), its mean serves as a consistent point estimate. Yet its credible regions require additional calibration, such as normalizing the expression by exp⁡(f(y∣θ))∫exp⁡(f(y′∣θ)) dy′\frac{\exp (f({y}\mid{\theta}))}{\int \exp (f({y}'\mid{\theta}))\,\mathrm{d}{y}'}."

In Section 5 LIMITATIONS AND FUTURE DISCUSSION, we also describe the future improvements:

" First, the discovered expressions are unnormalized densities. A further credible region calibration is needed for uncertainty quantification (Syring & Martin, 2019). "

To respond to your question directly, we fit the 90 held-out subjects (1000 trials for each) with each of the top five expressions. The HSSM-LAN serves as a reference. All fits used NUTS with 4 chains of 500 warm-up and 500 sampling iterations. Bias and RMSE are those of the posterior mean derived from unnormalized expressions{\colorbox{#dbeafe}{unnormalized expressions}} against the true value, while the 95% credible interval coverage rate is reported by normalized expressions with credible intervals calibrated by a bootstrap-chosen learning rate (Syring & Martin, 2019). Specifically, RMSE is given relative to the parameter range to unify the parameter unit.

LikelihoodBias (v / a/ z / t){\colorbox{#dbeafe}{Bias (v / a/ z / t)}}RMSE/range ( v / a/ z / t){\colorbox{#dbeafe}{RMSE/range ( v / a/ z / t)}}95% CI coverage ((v / a/ z / t)
HSSM LAN+0.018 / +0.051 / −0.008 / +0.0020.038 / 0.118 / 0.063 / 0.01576% / 76% / 70% / 53%
Top-five expressions, mean+0.096 / +0.106 / −0.000 / +0.004{\colorbox{#dbeafe}{+0.096 / +0.106 / −0.000 / +0.004}}0.112 / 0.133 / 0.083 / 0.027{\colorbox{#dbeafe}{0.112 / 0.133 / 0.083 / 0.027}}38% / 27% / 47% / 32%

Admittedly, the table shows that interval coverage remains low even for the normalized and calibrated expressions: on average 27–47% across parameters, against the nominal 95%. The HSSM LAN also falls short of the nominal (53–76%). So, part of the gap is common to approximate likelihoods, but the expressions are clearly worse. It is worth noting, however, that the posterior means of the unnormalized expressions (highlighted in blue) show little bias on average. Whether the speed–accuracy trade-off is acceptable, therefore, depends on what the analysis needs: it is very attractive for point estimates, but may not be economical when calibrated credible intervals are.

Thank you for your thoughtful comments. We hope our response addresses your concerns and are happy to discuss any further questions.

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

The authors introduce a novel symbolic distillation pipeline for converting neural likelihoods into expressions optimized for parameter efficient estimation and fast likelihood computation. The authors propose using a recovery directed objective to bias search towards expressions preserving parameter recovery accuracy. This change is empirically validated to be valuable in improving performance. The authors show over 100 x faster parameter evaluation over neural likelihoods and exact likelihood with manageable loss in accuracy.

Critical strengths
  • Clear presentation.
  • A novel use of the recovery directed loss function with empirical evidence of its value.
  • Compelling results. Notably, recovery of the DDM parameters with drastically improved computation.
Critical weaknesses
  • The authors acknowledge an important limitation, which is "approach matters most for models without a tractable likelihood, such as those with collapsing bounds, and applying the recovery-directed objective to such models is the most important next step." The results are currently scoped to recovery of DDM parameters.
Questions for authors
  • For reproducibility purposes, can you share the hardware details for results in table 2?
  • What is the total cost of symbolic distillation and how does this change the practical interpretation of results in 4.3?
2 comments
AuthorsOpenReview ↗

Response regarding hardware details and time cost of symbolic distillation.

Thank you for acknowledging our research’s presentation, methodological novelty, and results. Each weakness and question is addressed and clarified point by point and grouped by topic below. Revised content from the manuscript is quoted in our responses.

Regarding hardware details for Table 2.{\color{#f26921}{\text{Regarding hardware details for Table 2.}}}

Question-1:{\color{#f26921}{\text{Question-1:}}} For reproducibility purposes, can you share the hardware details for results in table 2?

Response:{\color{#f26921}{\text{Response:}}} All analyses were run on an Apple MacBook Pro with an Apple M5 Max processor(18 CPU cores) and 128 GB of unified memory, running macOS 26.6.1. The expression and the analytical series (HSSM, PyTensor) and the LAN (HSSM, JAX) were timed in Python 3.12.13 with HSSM 0.4.0, and the neural likelihood in Python 3.11 with BayesFlow 2.0.12 on the PyTorch backend. Each entry is the median of repeated calls after warm-up (15 calls, or 3 for the neural likelihood). The empirical fits in Figure 5b ran on the same machine. We have updated this information in the REPRODUCIBILITY STATEMENT.

Regarding the cost of symbolic distillation.{\color{#f26921}{\text{Regarding the cost of symbolic distillation.}}}

Question-2:{\color{#f26921}{\text{Question-2:}}} What is the total cost of symbolic distillation and how does this change the practical interpretation of results in 4.3?

Response:{\color{#f26921}{\text{Response:}}} The cost of symbolic distillation is one-time, similar to the amortized SBI training. The discovered expression is reused for every later fit and dataset, and the per-evaluation savings grow with dataset size (Table 2). This suits cognitive science well, where the same model is often refit many times: across participants and experiments, in hierarchical analyses, and in parameter recovery studies. Such a one-time investigation on symbolic distillation will help researchers to quickly validate their proposed model quickly.

This upfront cost does not change the main message of Section 4.3. The neural likelihood and the LAN in our benchmark also need one-time training, and distillation adds a further one-time cost that is recovered over repeated fits. Admittedly, once the costs of training and distillation are both counted, the analytical likelihood remains the better choice for the plain DDM. However, the plain DDM is used as a validation against known ground truth, and the proposed pipeline is aimed at models whose likelihood has no closed form.

AuthorsOpenReview ↗

Regarding application on models without a tractable likelihood.

Weakness-1:{\color{#f26921}{\text{Weakness-1:}}} The authors acknowledge an important limitation, which is "approach matters most for models without a tractable likelihood, such as those with collapsing bounds, and applying the recovery-directed objective to such models is the most important next step." The results are currently scoped to recovery of DDM parameters.

Response:{\color{#f26921}{\text{Response:}}} We currently include an additional example, angle DDM in the METHOD section of the manuscript:

" In addition, we introduce angle-DDM, whose boundaries b(τ)b(\tau) collapse linearly at θcollapse\theta_{collapse} over time τ\tau: b(τ)=a−τtan⁡θcollapse,b(\tau) = a - \tau \tan\theta_{collapse}, θ=(v,a,z,t,θcollapse){\theta} = (v, a, z, t, \theta_{collapse}) comprises the four standard DDM parameters and the angle θcollapse\theta_{collapse} at which the bounds collapse linearly over time.Angle-DDM has no analytical likelihood, and hence its likelihood is mostly approximated by SBIs. "

The same pipeline as for the DDM is applied, with results discussed in Section 4.5 EXTENSION TO INTRACTABLE LIKELIHOODS :

"Above, we have validated our framework on the plain DDM against its analytical solution. However, intractable likelihoods are a major motivation for this technique. We now show that the same pipeline, with only the simulator changed to angle-DDM (see Simulator Specification, Section 2.1), also yields closed-form expressions that recover the parameters. Stage 2 used two neural likelihoods as references: a pretrained LAN distributed with HSSM, and our own neural likelihood. Even for these, angle-DDM is harder to estimate than the DDM. The LAN's weakest parameter is aa (r=.54r = .54), and ours is θcollapse\theta_{\mathrm{collapse}} (r=.48r = .48; Figure 6a, top two rows). Ranked by weakest correlation, the top expression recovered every parameter at r≥.73r \ge .73: f = \frac{1}{\sqrt{\mathrm{RT}}}\left[-\tau e^{z^{2}} + \frac{a - \tau + \big(a - \rho\,(\tau + z^{4})\big)^{2}\log z}{a}\right], \tag{9} where τ=RT−t\tau = \mathrm{RT} - t and ρ=θcollapse+z (v+RT θcollapse4)\rho = \theta_{\mathrm{collapse}} + \sqrt{z}\,(v + \sqrt{\mathrm{RT}}\,\theta_{\mathrm{collapse}}^{4}), shown in Figure 6a (bottom row). As in the DDM expression (Equation 9), the squared term penalizes a mismatch between the boundary aa and a rate multiplied by a time, and here θcollapse\theta_{\mathrm{collapse}} adds to the drift in that rate, as expected when evidence drifting at vv meets a boundary falling at tan⁡θcollapse\tan\theta_{\mathrm{collapse}}. The recovery term was again what made recovery possible. Without it (λ=0\lambda = 0), none of 10,750 expressions from 384 independent searches recovered all five parameters. With it (λ=1\lambda = 1), 548 of 52,333 (1.05%) did (Figure 6a). The collapse angle was the main obstacle, recovered by 0.1% of expressions without the recovery term and 8.3% with it (Figure 6b). Finally, we refit all 583 expressions on 90 additional subjects never used for selection. The top expression again recovered all five parameters (weakest r=.68r = .68, against .75.75 for the LAN). Selection on 60 subjects was optimistic, as expected: 186 of the 583 (32%) recovered all five parameters on the additional subjects, and 349 of the 397 that did not failed on θcollapse\theta_{\mathrm{collapse}} alone (Appendix I)."

As figures cannot be displayed in OpenReview responses, please refer to the revised manuscript for Figures 6 and 7.

We value the discussions with you and hope our response addresses your concerns. Please let us know if there are any additional questions.

Reviewer_xQM1OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a pipeline that distills a trained amortized neural likelihood into a closed-form expression optimized for parameter recovery rather than density fit. Stage 1 trains a normalizing flow for reaction times and a classifier for choices, and admits the neural likelihood only after diagnostic checks. Stage 2 runs PySR with a recovery-directed objective: MSE on the log-likelihood plus λ(1−R(f)) plus a hard constraint term B(f), where R(f) is a grid-based parameter-recovery proxy and B(f) penalizes expressions that omit a parameter or contain constant leaves. Stages 3–4 screen candidates with NUTS on held-out subjects, require r > 0.5 for every parameter and limited boundary pile-up, and rank by the weakest parameter recovery. The pipeline is evaluated on the diffusion decision model (DDM) with parameters (v, a, z, t), where the exact likelihood provides ground truth, and is demonstrated on a hierarchical empirical dataset.

Critical strengths
  1. Well-motivated and novel problem. Distilling a neural likelihood into an interpretable, fast expression is a compelling way to open the black box of amortized SBI. The key insight—that density fit and parameter recovery are different objectives—is important and well argued.

  2. Careful pipeline design. The diagnostic gate before distillation, the recovery proxy, the hard constraints, the NUTS screening, and the PPC/SBC checks are all sensible. The paper is more rigorous than typical “symbolic regression for science” work.

  3. Strong practical results. A 25-node expression recovers all four DDM parameters with r = .94/.91/.88/.97, and the speedups are substantial (0.010 ms vs. 0.27–1144 ms at 1000 trials). The hierarchical demo reports a 27× speedup for the simpler model and good convergence for the full model.

  4. Effective and simple objective. The recovery term raises full recovery from 1.2% at λ=0 to 19.5% at λ=1, and the λ sweep provides a clear picture of the trade-off. The constraint term is a simple way to prevent degenerate expressions.

  5. Honest limitations. The paper explicitly discusses unnormalized densities, the generalized posterior, poor z ranking, and NUTS convergence failures. This transparency is commendable.

  6. Good exposition and useful appendix. The pipeline figure, the expressions in Table 1, and the diagnostic details make the method reproducible in principle.

Critical weaknesses
  1. Only one model with an exact likelihood is evaluated. The main motivation is models without a tractable likelihood, but the only application is the DDM, whose exact likelihood is known. This is a reasonable controlled validation, but it does not demonstrate the method’s value in the setting it targets. At minimum, the authors should apply the pipeline to a DDM extension without a closed-form likelihood (e.g., collapsing bounds or a leaky accumulator), compare parameter recovery to amortized SBI, and report the cost–accuracy trade-off.

  2. The expressions are unnormalized, so they define a generalized posterior. The paper acknowledges this, but the practical consequences are not quantified: posterior intervals have no calibration guarantee, marginal likelihoods/model comparisons are unavailable, and the missing θ-dependent normalizer may explain some group-level differences. The speed advantage is real, but users may need a normalized density. The paper should discuss when the generalized posterior is sufficient and when it is not.

  3. The recovery proxy is coarse and is also the selection target. R(f) uses a grid over each parameter and a correlation-based score, and the search optimizes this proxy. This creates a shortcut risk: expressions may fit the proxy rather than true recovery. The NUTS screening mitigates this, but the paper should report the correlation between the proxy and NUTS recovery across all 3371 candidates, and show that the proxy-selected expressions generalize to held-out subjects.

  4. λ is highly sensitive, and there is no principled selection rule. Full recovery is 1.2% at λ=0, 19.5% at λ=1, 15.6% at λ=3, and 2.1% at λ=10; the best λ also differs by parameter (z at λ=3, v at λ=1, t degrades at λ=10). The paper reports the sweep but does not provide a validation-based way to choose λ. A practical user would not know which λ to use.

  5. NUTS convergence is imperfect. For Expressions 1, 3, and 4, R̂ remains above 1.01 for 25–36 of 90 subjects; Expression 2 has 3303 divergent transitions. The reported recovery correlations may be affected. The paper should report results excluding non-converged subjects and discuss whether the ranking changes.

  6. The empirical hierarchical demo is limited. It uses a single dataset and compares only to the analytical likelihood, not to the neural likelihood or a standard DDM package. The 73 s vs. 174 s median is a modest speedup; the 27× claim is for a simpler model and a different comparison. A broader empirical evaluation would strengthen the practical claims.

  7. The t parameter is estimated from the fastest responses. This makes it sensitive to fast guesses and RT contaminants. The paper should report sensitivity to RT contamination and to the treatment of the lower RT bound.

  8. The symbolic search is treated as a black box. The method relies on PySR with a custom objective, but the paper does not justify the choice of search engine, operator set, or constant optimization procedure, nor does it test a second SR engine. Because the recovery term changes the search landscape, the conclusions may be specific to PySR's inductive bias. A brief justification and, if feasible, a comparison with a second engine (e.g., a GP or MCTS-based SR implementation) on a subset of candidates would strengthen the claim that the gains come from the objective rather than the search engine.

Questions for authors
  1. Can you apply the pipeline to a DDM extension without a tractable likelihood and compare parameter recovery and speed to amortized SBI? This would directly test the core motivation.
  2. How should a practitioner choose λ? Can you use a validation-based criterion rather than reporting the best λ?
  3. What is the correlation between the grid recovery proxy R(f) and NUTS recovery across the 3371 candidates? Does optimizing the proxy select expressions that generalize?
  4. How do you handle NUTS convergence failures? Are the reported correlations robust to excluding divergent or high-R̂ subjects?
  5. The top expression is ranked on 60 subjects and then checked on 90 held-out subjects. Is the final selection made before seeing the 90, and are all reported correlations on the held-out set?
  6. How sensitive are the results to RT contamination and to the lower RT bound?
  7. Can you compare the hierarchical inference to the neural likelihood and to a standard DDM package, not only the analytical likelihood?
  8. Will the code, data, and trained neural likelihood be released?
6 comments
AuthorsOpenReview ↗

Regarding an example that the model has intractable likelihood.

We sincerely appreciate your constructive comments, which have helped us improve the manuscript substantially. We are grateful for our pipeline design’s novelty, rigor, and empirical contributions. Each weakness and question is addressed and clarified point by point, grouped by topic below. Revised content from the manuscript is quoted in our responses.

Weakness-1:{\color{#f26921}{\text{Weakness-1:}}} Only one model with an exact likelihood is evaluated. The main motivation is models without a tractable likelihood, but the only application is the DDM, whose exact likelihood is known. This is a reasonable controlled validation, but it does not demonstrate the method's value in the setting it targets. At minimum, the authors should apply the pipeline to a DDM extension without a closed-form likelihood (e.g., collapsing bounds or a leaky accumulator), compare parameter recovery to amortized SBI, and report the cost–accuracy trade-off.

Question-1:{\color{#f26921}{\text{Question-1:}}} Can you apply the pipeline to a DDM extension without a tractable likelihood and compare parameter recovery and speed to amortized SBI? This would directly test the core motivation.

Response:{\color{#f26921}{\text{Response:}}}

We currently include an additional example, angle DDM, in the METHOD section of the manuscript:

" In addition, we introduce angle-DDM, whose boundaries b(τ)b(\tau) collapse linearly at θcollapse\theta_{collapse} over time τ\tau: b(τ)=a−τtan⁡θcollapse,b(\tau) = a - \tau \tan\theta_{collapse}, θ=(v,a,z,t,θcollapse){\theta} = (v, a, z, t, \theta_{collapse}) comprises the four standard DDM parameters and the angle θcollapse\theta_{collapse} at which the bounds collapse linearly over time.Angle-DDM has no analytical likelihood, and hence its likelihood is mostly approximated by SBIs. "

The same pipeline as for the DDM is applied, with results discussed in Section 4.5 EXTENSION TO INTRACTABLE LIKELIHOODS :

"Above, we have validated our framework on the plain DDM against its analytical solution. However, intractable likelihoods are a major motivation for this technique. We now show that the same pipeline, with only the simulator changed to angle-DDM (see Simulator Specification, Section 2.1), also yields closed-form expressions that recover the parameters. Stage 2 used two neural likelihoods as references: a pretrained LAN distributed with HSSM, and our own neural likelihood. Even for these, angle-DDM is harder to estimate than the DDM. The LAN's weakest parameter is aa (r=.54r = .54), and ours is θcollapse\theta_{\mathrm{collapse}} (r=.48r = .48; Figure 6a, top two rows). Ranked by weakest correlation, the top expression recovered every parameter at r≥.73r \ge .73: f = \frac{1}{\sqrt{\mathrm{RT}}}\left[-\tau e^{z^{2}} + \frac{a - \tau + \big(a - \rho\,(\tau + z^{4})\big)^{2}\log z}{a}\right], \tag{9} where τ=RT−t\tau = \mathrm{RT} - t and ρ=θcollapse+z (v+RT θcollapse4)\rho = \theta_{\mathrm{collapse}} + \sqrt{z}\,(v + \sqrt{\mathrm{RT}}\,\theta_{\mathrm{collapse}}^{4}), shown in Figure 6a (bottom row). As in the DDM expression (Equation 9), the squared term penalizes a mismatch between the boundary aa and a rate multiplied by a time, and here θcollapse\theta_{\mathrm{collapse}} adds to the drift in that rate, as expected when evidence drifting at vv meets a boundary falling at tan⁡θcollapse\tan\theta_{\mathrm{collapse}}. The recovery term was again what made recovery possible. Without it (λ=0\lambda = 0), none of 10,750 expressions from 384 independent searches recovered all five parameters. With it (λ=1\lambda = 1), 548 of 52,333 (1.05%) did (Figure 6a). The collapse angle was the main obstacle, recovered by 0.1% of expressions without the recovery term and 8.3% with it (Figure 6b). Finally, we refit all 583 expressions on 90 additional subjects never used for selection. The top expression again recovered all five parameters (weakest r=.68r = .68, against .75.75 for the LAN). Selection on 60 subjects was optimistic, as expected: 186 of the 583 (32%) recovered all five parameters on the additional subjects, and 349 of the 397 that did not failed on θcollapse\theta_{\mathrm{collapse}} alone (Appendix I)."

As figures cannot be displayed in OpenReview responses, please refer to the revised manuscript for Figures 6 and 7.

AuthorsOpenReview ↗

Response to the unnormalized expressions + NUTS convergence failure

Regarding to the unnormalized expressions.{\color{#f26921}{\text{Regarding to the unnormalized expressions.}}}

Weakness-2:{\color{#f26921}{\text{Weakness-2:}}} The expressions are unnormalized, so they define a generalized posterior. The paper acknowledges this, but the practical consequences are not quantified: posterior intervals have no calibration guarantee, marginal likelihoods/model comparisons are unavailable, and the missing θ-dependent normalizer may explain some group-level differences. The speed advantage is real, but users may need a normalized density. The paper should discuss when the generalized posterior is sufficient and when it is not.

Response:{\color{#f26921}{\text{Response:}}} We first note that the timing comparison in Figure 4 won't change under a normalized posterior. Both our symbolic expression and the LAN likelihoods in HSSM are unnormalized log-likelihoods. In HSSM, the LAN log-likelihood is the raw network output. distribution_utils/onnx.py concatenates the parameters with each trial’s (rt, choice), passes them through the network, and returns the output as that trial’s log-likelihood (for implementation, see LAN-JAX,LAN-PyTensor). Downstream code never normalizes it over (rt, choice).

We now state in Section 2 METHOD when the generalized and the normalized posteriors are used:

"Figure 1 summarizes the pipeline,...Finally, the top-ranked expression fbest∗(y∣θ)f_{\text{best}}^{*}(y \mid \theta) is used in MCMC sampling. Since q(θ∣y)q(\theta \mid y) is a generalized posterior (Bissiri et al., 2016), its mean serves as a consistent point estimate. Yet its credible regions require additional calibration, such as normalizing the expression by exp⁡(f(y∣θ))∫exp⁡(f(y′∣θ)) dy′\frac{\exp (f({y}\mid{\theta}))}{\int \exp (f({y}'\mid{\theta}))\,\mathrm{d}{y}'}."

In Section 5 LIMITATIONS AND FUTURE DISCUSSION, we also describe the future improvements:

"First, the discovered expressions are unnormalized densities. A further credible region calibration is needed for uncertainty quantification (Syring & Martin, 2019)."

Regarding to NUTS convergence failure.{\color{#f26921}{\text{Regarding to NUTS convergence failure.}}}

Weakness-5:{\color{#f26921}{\text{Weakness-5:}}} NUTS convergence is imperfect. For Expressions 1, 3, and 4, R̂ remains above 1.01 for 25–36 of 90 subjects; Expression 2 has 3303 divergent transitions. The reported recovery correlations may be affected. The paper should report results excluding non-converged subjects and discuss whether the ranking changes.

Question-4:{\color{#f26921}{\text{Question-4:}}} How do you handle NUTS convergence failures? Are the reported correlations robust to excluding divergent or high-R̂ subjects?

Response:{\color{#f26921}{\text{Response:}}} We monitored R^\hat{R} and divergences for every subject but did not exclude any. Instead, we tested whether non-convergence could change the verdicts: 473 expressions near the criterion or in the worst decile of R^\hat{R} or divergences were refit with 4× the draws and 2× the chains, and 458 verdicts were unchanged (Appendix H.2).

Recovery is robust to excluding non-converged fits. we now recompute Table 6 in Appendix H.2 on the 90 held-out subjects after excluding non-converged subjects. Each cell gives the lowest rr across the four parameters, with the number of subjects kept in parentheses. All five expressions still recover every parameter under both exclusions, and the lowest rr changes by −0.01 to +0.03.

all subjects (reported in Table 6)R^≤1.01\hat{R} \le 1.01no divergences
Expression 10.87 (90)0.86 (65)0.86 (65)
Expression 20.90 (90)0.90 (84)0.90 (74)
Expression 30.81 (90)0.84 (54)0.83 (49)
Expression 40.84 (90)0.87 (57)0.87 (56)
Expression 50.86 (90)0.86 (90)0.86 (89)
AuthorsOpenReview ↗

Regarding the recovery proxy validity.

Weakness-3:{\color{#f26921}{\text{Weakness-3:}}} The recovery proxy is coarse and is also the selection target. R(f) uses a grid over each parameter and a correlation-based score, and the search optimizes this proxy. This creates a shortcut risk: expressions may fit the proxy rather than true recovery. The NUTS screening mitigates this, but the paper should report the correlation between the proxy and NUTS recovery across all 3371 candidates, and show that the proxy-selected expressions generalize to held-out subjects.

Question-3:{\color{#f26921}{\text{Question-3:}}} What is the correlation between the grid recovery proxy R(f) and NUTS recovery across the 3371 candidates? Does optimizing the proxy select expressions that generalize?

Response:{\color{#f26921}{\text{Response:}}} We now compute R(f)\mathcal{R}(f) for all 3,371 candidates across five different λ\lambda values and compare it with NUTS recovery on the 60 test subjects (columns 3-4). To check that R(f)\mathcal{R}(f) itself generalizes, we also recomputed it on 200 training groups the search never used and correlated it with the search value (last column\colorbox{#dbeafe}{last column}).

λ\lambdamean R(f)\mathcal{R}(f)correlation with weakest rrcorrelation with mean rrall r>0.5r>0.5 (selected / pool){\colorbox{#dbeafe}{\mathcal{R}(f)$ generalization}}$
0−0.190.290.3912 / 968 (1.2%)0.98{\colorbox{#dbeafe}{0.98}}
0.30.080.420.5455 / 560 (9.8%)0.96{\colorbox{#dbeafe}{0.96}}
10.240.610.5995 / 486 (19.5%)0.92{\colorbox{#dbeafe}{0.92}}
30.380.110.1199 / 634 (15.6%)0.92{\colorbox{#dbeafe}{0.92}}
100.47−0.14−0.0715 / 723 (2.1%)0.84{\colorbox{#dbeafe}{0.84}}

Correlation between R(f)\mathcal{R}(f) and NUTS recovery performance is not always postive. We report spearman correlation rate within each λ\lambda, since λ\lambda shifts R(f)\mathcal{R}(f) itself (see mean R(f)\mathcal{R}(f) column).

  • λ=0\lambda = 0: Only the plain MSE objective is used, yet it still predicts recovery (0.29/0.39). That means the proxy carries some signal when the expression is close to the learnt likelihood.
  • λ=0.3–3\lambda = 0.3–3: the agreement is high, and the expression NUTS ranks first is in the top 2–6% by R(f)R(f).
  • λ=10\lambda = 10: the search exploits a shortcut in the proxy: the mean R(f)\mathcal{R}(f) across candidates reaches 0.47, yet only 2.1% recover all four parameters under NUTS. R(f)\mathcal{R}(f) no longer predicts NUTS recovery (ρ=−0.14\rho = -0.14 / −0.07-0.07).

As a result, we never select expressions on R(f)\mathcal{R}(f). Instead, we include stage 3 to screen and rank expressions by NUTS recovery. Because λ\lambda balances likelihood approximation against parameter recovery, it should be kept moderate and swept over several values. The proxy itself transfers to new groups at every λ\lambda (last column), and under moderate λ\lambda the expressions it favors do recover under NUTS on held-out subjects. The top five expressions in Table 1, all found at λ\lambda = 1 or 3, rank in the top 15% of their searches by R(f)\mathcal{R}(f), and each still recovers every parameter on 90 further held-out subjects (Table 5, Appendix H.2).

We have added this analysis to Appendix A.2 VALIDITY OF THE RECOVERY PROXY in the manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.