acceptodds
Under review as a conference paper at ICLR 2027

An Architecture for Predicate Learning and its Uses in Planning

Abstract

We introduce a new architecture for learning unary and binary predicates from positive and negative examples over relational structures with a fixed set of primitive predicates. The architecture, called Neural Description Logic Machine or NDLM, is based on a rich description logic which has the expressive power of 3-variable first-order logic, FO. The architecture is restricted to predicates of arity one and two, and unlike related architectures such as neural logic machines (NLMs), edge-transformers, and (oblivious) 3-GNNs, it achieves this expressivity requiring quadratic rather than cubic space. The predicate learning task considered is a generalization and variation of inductive logical tasks such as concept learning and inductive logic programming but where the representation of the learned predicates is not given in terms of a first-order formula or logic program. Still we show that the learned, neural representation of concepts and roles is not fully opaque either, and in many cases, can be understood logically as well. We also show that a number of tasks in "learning for planning" can be reduced to predicate learning tasks of this form, including lifted STRIPS model learning from state-action traces, and general policy learning from plans. The performance of NDLMs is evaluated experimentally over a number of planning and concept learning tasks where it is compared with NLMs and state-of-the-art symbolic and neural unary and binary-relational learners.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
concept learning, predicate learning, classical planning, learning for planning

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_PDe2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Neural Description Logic Machines (NDLMs), an architecture for predicate learning and its applications to planning. It represents relational structures with unary and binary predicates. Fixed operations, including role composition and quantification, generate intermediate logical features, which are then combined by learnable feed-forward networks. The main idea is to support reasoning with three logical variables without explicitly storing ternary predicate tensors, yielding quadratic activation memory in the number of objects. The paper analyzes NDLM’s expressivity with respect to a three-variable fragment of first-order logic and evaluates it on graph classification, action-model learning, and general-policy learning. The performances are comparable on several planning tasks and show generalization across instance sizes, but NDLM does not consistently outperform existing methods on graph classification and concept learning.

Critical strengths
  1. NDLM uses fixed relational operations to construct logical features and FFNs to learn their combinations. This approach has a clear logical motivation and an accompanying expressivity analysis.
  2. NDLM performs well in action-model learning and demonstrates generalization across instance sizes. Its quadratic activation memory is a practical advantage over NLMs that explicitly maintain ternary tensors.
Critical weaknesses
  1. Some key notations lack references, such as the interpretation and closed-world assumptions, etc.
  2. The Application in Planning section can be merged with Related Work or Experiment.
  3. The distinct contribution over prior work remains unclear. Methods such as PPGN also enable 3-wl expressiveness relational reasoning with quadratic space complexity. The additional benefits of NDLM’s logic-informed structure need to be explained more clearly.
  4. The expressivity is still limited to a three-variable fragment of first-order logic with predicates of arity at most two.
Questions for authors
  1. Section 6 assumes that pointwise FFNs can implement the required Boolean functions exactly, whereas the experiments use relatively simple single-layer FFNs. Could the authors clarify the relationship between the theory and the implemented configuration, especially any requirements on activation functions, width, and depth?
  2. Since PPGN and related methods also support higher-order relational reasoning with quadratic storage, could the authors clarify NDLM’s distinctive inductive bias and the settings in which it is beneficial? More direct comparisons of performance, memory, and runtime would help assess the significance of the contribution.
  3. Moreover, can the authors compare the results with Relational Networks [1] and Slot Attention [2] encoders? Besides, can the authors explain the difference in expressiveness of the proposed methods with the above two models?
  4. Some general-policy experiments use an additional transitive-closure module. Could the authors distinguish the capabilities of the base NDLM from the extended version and explain how transitive closure affects computational cost and the reported results?

Reference: [1] Santoro A, Raposo D, Barrett DG, Malinowski M, Pascanu R, Battaglia P, Lillicrap T. A simple neural network module for relational reasoning. Advances in neural information processing systems. 2017;30. [2] Locatello F, Weissenborn D, Unterthiner T, Mahendran A, Heigold G, Uszkoreit J, Dosovitskiy A, Kipf T. Object-centric learning with slot attention. Advances in neural information processing systems. 2020;33:11525-38.

1 comment

Thank you for your constructive feedback. The restriction to predicates of arity at most two allows the architecture to use quadratic memory, and the limited expressivity of FO3FO^3 helps the network generalize to instance sizes not seen during training (see our answer to Q2). We added the references for the notation (Section 2), an explanation of how the expressivity result relates to the implemented FFNs (Appendix E.2), controlled comparisons with PPGN (Appendix C), resource comparisons (Appendix B.4.1 and B.5.1), and Relation Networks as a baseline (Tables 2 and 3). Below, we describe these additions and address your questions.

Q1

Thank you for raising this point. We added an appendix subsection (Appendix E.2) explaining how the exact Boolean-FFN assumption simplifies the analysis and how the construction transfers to our single-layer FFNs. Sigmoid FFNs can approximate elementary Boolean connectives arbitrarily closely. Pairwise multiplications can increase approximation errors, while min/max aggregations do not amplify them; suitably parameterized sigmoid gates restore the chosen error bound after each layer. Thus, every fixed formula can be approximated arbitrarily closely. Identity FFNs, together with the nonlinear Stage 1 operators, realize Boolean operations exactly on domains with at least two objects. These constructions require channels for intermediate predicates and may add layers, so the precise depth bound remains specific to the idealized assumption.

Q2

NDLMs aggregate with min and max instead of sums, so they cannot count. This is the inductive bias that distinguishes them from PPGNs, and it is beneficial when the target must generalize to larger instances, as in planning (see our answer to Q4 of Reviewer yH7B for details).

We also added direct comparisons with PPGN. When only irrelevant objects are added at test time (Appendix C.1), PPGN drops to 75% accuracy while NDLM stays at 100%. PPGN also fails to generalize to larger instances on most planning tasks that NDLM solves (Tables 2 and 3), and for some domains even fails to fit the training data. Parameters, epoch times, and peak memory are reported in Appendix B.4.1 and B.5.1.

Q3

We have added Relation Networks as baselines for both action-model and general-policy learning, adapting our inputs to their expected data format while preserving the supplied unary and binary predicates. In action-model learning, they achieved perfect training and test predictions in only 1/10 runs for Blocksworld preconditions and 1/10 runs for Blocksworld effects, and in 0/10 runs for all other domain–task combinations. In general-policy learning, they solved 12.8 ±\pm 0.6/13 training problems in Gripper but only 1.0 ±\pm 1.0/18 test problems, indicating poor generalization to larger instances. In Logistics, they struggled even on the training instances, solving only 0.3 ±\pm 0.5/24 training problems and 0/20 test problems.

Slot Attention learns object representations from input features. This does not exactly fit our predicate learning setup, where the task is to infer additional, potentially more complex, predicates from given predicates. We therefore view it as a potentially complementary representation-learning component, particularly for perceptual inputs.

To the best of our knowledge, neither Relation Networks nor Slot Attention has an exact characterization in terms of a logical language comparable to our characterization of NDLMs in FO≤23FO^3_{\leq 2}.

Q4

Adding transitive closure to the architecture increases the runtime to O(n4)O(n^4). Without the transitive closure it would not be possible to solve Blocksworld, Miconic, and Navigation-xy with the architecture in a generalizing way. In Blocksworld for example, to free a block xx that lies under a tower, the policy must unstack the top block of that tower, i.e., the clear block above xx. “Above” is the transitive closure of onon. Without TC, a network of depth kk can only find this block in towers up to a fixed height (roughly 2k2^k, see Appendix B.7). With TC, it finds it in towers of any height.

Reviewer_8zj8OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces Neural Description Logic Machines (NDLMs), a neurosymbolic model for learning unary and binary predicates from relational data with FO³ expressive power. The model has two stages: a parameter-free stage that generates relational features through operations such as role composition, transposition, and quantification, followed by shared feed-forward networks that learn Boolean combinations. This design achieves quadratic activation memory while preserving FO³ expressivity. The authors apply NDLMs to lifted STRIPS(+)-style action model learning and general policy learning, showing strong results on planning tasks and competitive performance on graph classification benchmarks.

Critical strengths
  1. The paper is well-organized, with a clear task definition (Definition 2.1), an explicit DL grammar, well-defined operators (including generalized min/max matrix products), and a clear complexity analysis.

  2. The method supports three-variable reasoning through staged role composition without using ternary tensors, reducing activation memory to quadratic rather than cubic, as in NLMs and many 3-GNN variants.

  3. The interpretability analysis shows that pruned linear combinations of intermediate DL-like features can produce readable explanations consistent with intuitive STRIPS(+)-style rules.

Critical weaknesses
  1. The method lacks support for counting quantifiers, which are important for many graph tasks and 3-WL problems. This may explain its weaker results on some TU datasets and limit its use in complex planning domains.

  2. The planning experiments exclude recent methods such as DIAS and SYNTH+, which may provide stronger AML baselines under sparse or implicit arguments.

  3. Relevant methods, including Differentiable Logic Machines (DLM, 2021), DIAS (2026), and SYNTH+ (2026), are neither discussed nor compared, despite their relevance to neurosymbolic reasoning and lifted model learning.

Questions for authors
  1. How are the thresholds for policy rollout determined? Are they set separately for each domain or instance, or selected through validation? To what extent do different threshold choices affect the success rates?

  2. Could the authors conduct ablation studies by individually removing the Stage-1 operators (e.g., RC, RR, and CE) to better understand their respective contributions to AML and policy learning?

  3. How would replacing the min/max operations with smooth alternatives, such as LogSumExp or softmin, influence optimization stability and final model accuracy?

  4. How consistent are the interpretability findings across different random seeds and domains? Could the authors quantify the fidelity of the pruned formulas by measuring their agreement with the full models?

  5. What is the rationale for excluding DIAS and SYNTH+ from the AML comparisons? If these methods were included, how would the authors expect them to compare in terms of performance and data efficiency?

  6. Could NDLM incorporate lightweight counting features, such as degree-threshold concepts, while preserving its quadratic memory complexity? Are there any theoretical considerations regarding the feasibility of such extensions?

3 comments
AuthorsOpenReview ↗

Answer to Questions (1/2)

Thank you for your detailed and constructive feedback. The absence of counting quantifiers is a design choice that we believe helps NDLMs generalize to instance sizes not seen during training (see Q4 of our response to Reviewer yH7B). In response to your comments, we have added DLMs as a baseline, an evaluation of smooth alternatives to min⁡/max⁡\min/ \max, an expanded pruning analysis across domains and random seeds, and a discussion about the relation to SYNTH+ and DIAS. Below, we report ablation results and address your questions.

Q1

For sigmoid outputs we choose 0.5 as the threshold, for identity activation networks we choose 0. Those were fixed in advance and independent of the domain. As we choose the action with maximum score, all successful runs remain successful when removing the threshold. Rerunning the experiments without the threshold increased performance for EdgeTransformers in Gripper from 14.3 ±\pm 6.1/18 to 15.0 ±\pm 4.9/18. All other results remain unchanged. Higher thresholds were not tested since they cannot increase performance.

Q2

We performed operator ablations for both action-model and general-policy learning, retraining the networks after individually removing each operator. In the tested Blocksworld action-model configurations, removing RQ, RC, or RR preserved perfect test predictions, whereas removing CE reduced performance. CE transfers unary concept information to binary role channels, allowing predicted relations to depend on object properties or action-argument indicators. The limited effect of removing RQ may reflect overlap with the information available through RR.

We additionally evaluated policy ablations on Blocksworld, Gripper, Logistics, Miconic, and Navigation-xy, with three paired seeds (seeds differ from Table 3 and Table 4). Removing CE reduced Miconic test success from 19.67 ±\pm 0.58 to 0/20 problems. On Logistics, removing RC or CE reduced success from 20/20 to 1.00 ±\pm 1.41/20 and 4.50 ±\pm 3.54/20, respectively, across the two completed seeds. Logistics uses no transitive closure, so RC is the only operator that composes binary relations through a third object. Without RC, the architecture is restricted to FO2FO^2 expressivity. Since the required Logistics policy is not expressible in C2C^2, and therefore not in FO2FO^2, removing RC prevents the architecture from representing a general solution. The observed decrease is thus consistent with this expressivity limitation.

Removing RQ reduced Navigation-xy success from 96/96 to 87.67 ±\pm 6.35/96. While Navigation-xy is also known to require reasoning about three variables, removing RC can be offset by the transitive-closure operator, which provides the required reachability relations and supports reasoning beyond two-variable logic. All individual removals preserved 18/18 success on Gripper. These results indicate that the contribution of each operator depends on the domain.

Q3

Smooth alternatives are indeed applicable. We tested a LogSumExp smoothing on one PROTEINS fold with three paired seeds with different temperatures (0.01, 0.1, 0.5). The accuracy using mild smoothing was higher by 2.10 percentage points, while larger temperatures reduced accuracy (see Appendix B.1.1). This establishes feasibility, but not a general advantage.

Q4

We added a table in Section 8 reporting pruning rates and held-out whole-example agreement across ten training seeds per domain. Depending on the domain, pruning removed an average of 93.92%–98.77% of parameters. Mean whole-example agreement with the original models was 99.73% for Blocksworld, 92.67% for n-Puzzle, and 100% for Delivery and Logistics. Perfect agreement on every evaluated test example was maintained in 8/10, 5/10, 10/10, and 10/10 seeds, respectively.

AuthorsOpenReview ↗

Answer to Questions (2/2)

Q5

Thank you for pointing out the recent work on SYNTH+ and DIAS, which address closely related model-learning tasks in planning. SYNTH+ extends SYNTH with predicate invention based on SIFT. When all predicates are observed, it is equivalent to SYNTH. SYNTH and DIAS learn symbolic STRIPS(+) action schemas from state transitions. NDLMs instead learn to predict the applicability of each action and the denotation of each domain predicate after the action, using positive and negative examples drawn from the transitions. The two tasks are closely related. On the domains in our paper, we expect both methods to learn the correct domain, as NDLMs do. Regarding data efficiency, we expect the symbolic approach (SYNTH) to need fewer examples than the neural approaches (DIAS and NDLMs), although these are sample-efficient too.

The methods differ in the preconditions they can capture. SYNTH is limited to stratified preconditions, which can be expressed in FO2FO^2, since the implicit arguments are bound one at a time. NDLMs can represent preconditions in FO≤23FO^3_{\leq 2}. Moreover, not every FO2FO^2 precondition is stratified. For example, an action that moves an agent made up of two yellow cells xx and yy, one above the other, has the precondition yellow(x)∧yellow(y)∧above(x,y)yellow(x) \wedge yellow(y) \wedge above(x, y). NDLMs can express this precondition, but it is not stratified if the grid contains other, isolated yellow cells, so SYNTH cannot learn it when no action argument is given.

DIAS infers the implicit action arguments with a message-passing GNN. Such GNNs are limited by C2C^2 (FO2FO^2 with counting) and cannot properly recognize preconditions like p(x,y),r(y,z),t(x,z)p(x,y), r(y,z), t(x,z) when xx, yy, and zz are implicit (determined) action arguments. Such preconditions are in FO3FO^3, so NDLMs can express them.

We will add experiments with SYNTH and DIAS to the paper.

Q6

Yes, this is a nice idea. It can be realized by adding operators that work like the current ones but select the k-th largest and k-th smallest value instead of the maximum and minimum. When adding a constant number of these 'threshold' computations, the memory will stay quadratic. While it is unclear whether adding these features will improve results on, for example, the TUDataset, it will boost the expressivity. The network is then expected to be more expressive than FO3FO^3 but still less expressive than C3C^3 since only a constant number of thresholds are added and therefore not all thresholds can be expressed.

Reviewer_8zj8OpenReview ↗

For authors

Thank you to the authors for addressing my concerns. I maintain my original positive score.

Reviewer_yH7BOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Neural Description Logic Machines (NDLMs), a neural architecture for learning unary and binary predicates from relational structures. The architecture is based on description logic and combines fixed relational operations with learnable predicate transformations.

The main technical contribution is that NDLMs can represent the three-variable fragment of first-order logic while requiring quadratic rather than cubic activation memory. This is an important difference from Neural Logic Machines (NLMs), which explicitly maintain higher-arity predicate tensors.

The authors evaluate NDLMs on graph classification, concept learning, action model learning, and general policy learning. The results are particularly strong on planning-related predicate learning tasks, where NDLMs achieve good generalisation to larger problem instances.

Overall, I find this a technically sound and interesting paper. My main question concerns whether the proposed architecture also provides better parameter efficiency and learning efficiency compared with NLMs, beyond its established memory advantage.

Critical strengths
  1. The paper presents a clear architectural contribution. Using description-logic operations to avoid explicitly constructing ternary predicates is an elegant way to reduce memory requirements while retaining the intended logical expressivity.

  2. The theoretical analysis is a strength. The expressivity result provides a useful explanation of what the architecture can represent, rather than relying only on empirical performance.

  3. The planning experiments are encouraging. In particular, the results on action model learning show that NDLMs can learn relational predicates that generalise to instances with different numbers of objects.

  4. The memory comparison with NLMs is convincing and demonstrates a practical advantage of the proposed architecture. The inclusion of additional graph-learning and symbolic concept-learning benchmarks is also useful.

Critical weaknesses

1. The parameter efficiency and learning advantage over NLMs are not sufficiently clear

This is my main concern.

The paper establishes that NDLMs can achieve quadratic activation memory, compared with the cubic memory required by NLMs when explicitly representing ternary predicates. I consider this a meaningful contribution.

However, lower activation memory does not necessarily imply fewer trainable parameters or better learning efficiency. It is therefore important to understand whether the proposed architecture has further advantages beyond memory scaling.

In particular, Appendix B.4, Table 15 shows that NDLM and NLM use different configurations in the action-model experiments. NDLM uses three layers for effect learning and four layers for precondition learning, whereas NLM uses six layers for both. Their hidden widths and activation settings also differ.

Different architectures do not need to use identical configurations. However, the paper does not sufficiently explain how these configurations were selected, or whether the reported differences in learning success are affected by model capacity and hyperparameter choices.

I would appreciate a more direct comparison of trainable parameter counts, training time, and prediction performance. A parameter–performance curve across different depths and widths would be particularly informative.

Such a comparison would help establish whether NDLM's relational operations and inductive bias also improve learning efficiency, or whether the main practical gain is primarily reduced activation memory. Either outcome would be interesting, but the distinction should be made clear.

2. The contribution should be positioned more precisely relative to existing relational architectures

NDLM is a meaningful improvement over NLM in terms of memory complexity. However, quadratic-memory architectures capable of relational composition already exist, such as PPGNs.

The paper discusses these approaches, but I think the particular advantage of using description-logic operators could be explained more explicitly.

For example, it would be useful to clarify whether the main benefit is the logical inductive bias, more efficient optimisation, better generalisation from limited examples, or a combination of these factors.

The experiments suggest that the inductive bias is useful for planning-related predicate learning, although the results on general graph classification are mixed.

3. The scope of the planning contribution could be stated more carefully

The planning applications are interesting, but they assume that states are already represented using a predefined set of relational predicates.

The task is therefore closer to traditional ILP or description-logic concept learning than to learning symbolic abstractions directly from raw observations.

This is not a weakness of the experimental setup itself. The paper defines its predicate-learning task clearly, and the proposed architecture is intended for this setting.

Nevertheless, I think the conclusions should emphasise that the contribution is an efficient neural architecture for learning predicates over existing relational representations, rather than the discovery of new relational representations.

Questions for authors
  1. Could the authors report the trainable parameter counts of NDLM and NLM for the main planning experiments, alongside peak memory usage, training time, and test performance?

  2. How were the network depths and widths in Appendix B.4, Table 15 selected? Were comparable hyperparameter search budgets used for both architectures?

  3. Does NDLM achieve comparable or better learning performance with fewer trainable parameters? A parameter–performance comparison across several configurations would help clarify the practical advantage of the architecture.

  4. Beyond the reduction in activation memory, which benefits are specifically attributable to the description-logic operations? A clearer discussion of the difference from NLM and PPGN would strengthen the positioning of the contribution.

2 comments

Thank you for your detailed and constructive feedback. We agree that parameter efficiency and learning efficiency should be assessed separately from activation-memory requirements. We have added resource comparisons to the paper and conducted a depth–width sweep, whose results we report below. The memory advantage applies during both training and inference, enabling NDLMs to handle substantially larger instances on the same hardware (Figure 4, Appendix B.6). We have also expanded the comparison with PPGN through additional experiments and discussion, and clarified in the conclusion that our approach learns predicates over existing symbolic representations. Below, we describe these additions and address your questions.

Q1

Thank you for this suggestion. We have added resource comparisons across all four action-model domains to Appendix B.4.1 ("Parameters, Epoch Times, and Memory"), covering NDLM, NLM, PPGN, EdgeTransformer and 2-GNN. The tables report trainable parameters, training and evaluation epoch times, and peak allocated memory.

In the reported experiments the selected NDLM configurations have more trainable parameters than NLM and take longer per training epoch, while using less peak allocated memory. For example, on Delivery effects, NDLM/NLM use 25,034/11,954 parameters, 75.03/102.68 MiB of peak training memory, and 3.69/2.72 seconds per epoch. These configurations were not optimized for parameter efficiency (see Q3). The advantage we establish concerns activation memory as the number of objects grows; it does not imply fewer parameters or faster training.

Q2

For both architectures we ran only a small hyperparameter search, sufficient for both architectures to fit the training data for preconditions and effects. For a comparison of the parameter efficiency of both architectures across depths and widths, please see our answer to Q3.

From a theoretical point of view, NLMs may need up to twice the number of layers to learn the same function as an NDLM, which explains why NLMs use more layers than NDLMs. This is because an NDLM combines two binary predicates in one layer, and an NLM needs two, one that combines the predicates into a ternary and one that reduces this to a binary. For this theoretical argument we have also added experimental support (Appendix B.7).

Q3

We investigated parameter efficiency through an architecture sweep.
For Blocksworld effects learning, we tested widths {1, 2, 4, 6, 8, 10} and depths {1, 2, 3, 4, 6}, using five random seeds. The smallest tested models achieving zero training and test errors across all five seeds required 1,664 parameters for NDLM, 3,066 for NLM with maximum arity 2, and 2,956 for NLM with maximum arity 3. Thus, NDLM achieved the same perfect prediction performance with approximately 44–46% fewer parameters.

These results suggest some parameter-efficiency advantage for NDLM in the tested settings. We will add this experiment to the paper, extended to the other domains, precondition learning, and policy learning.

Q4

Two key differences from PPGN are that NDLMs generalize better to larger instances, as seen in the planning experiments and in a new experiment (Appendix C.1), and that a learned NDLM corresponds to a logical formula, which we can partly extract (Section 8).

The logical inductive bias leads to better generalization from limited examples. For NDLMs we have an exact characterization of expressivity (Theorem 6.1): every thresholded NDLM computes a formula of D≡FO≤23\mathcal{D} \equiv FO^3_{\leq 2}, and every formula of depth dd is computed by an NDLM of depth at most 2d2d. Moreover, since min⁡\min and max⁡\max ignore multiplicities, a thresholded NDLM of depth kk can only compute finitely many functions. Their number depends on kk and the input predicates, not on the width, the number of parameters, or the size of the instance (follows from Proposition E.2). A standard Occam/PAC bound for finite hypothesis classes then gives a sample-complexity bound. The bound is far too large to be practical, but it gives an intuition for why NDLMs generalize.

PPGNs' expressivity is lower-bounded by C3C^3, but the computed functions may not be C3C^3-definable. With sum aggregation they can, for example, compare two unbounded counts, which no C3C^3 formula can. Already the C3C^3 formulas of bounded depth form an infinite class, one for each counting threshold. This can help on graph classification tasks (Table 1), but larger planning instances contain counts never seen during training, which hinders generalization.

You are right that the main difference between NDLMs and NLMs is that NDLMs need less activation memory to be able to express three-variable formulas. Also, as mentioned in Q2, NLMs need up to twice as many layers to compute the same function as NDLMs.

Thanks for the rebuttal, and I'll keep my positive rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.