acceptodds
Under review as a conference paper at ICLR 2027

Neurosymbolic Learning of Interpretable World Models

Abstract

World models aim to learn compact representations of an environment from observations, enabling agents to predict future states and plan their behaviour. However, most existing approaches learn black-box representations, with limited ability to incorporate prior knowledge about the world. We introduce Concept-Grounded World Models (CGWMs), a neurosymbolic world model design that instead learns interpretable concept representations. We learn these concepts from observations using explicit and complete knowledge, expressed in temporal logic, that specifies the valid world states and the temporal relations governing their dynamics. Furthermore, we directly use this knowledge by formulating planning as differentiable reasoning, optimising the agent’s actions by gradient ascent to satisfy a goal. We experimentally compare CGWM with a purely neural baseline and a neurosymbolic baseline on two game environments. We find that CGWMs ground concepts much more accurately, even over long horizons, resulting in significantly more robust planning both in and out of distribution.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
neurosymbolic, world models, temporal logic, fuzzy logics, planning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_5qgCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

World models let an agent predict how its environment will change and plan ahead. Most current ones are purely neural and represent the state as an uninterpretable vector, which makes them hard to inspect and hard to combine with what we already know about the world. The paper places itself in neurosymbolic learning, which combines neural networks with logical rules. It proposes Concept-Grounded World Models (CGWMs), in which the state is a set of named, human-readable facts such as "the agent is in this cell" or "this cell is a wall". A neural network learns to read these facts from images. How the world changes is not learned: it is written by hand as logical rules about how the facts evolve over time.

The network is trained with very little labelled data. It sees only a few yes/no labels per episode, such as "the agent eventually reached the goal", and must otherwise make its readings consistent with the hand-written rules. To make this trainable, the rules are scored with a soft, differentiable form of logic, and two training tricks help avoid degenerate solutions. Planning uses the same soft logic: a sequence of actions is adjusted by gradient ascent until the rules predict that the goal is reached. On two small grid-world puzzles, the method reads the facts almost perfectly, even with only 50 training episodes, and stays accurate over long imagined rollouts. It also plans much more successfully than a purely neural world model, including on a new layout and on a variant whose rules were rewritten by hand. A second finding is that different soft logics work best for learning and for planning.

Critical strengths
  1. Concepts are recognised almost perfectly from a few labels per episode, and the ablations show which training choices make this work. Trained with the product form of fuzzy logic, CGWM recognises the concepts with F1 of 0.98–1.00, even from 50 training trajectories (Tables 6–7). Both training tricks are needed. In Two Rooms with 3200 trajectories, F1 drops to 0.21 without blocking gradients through the "if" part of each rule, and to 0.51 without the entropy term (Table 6). Training with logic-based losses often gets stuck in solutions that satisfy the rules trivially, for example by predicting that the agent is in no cell. This is concrete guidance for anyone training models this way.
  2. The best fuzzy logic for learning differs from the best one for planning, and one model can use both (RQ4). With the min/max (Gödel) form of fuzzy logic the encoder fails to learn (F1 about 0.2–0.4, Tables 6–7), while the product form succeeds. For planning, the noisy min/max variant solves more instances than the product form, for example 88.4% versus 68.6% on Two Rooms (lava dense) (Tables 1, 8). The same trained encoder and the same rules serve both phases, so the logic can be chosen separately for each. The finding is clearly supported and useful to others.
Critical weaknesses
  1. CGWM's advantage is largely guaranteed by the setup. CGWM is given the environment's complete and exact transition rules (Tables 4–5). Once it reads the first frame correctly, its predictions must stay correct at every later step, as Sec. 5 itself notes, so the long-horizon accuracy result (RQ2) does not test the method. On the first frame alone, a linear probe on the neural baseline (LeWM) is as accurate as CGWM, even with 50 training trajectories (F1 = 1.00, Tables 6–7). The data-efficiency advantage claimed for RQ1 therefore comes entirely from CGWM not having to learn the dynamics. Likewise, the "zero-shot transfer" to the periodic-door variant (Table 1) consists of a person writing new rules (Table 5), while LeWM gets no new data or retraining. LeWM's planner was also replaced: its actions are optimised by gradient descent over the four discrete moves instead of by its original sampling-based planner (App. D.2). It then solves only 37.8% of Two Rooms (medium) instances, which suggests the replacement planner, not LeWM's representation, may be what limits it. The claims of "significantly more robust planning" and better data efficiency therefore mostly measure the knowledge handed to CGWM, not what it learns.
  2. There is no comparison with a standard search-based planner, which would solve every instance in this setting. The dynamics are known, deterministic and fully observable. Reading the first frame and then running breadth-first search over the rules would find an optimal plan for every instance; the authors themselves use breadth-first search to compute the shortest plans (App. D.3). CGWM's gradient-based planner still fails on 3.7–31.4% of instances in the harder Two Rooms variants (Table 1, Fig. 6). Its only justification is a future extension to continuous variables (Sec. 3.4, 7), which is not tested. Separately, the encoder classifies each grid cell from that cell's 8×8 pixels alone, and each object is a fixed, distinct sprite (App. D.1), so recognising concepts here is an easy problem. There is no experiment with varied visuals, objects spanning several cells, or incomplete rules, so the paper does not show that the approach works beyond these grid puzzles. This matters for the paper's overall argument, because the introduction motivates the work by robotics and AI in the physical world.
Questions for authors
  1. (Weakness 1: given versus learned dynamics) Can you add a variant of CGWM that keeps the same encoder, labels and constraints but learns its transition model from data instead of receiving the rules? Suppose this variant still recognises concepts over long rollouts and plans clearly better than the neural baseline. Then the advantage would come from the logic-based supervision rather than from the hand-written dynamics, and the first weakness would largely fall away.
  2. (Weakness 1: modified baseline planner) How does the neural baseline (LeWM) perform on Two Rooms with its original sampling-based planner instead of the gradient-based one used here? If its success rate rises well above the reported 22–38% on the medium and hard variants, much of the reported planning gap reflects the change to the baseline rather than a difference between the world models.
  3. (Weakness 2: no search-based planner) How does CGWM's gradient-based planner compare with reading the first frame and then running breadth-first search over the same rules? If search solves every instance, what does gradient-based planning offer in this paper's setting? A concrete case where search does not apply but the gradient planner works would answer this.
  4. (Weakness 2: dependence on complete hand-written knowledge) Does grounding still work when the knowledge is reduced or imperfect? Examples: only the transition rules and per-episode labels without the extra constraints (Tables 3–4), one transition rule removed or wrong, or cells whose appearance varies. If grounding collapses in these cases, the method needs a complete, hand-written simulator of the environment. That would undercut the "weak supervision" claim and the motivation from real-world settings.
9 comments

We thank reviewer 5qgC for the careful assessment and constructive suggestions. We have conducted additional experiments to test learning and planning when the explicitly provided knowledge is incomplete and incorrect. We address each question below.

Answering question 1.
Yes, CGWM can still ground and plan well under incomplete knowledge. 

The reviewer suggests learning the transition dynamics from data. As a step in this direction, we introduced two variants of the Two Rooms (hard) setting, maintaining the concept vocabulary while introducing transitions that are absent from, or contradict, the explicitly provided transition knowledge:

  • Wall-Jump: moving towards the central wall makes the agent jump over it, instead of remaining stationary.
  • Lava-Door: lava does not kill the agent, stepping on lava opens the door at the next step, and pressing the button has no effect.

In both environments we retain the original symbolic predictor, whose rules state that walls block the movement, lava is lethal, and button opens the door. These rules represent a reasonable initial model based on common-sense knowledge, but now make incorrect predictions about the transitions.

We add a residual predictor that corrects the symbolic predictor’s output and we train it on trajectories collected from each modified environment. We consider two settings: in the first, we freeze the encoder pretrained in the original Two Rooms (hard) environment and train only the residual predictor; in the second, we train the encoder and residual predictor jointly from scratch. The original symbolic rules remain unchanged in both settings. The target of the residual predictor is the encoding of the subsequent observation. The new transition knowledge and corrections to the incorrect rules are thus learned from data rather than provided explicitly. The corrected next state prediction is
z^t+1=Pred(zt,at)+δt,\hat{z}_{t+1} =\mathrm{Pred}(z_t,a_t) + \delta_t, where Pred(zt,at)\mathrm{Pred}(z_t,a_t) denotes the symbolic predictor’s output and δt\delta_t the correction provided by the residual predictor. During planning, both the encoder and residual predictor are frozen, and the actions are optimized with the same differentiable procedure. We obtain the following results:

EnvironmentOriginal symbolic predictor (with incorrect knowledge)Symbolic predictor + residual predictor, pretrained frozen encoderSymbolic predictor + residual predictor, jointly trained encoder
Wall-Jump59.3 (5.3)99.5 (0.7)73.9 (3.0)
Lava-Door1.3 (0.9)97.1 (1.5)95.6 (2.5)

Values are the average percentage of solved instances over 10 seeds, with standard deviation in parentheses. 

The high success rates obtained with the pretrained frozen encoder should also be interpreted in light of the changed tasks. Once the new transitions have been learned, jumping over the wall and safely stepping on lava can shorten the routes to the goal, and simplify planning compared to the original Two Rooms (hard) setting.
With the encoder jointly trained from scratch, planning success remains high on Lava-Door but is lower on Wall-Jump. The reason can be attributed to the grounding performance, which will be discussed in answering question 4. The Wall-Jump model grounds walls imperfectly, whereas the concept not grounded in Lava-Door is the button, which has no effect in the modified environment. These experiments show that CGWM can learn to correct incorrect transition knowledge and use the corrected predictions for planning without manually rewriting its symbolic rules.

Answering question 2. The CEM Planner for LeWM usually performs worse. We had evaluated the LeWM baseline with the CEM optimization algorithm before submission. It attains low planning performance on most settings. We therefore evaluated the spatial LeWM variant with reward-based differentiable planning described in the paper.

The table below reports the planning performance of LeWM with CEM using both its original vector representation and the spatial representation used in our baseline. We include the spatial variant with gradient-based planning for comparison:

EnvironmentVector LeWM + CEMSpatial LeWM + CEMSpatial LeWM + gradient planning
Two Rooms (easy)81.3 (3.7)88.0 (2.5)95.8 (1.2)
Two Rooms (medium)0.0 (0.0)0.0 (0.0)37.8 (2.8)
Two Rooms (hard)1.1 (1.0)30.6 (6.4)22.0 (9.0)
Two Rooms (lava dense)0.0 (0.0)6.4 (3.2)9.6 (6.9)
Two Rooms (periodic door)0.5 (0.7)9.1 (5.2)7.2 (3.6)
Ice Slider (T=20)0.4 (0.5)50.4 (6.4)69.1 (12.6)
Ice Slider (T=T*)0.0 (0.0)0.2 (0.4)3.6 (2.9)

Values are the mean percentage of solved instances over 10 seeds, with standard deviation in parentheses. Our planner outperforms the CEM planner in all but two rows, strengthening the baseline.
We will include the full CEM results in the camera ready.

To address the concern expressed in Weakness 1 about transfer to periodic-door, we also fine-tuned both LeWM variants on trajectories collected from that environment, starting from the model trained on Two Rooms (hard). We report the obtained planning results:

Fine-tuned modelPlannerSolved instances (%)
Spatial LeWMGradient7.9 (3.6)
Spatial LeWMCEM15.8 (4.3)
Vector LeWMCEM0.5 (0.5)

Values are the mean percentage of solved instances over 10 seeds, and standard deviation in parentheses. Fine-tuning improves planning performance for spatial LeWM, but both variants remain substantially below CGWM. We will include these results in the revised paper.

Answering question 3. BFS finds a shortest plan for every solvable instance. Given a correctly decoded initial state and complete, correct, deterministic transition rules, BFS returns a shortest plan for every solvable instance in these environments. Indeed, we already use BFS to compute optimal plan lengths (Appendix D.3). We now report CGWM results with BFS, extending Table 1.

The contribution of our differentiable planning procedure is to demonstrate that CGWM supports learning, prediction, and planning within the same differentiable reasoning framework. Although introduced to learn grounded world models, CGWM naturally supports action optimization through its own transition model: the same differentiable machinery used during learning can propagate goal-satisfaction gradients to action sequences during planning.

This reuse makes it valuable to investigate how the framework behaves across these phases. Our experiments show that the fuzzy semantics that performs best for learning the grounding differs from the one that performs best for planning, providing practical insight into the choice of semantics for differentiable planning. These findings can also inform the design of other differentiable neurosymbolic planners, including those operating over continuous action spaces (which standard BFS cannot directly enumerate).

We will add the BFS results and clarify both CGWM’s compatibility with classical planners and the contribution of demonstrating and studying planning within its differentiable framework in the revised paper.

Two Rooms (easy)Two Rooms (medium)Two Rooms (hard)Two Rooms (lava dense)Two Rooms (periodic door)Ice Slider (T=20)Ice Slider (T=T)*
CGWM (Godel Trick)100.0 (0.0)100.0 (0.0)96.3 (0.8)88.4 (2.8)94.9 (1.0)100.0 (0.0)100.0 (0.0)
CGWM (log-product)100.0 (0.0)100.0 (0.0)90.4 (1.9)68.6 (5.5)84.3 (3.3)99.8 (0.4)90.4 (1.7)
CGWM (BFS)100.0 (0.0)100.0 (0.0)100.0 (0.0)100.0 (0.0)100.0 (0.0)100.0 (0.0)100.0 (0.0)
LeWM95.8 (1.2)37.8 (2.8)22.0 (9.0)9.6 (6.9)7.2 (3.6)69.1 (12.6)3.6 (2.9)
Reviewer_bhZGOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Concept-Grounded World Models (CGWMs), a neurosymbolic framework for learning interpretable world models from visual observations using weak trajectory-level supervision and prior symbolic knowledge. Unlike conventional neural world models, CGWMs represent states explicitly through named concepts and use temporal logic to encode environmental constraints, transition dynamics, and task specifications.

The method has two main components: a neural encoder that maps observations to fuzzy concept assignments, and a non-parametric predictor that applies predefined symbolic transition rules. Training combines three principal objectives (trajectory-specification supervision, constraint satisfaction, and next-state prediction) with an entropy-based regularizer intended to prevent degenerate solutions. The authors also block gradients through the antecedents of selected logical implications to reduce trivial satisfaction of constraints. For planning, the model optimizes action sequences by gradient ascent on the satisfaction of a temporal-logic goal formula.

The authors evaluate CGWMs on Ice Slider and Two Rooms, comparing them with a neural world-model baseline (LeWM) and a neurosymbolic baseline (NeSy-MM). Their results suggest that CGWMs trained with log-product fuzzy semantics can learn highly accurate concept representations from relatively few trajectories and preserve these representations over long imagined rollouts. For planning, the Gödel Trick performs better than log-product semantics. The reported results also show promising transfer to layouts with more hazards and to a variant with different door dynamics, provided the transition rules are updated.

The central contribution is a unified approach that connects visual concept grounding, symbolic dynamics, and goal-directed planning. Its empirical scope, however, is deliberately restricted to deterministic, fully observable grid worlds in which the symbolic dynamics are assumed to be complete and correct.

Critical strengths

S1: A coherent integration of concept grounding, dynamics, and planning

The paper presents a clean architecture in which concepts have predefined semantics and the same symbolic representation is used for learning, prediction, and planning. This avoids treating interpretability as a post-hoc property of an otherwise opaque latent space. Temporal logic provides a common formalism for encoding environmental invariants, transition rules, and task objectives. The separation between a learned perception module and an explicit transition system is particularly appealing: it makes the source of prediction errors easier to diagnose and allows goals to be changed without retraining the encoder.

S2: Strong empirical evidence for long-horizon grounding and data efficiency

The experiments evaluate concept recognition both on observed states and across imagined rollouts, rather than relying only on task success. The reported results show that log-product CGWMs achieve near-perfect concept grounding across training-set sizes ranging from 50 to 3,200 trajectories, with only a modest degradation in the more complex setting at the smallest data scale. The results also support an important architectural argument: when the initial state is grounded correctly and the symbolic transition rules are correct, prediction errors do not accumulate through a learned dynamics model in the same way they do in neural world models. The comparison with NeSy-MM further suggests that explicitly representing concepts is not sufficient by itself; the training objective and optimization strategy matter.

S3: Promising planning performance and transfer without retraining the encoder

The planning results are compelling within the evaluated domains. With the Gödel Trick, CGWM reports 96.3% success on Two Rooms (hard), 88.4% on the denser-lava variant, and 94.9% on the periodic-door variant. By comparison, LeWM reports 22.0%, 9.6%, and 7.2%, respectively. The paper also identifies a useful distinction between the best semantics for learning and planning: log-product semantics are effective for concept grounding, whereas the Gödel Trick is more effective for discrete action optimization. The ability to combine these choices without retraining the encoder is a practical strength. Importantly, the transfer experiments modify the layout or transition rules while preserving the concept vocabulary, making the intended mechanism for generalization explicit.

Critical weaknesses

W1: Strong assumptions limit the scope of the contribution

The approach assumes that the symbolic transition rules are complete and correct, the environment is deterministic, and all relevant concepts can be inferred from a single observation. These assumptions substantially simplify both learning and planning. In this setting, once the encoder has correctly identified the initial state, the symbolic predictor can propagate it using the supplied rules. The experiments therefore demonstrate the value of grounding perception into a known symbolic model, but provide limited evidence that CGWM can serve as a general-purpose world model when the dynamics are unknown, partially specified, stochastic, or partially observable. The paper acknowledges these limitations, but they affect the scope of its central claims about robust world modeling and transfer. It would help to establish how performance degrades when the supplied knowledge is incomplete or slightly incorrect.

W2: The experimental comparison does not fully isolate the source of the gains

CGWM benefits from an explicitly specified transition system, whereas LeWM must learn its dynamics from data. This is a meaningful comparison of two modeling paradigms, but it does not isolate the contribution of the proposed grounding objective from the advantage of having a hand-specified dynamics model. The baseline protocols also differ: LeWM receives frame-level rewards and can re-plan using new observations every five actions, whereas CGWM receives trajectory-level labels and plans from the initial observation. These differences favor LeWM in some respects, but the comparison still leaves open how much of CGWM's advantage comes from symbolic dynamics, the grounding losses, or the optimization procedure. NeSy-MM is a useful comparison for concept grounding, but it is not evaluated on planning. A stronger set of controlled comparisons would make the empirical case for the specific methodological contributions more convincing.

W3: Differentiable planning is effective but not yet fully characterized

The planner optimizes a continuous parameterization of actions using noisy gradient-based optimization and then discretizes the resulting action distributions. This introduces potential sensitivity to initialization, optimization hyperparameters, planning horizon, and local optima. Although the reported success rates are strong, CGWM does not solve every Two Rooms instance, and failures concentrate on instances requiring longer plans. The results therefore leave questions about planning reliability and computational scaling unanswered. Moreover, the reported success rate establishes that a plan reaches the goal in the tested environments; it does not, by itself, establish a general guarantee of safety or optimality. The paper would be stronger with a more systematic analysis of planning runtime, sensitivity to optimization choices, and performance against a classical symbolic planner in these small domains.

Questions for authors

Q1. Robustness to incomplete or incorrect knowledge. How does CGWM perform when some transition rules or constraints are missing, or when a small fraction of the supplied rules are incorrect? Can the authors quantify the degradation in concept-grounding F1 and planning success as the quality of the knowledge specification decreases?

Q2. Attribution of the empirical gains. Can the authors provide controlled ablations that separate the benefit of the symbolic transition system from the benefit of the proposed grounding objective? In particular, how does CGWM compare with a model using the same symbolic predictor but a simpler training objective, and with a neural baseline given equivalent transition knowledge or supervision?

Q3. Planning reliability and computational cost. How sensitive is the differentiable planner to the number of random initializations, optimization steps, noise schedule, and planning horizon? Can the authors report planning runtime and compare against an exact or classical symbolic planner on the same environments, including success and optimality rates?

Q4. Scope of the generalization claim. The transfer experiments preserve the concept vocabulary and rely on updating the transition rules when dynamics change. What happens when a new environment introduces an unseen concept, requires history beyond the current observation, or contains stochastic transitions? Which parts of CGWM would need to change, and is there evidence that the grounding learned in the current experiments transfers to these settings?

6 comments

We thank reviewer bhZG for the careful assessment and constructive suggestions. We have conducted additional experiments on grounding and planning with imperfect transition knowledge, the role of grounding constraints, and planning reliability and computational cost. We address each question below.

Answering question 1 (part 1/2). Yes, CGWM can still ground and plan well under incomplete knowledge.  To address the concerns about the robustness to incomplete or incorrect knowledge, we introduced experiments in which CGWM must learn new transitions and fix incorrect supplied knowledge. These experiments test whether the model can learn to ground symbols and whether it can adapt its predictions and use them for planning without manually rewriting its symbolic rules.
We introduced two variants of the Two Rooms (hard) setting, maintaining the concept vocabulary while changing the environment’s transitions:

  • Wall-Jump: moving towards the central wall makes the agent jump over it, instead of remaining stationary.
  • Lava-Door: lava does not kill the agent, stepping on lava opens the door at the next step, and pressing the button has no effect.

In both environments we retain the original symbolic predictor, whose rules state that walls block the movement, lava is lethal, and button opens the door. These rules represent a reasonable initial model based on common-sense knowledge, but now they make incorrect predictions about the transitions.
We add a residual predictor that corrects the symbolic predictor’s output and we train it on trajectories collected from each modified environment. We consider two settings: in the first, we freeze the encoder pretrained in the original Two Rooms (hard) environment and train only the residual predictor; in the second, we train the encoder and residual predictor jointly from scratch. The target of the residual predictor is the encoding of the subsequent observation. In both settings, the original symbolic rules remain unchanged, so the corrections to the incorrect rules are learned from data rather than provided explicitly. The corrected next state prediction is z^t+1=Pred(zt,at)+δt,\hat{z}_{t+1} =\mathrm{Pred}(z_t,a_t) + \delta_t, where Pred(zt,at)\mathrm{Pred}(z_t,a_t) denotes the symbolic predictor’s output and δt\delta_t the correction provided by the residual predictor.

For grounding, we trained the encoder from scratch, using either only the original symbolic predictor or the original symbolic predictor plus the residual neural predictor. We report the grounding results achieved on the two new settings.

ConceptWall-Jump, symbolicWall-Jump, + residualLava-Door, symbolicLava-Door, + residual
Agent1.00 (0.00)1.00 (0.00)1.00 (0.00)1.00 (0.00)
Goal1.00 (0.00)1.00 (0.00)1.00 (0.00)1.00 (0.00)
Wall0.74 (0.02)0.80 (0.10)0.99 (0.02)0.91 (0.17)
Lava1.00 (0.00)1.00 (0.00)1.00 (0.00)0.80 (0.40)
Button1.00 (0.00)1.00 (0.00)0.30 (0.46)0.10 (0.30)
Door1.00 (0.00)1.00 (0.00)0.80 (0.40)0.90 (0.30)
Average0.96 (0.00)0.97 (0.02)0.85 (0.05)0.79 (0.09)

Values are the average test grounding F1 score over ten seeds, with standard deviation in parentheses.
The grounding errors generally affect the concepts most directly involved in the incorrect transition knowledge. In Wall-Jump, the errors only involve the wall concept, and the other 5 concepts remain perfectly grounded. In Lava-Door, button recognition is particularly affected.

Answering question 1 (part 2/2). For planning, we compare the original symbolic predictor with the two residual settings described above. The world model is kept fixed during planning, and only the actions are optimized with the same differentiable procedure. 

EnvironmentOriginal symbolic predictor (with incorrect knowledge)Symbolic predictor + residual predictor, pretrained frozen encoderSymbolic predictor + residual predictor, jointly trained encoder
Wall-Jump59.3 (5.3)99.5 (0.7)73.9 (3.0)
Lava-Door1.3 (0.9)97.1 (1.5)95.6 (2.5)

Values are the average percentage of solved instances over ten seeds, with standard deviation in parentheses. 

These experiments show that CGWM can learn transition corrections from data and use them for planning both with a pretrained frozen encoder and with an encoder jointly trained from scratch. With the encoder jointly trained from scratch, planning success remains high on Lava-Door but is lower on Wall-Jump. The reason can be attributed to the grounding performance. The Wall-Jump model grounds walls imperfectly, whereas the concept not grounded in Lava-Door is the button, which has no effect in the modified environment.

We will add these experiments to the paper, reporting the effects of incorrect knowledge on grounding and the planning results obtained with learned transition corrections.

Answering question 2.
We understand this question as asking how much of CGWM’s performance depends on the provided symbolic transition. One interpretation of “a simpler training objective” is to remove the constraint loss while retaining the symbolic predictor, specification labels and prediction loss. We have evaluated this variant on Two Rooms (hard). In this setting, we trained the model from scratch on 3200 trajectories using 10 seeds, and evaluated the grounding on 800 test trajectories. We report in the obtained grounding F1 score for each concept, averaging over the seeds, with standard deviation in parentheses.

ModelagentgoalwalllavabuttondoorOverall F1
CGWM (without constraint loss)0.71 (0.45)0.81 (0.38)0.75 (0.40)0.72 (0.43)0.32 (0.45)1.00 (0.00)0.72 (0.31)
CGWM (with constraint loss)1.00 (0.00)1.00 (0.00)1.00 (0.00)1.00 (0.00)1.00 (0.00)1.00 (0.00)1.00 (0.00)

Three runs achieve a perfect F1 score on all six concepts. Thus, grounding can be learned without the constraint loss, but its reliability varies substantially across seeds.

For the second part of the question, we kindly ask the reviewer if they could clarify how they envision “equivalent transition knowledge or supervision” to a neural baseline? A specific baseline would help us address the intended attribution question and assess what is feasible within the rebuttal period.

Answering question 3.
It is mostly sensitive to number of steps and noise scale. We evaluated planning sensitivity on Two Rooms (hard), using the Gödel trick and varying one parameter at a time around the paper configuration (first row of the following table). Since our planner based on the Gödel-trick uses a fixed noise distribution throughout optimization, we evaluate sensitivity to the noise scale rather than to its schedule. 

Number of initializationsNumber of stepsNoise scalePlanning horizonSolved instances (%)Planning time (s/batch)
12830014096.3 (0.8)206.9 (23.9)
6430014095.5 (0.8)130.4 (3.3)
25630014096.9 (0.8)301.4 (2.5)
12815014091.3 (1.4)134.3 (2.1)
12860014097.7 (0.8)320.7 (5.6)
1283000.54098.1 (1.0)195.4 (3.5)
12830024092.3 (1.7)197.6 (4.1)
12830016096.5 (1.0)260.3 (4.6)
12830018096.4 (1.5)376.9 (33.6)

Within these ranges, planning success is more sensitive to the number of optimization steps and noise scale than to the number of initializations or planning horizon.  We will add this sensitivity analysis and the measured runtimes to the revised paper. 

BFS finds a shortest plan for every solvable instance. Regarding the comparison against a classical planner, given a correctly decoded initial state and complete, correct, deterministic transition rules, breadth-first search returns a shortest plan for every solvable instance in these environments. Indeed, we already use BFS to compute optimal plan lengths (Appendix D.3). We now report CGWM results with BFS, extending Table 1.

The contribution of our differentiable planning procedure is to demonstrate that CGWM supports learning, prediction, and planning within the same differentiable reasoning framework. Although introduced to learn grounded world models, CGWM naturally supports action optimization through its own transition model: the same differentiable machinery used during learning can propagate goal-satisfaction gradients to action sequences during planning.

This reuse makes it valuable to investigate how the framework behaves across these phases. Our experiments show that the fuzzy semantics that performs best for learning the grounding differs from the one that performs best for planning, providing practical insight into the choice of semantics for differentiable planning. These findings can also inform the design of other differentiable neurosymbolic planners, including those operating over continuous action spaces (which standard BFS cannot directly enumerate).

We will add the BFS results and clarify both CGWM’s compatibility with classical planners and the contribution of demonstrating and studying planning within its differentiable framework in the revised paper.

Two Rooms (easy)Two Rooms (medium)Two Rooms (hard)Two Rooms (lava dense)Two Rooms (periodic door)Ice Slider (T=20)Ice Slider (T=T)*
CGWM (Godel Trick)100.0 (0.0)100.0 (0.0)96.3 (0.8)88.4 (2.8)94.9 (1.0)100.0 (0.0)100.0 (0.0)
CGWM (log-product)100.0 (0.0)100.0 (0.0)90.4 (1.9)68.6 (5.5)84.3 (3.3)99.8 (0.4)90.4 (1.7)
CGWM (BFS)100.0 (0.0)100.0 (0.0)100.0 (0.0)100.0 (0.0)100.0 (0.0)100.0 (0.0)100.0 (0.0)
LeWM95.8 (1.2)37.8 (2.8)22.0 (9.0)9.6 (6.9)7.2 (3.6)69.1 (12.6)3.6 (2.9)
Reviewer_g5t4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Concept-Grounded World Models. Here, the state is a set of fuzzy truth values for named concepts in each grid cell. A CNN encoder grounds these concepts from images and a fixed set of hand-written LTLf transition rules serves as the predictor. The encoder is trained from trajectory-level labels and LTLf constraints, using stop-gradients on implication antecedents and an entropy term to avoid poor local minima. Planning is done by gradient ascent on the fuzzy truth value of a goal formula. The experiments are conducted on two grid world environments -- ice slider and two rooms. In these, CGWM grounds the concepts almost perfectly at every rollout horizon and solves 88% to 100% of planning instances, compared with 7% to 96% for LeWM.

Critical strengths
  • The comparison with NeSy-MM is fair and convincing. Both methods receive the same knowledge, labels and encoder. Under this comparison, CGWM reaches an all-concept F1 of 0.98 to 1.00 where NeSy-MM reaches 0.46 to 0.75. The ablations and the use of 10 seeds make this result solid.

  • The finding about fuzzy semantics is useful for the neurosymbolic community. Log-product semantics works for learning and the Godel Trick works for planning, and the same trained encoder can be used with both. The paper is also clearly written and fully documented.

Critical weaknesses
  • The main concern is that the experimental setting is very favorable to the method. The dynamics are complete and written by hand, the encoder looks at each grid cell in isolation, and the constraints effectively provide dense supervision at every step. Grounding also depends on stop-gradients placed by hand, and without them F1 drops to about 0.2. As a result, the comparison with LeWM and the periodic-door transfer, where a human rewrites the rules, mostly measure the benefit of being given the dynamics. So, the realism of the environments and the problem setting is a serious limitation of the current work.

  • Also, the paper does not compare against classical planning. With the grounded initial state and the complete, deterministic transition rules, breadth-first search would find an optimal plan for every solvable instance. The differentiable planner on the other hand underperform this for the three Two Rooms environment. The authors motivate gradient-based planning by future hybrid models with continuous variables, but the current experiments show no advantage over classical search in the setting they study.

Questions for authors
  1. Comparing BFS baseline and LeWM with CEM, how would breadth-first search on the decoded CGWM state perform, and how does LeWM perform with its original CEM planner?
  2. On generalization of the method, how did you choose where to place the stop-gradients and how well does the method generalize? Is there a general rule that would carry over to new environments and is it something the methods can learn?
  3. The paper describes the supervision as weak, since only trajectory-level labels are given. However, the constraints in Tables 3 and 4 relate concept values across consecutive steps for every cell, and the ablations show that grounding fails without carefully placed stop-gradients. This suggests that the constraint set does most of the work, more than the labels. Is this reading of the findings correct? Would it be possible to train CGWM on Two Rooms (hard) without the constraint loss, keeping only the specification labels and the prediction loss, and report the resulting F1?
  4. The encoder setting indicates that it sees each grid cell in isolation, so grounding reduces to classifying one sprite per cell. Is it possible to report F1 with an encoder whose receptive field spans neighboring cells? Near-perfect grounding there would show that the results do not depend on this design choice and make it independent of the encoder choices.
3 comments

We thank reviewer g5t4 for the careful assessment and constructive suggestions. We have conducted additional experiments addressing the role of the supplied knowledge, the constraint loss, and the encoder’s receptive field. We address each question below.

Answering question 1. The CEM Planner for LeWM usually performs worse. We had evaluated the LeWM baseline with the CEM optimization algorithm before submission. It attains low planning performance on most settings. We therefore evaluated the spatial LeWM variant with reward-based differentiable planning described in the paper.

The table below reports the planning performance of LeWM with CEM using both its original vector representation and the spatial representation used in our baseline. We include the spatial variant with gradient-based planning for comparison:

EnvironmentVector LeWM + CEMSpatial LeWM + CEMSpatial LeWM + gradient planning
Two Rooms (easy)81.3 (3.7)88.0 (2.5)95.8 (1.2)
Two Rooms (medium)0.0 (0.0)0.0 (0.0)37.8 (2.8)
Two Rooms (hard)1.1 (1.0)30.6 (6.4)22.0 (9.0)
Two Rooms (lava dense)0.0 (0.0)6.4 (3.2)9.6 (6.9)
Two Rooms (periodic door)0.5 (0.7)9.1 (5.2)7.2 (3.6)
Ice Slider (T=20)0.4 (0.5)50.4 (6.4)69.1 (12.6)
Ice Slider (T=T*)0.0 (0.0)0.2 (0.4)3.6 (2.9)

Values are the mean percentage of solved instances over 10 seeds, with standard deviation in parentheses. Our planner outperforms the CEM planner in all but two rows, strengthening the baseline.
We will include the full CEM results in the camera ready.

BFS finds a shortest plan for every solvable instance. Given a correctly decoded initial state and complete, correct, deterministic transition rules, BFS returns a shortest plan for every solvable instance in these environments. Indeed, we already use BFS to compute optimal plan lengths (Appendix D.3). We now report CGWM results with BFS, extending Table 1.

The contribution of our differentiable planning procedure is to demonstrate that CGWM supports learning, prediction, and planning within the same differentiable reasoning framework. Although introduced to learn grounded world models, CGWM naturally supports action optimization through its own transition model: the same differentiable machinery used during learning can propagate goal-satisfaction gradients to action sequences during planning.

This reuse makes it valuable to investigate how the framework behaves across these phases. Our experiments show that the fuzzy semantics that performs best for learning the grounding differs from the one that performs best for planning, providing practical insight into the choice of semantics for differentiable planning. These findings can also inform the design of other differentiable neurosymbolic planners, including those operating over continuous action spaces (which standard BFS cannot directly enumerate).

We will add the BFS results and clarify both CGWM’s compatibility with classical planners and the contribution of demonstrating and studying planning within its differentiable framework in the revised paper.

Two Rooms (easy)Two Rooms (medium)Two Rooms (hard)Two Rooms (lava dense)Two Rooms (periodic door)Ice Slider (T=20)Ice Slider (T=T)*
CGWM (Godel Trick)100.0 (0.0)100.0 (0.0)96.3 (0.8)88.4 (2.8)94.9 (1.0)100.0 (0.0)100.0 (0.0)
CGWM (log-product)100.0 (0.0)100.0 (0.0)90.4 (1.9)68.6 (5.5)84.3 (3.3)99.8 (0.4)90.4 (1.7)
CGWM (BFS)100.0 (0.0)100.0 (0.0)100.0 (0.0)100.0 (0.0)100.0 (0.0)100.0 (0.0)100.0 (0.0)
LeWM95.8 (1.2)37.8 (2.8)22.0 (9.0)9.6 (6.9)7.2 (3.6)69.1 (12.6)3.6 (2.9)

Answering question 2. We always place stop-gradients on the concepts in the antecedent of the implication.
While learning the correct grounding of the concepts, the current confidence in the antecedent’s concepts provides evidence for learning the concepts in the consequent.
The choice of placing stop-gradients on the antecedent’s concepts prevents specific optimization shortcuts. A fuzzy implication can become more satisfied either by increasing the consequent’s truth value or by decreasing the antecedent’s. The latter allows the model to lower the evidence on the antecedent instead of learning the consequent [1].
The placement of stop-gradients therefore is independent of particular concepts or environments. We apply this same criterion in both Two Rooms and Ice Slider, and it extends directly to implication constraints in new domains.
We will expand the explanation of the stop-gradient placement in Section 3.3 to make the general criterion and its role in preventing optimization shortcuts more explicit.

1: Analyzing Differentiable Fuzzy Logic Operators, Emile van Krieken, Erman Acar, and Frank van Harmelen, AIJ 2022

Answering question 3. Yes, this is possible, but it is much less reliable. We conducted the requested experiments on Two Rooms (hard), removing the constraint loss while keeping the specification labels and the prediction loss. We trained the model from scratch on 3200 trajectories using 10 seeds, and evaluated the grounding on 800 test trajectories. We report the obtained grounding F1 score for each concept, averaging over the seeds, with standard deviation in parentheses.

ModelagentgoalwalllavabuttondoorOverall F1
CGWM (without constraint loss)0.71 (0.45)0.81 (0.38)0.75 (0.40)0.72 (0.43)0.32 (0.45)1.00 (0.00)0.72 (0.31)
CGWM (with constraint loss)1.00 (0.00)1.00 (0.00)1.00 (0.00)1.00 (0.00)1.00 (0.00)1.00 (0.00)1.00 (0.00)

Correct grounding is possible without the constraint loss, as three runs achieve a perfect F1 score on all six concepts. The large variability however shows the learning is substantially less reliable, suggesting the presence of reasoning shortcuts [1] without the constraint loss.
We will add this experiment and clarify the role of the constraints in the revised paper.

[1] Marconato, Emanuele, et al. "Symbol grounding in neuro-symbolic ai: A gentle introduction to reasoning shortcuts." Journal of Artificial Intelligence Research 86 (2026).

Answering question 4. The encoder also works if it sees neighbouring cells. We replaced the 1x1 convolution in the last layer of the encoder with a 3x3 convolution, so that each prediction receives features also from the neighboring cells. We trained this encoder from scratch on the Two Rooms (hard) and Ice Slider environments on 3200 trajectories using 10 seeds, and we evaluated the grounding on 800 test trajectories. 

Two Rooms (hard):

EncoderagentgoalwalllavabuttondoorOverall F1
3x3 convolution1.00 (0.00)1.00 (0.00)0.99 (0.02)1.00 (0.00)0.99 (0.05)0.99 (0.04)0.99 (0.08)
1x1 convolution (original)1.00 (0.00)1.00 (0.00)1.00 (0.00)1.00 (0.00)1.00 (0.00)1.00 (0.00)1.00 (0.00)

Ice Slider:

EncoderagentgoalrockOverall F1
3x3 convolution1.00 (0.00)1.00 (0.00)1.00 (0.00)1.00 (0.00)
1x1 convolution (original)1.00 (0.00)1.00 (0.00)1.00 (0.00)1.00 (0.00)

Values are means over 10 training seeds, with population standard deviation in parentheses. These results show that near-perfect grounding can also be obtained when the encoder receives information from neighboring cells. We will add these results to the revised paper.

Dear Reviewer g5t4,

Thank you again for the insightful review. Following your questions, the revised paper (changes in blue) includes breadth-first search on the learned world models, LeWM with its original CEM planner, the ablation without the constraint loss, and an encoder that also sees the neighbouring cells. We also added experiments in which the transition knowledge given to CGWM is incorrect, which directly address your main concern about hand-written dynamics.

We would be grateful to know whether our responses address your concerns, and we are happy to provide any further clarification. If they do, we would appreciate it if you could take them into account in your assessment.

Best regards, The Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response to all Reviewers and the Area Chair

We thank the reviewers for their careful assessment and constructive questions. These questions led us to new experiments, most notably on learning and planning with incorrect transition knowledge. The new results confirm and extend the original findings, and none of them changes the conclusions of the paper. All changes are highlighted in blue in the revised paper.

Strengths recognised by the reviewers. All three reviewers found our results on fuzzy semantics useful beyond this work: log-product semantics grounds the concepts during learning, the Gödel Trick is more effective for planning, and the same trained model supports both (g5t4, bhZG, 5qgC). The reviewers also highlighted the near-perfect grounding from weak supervision and few trajectories (bhZG, 5qgC), the ablations showing which training choices make it work (g5t4, 5qgC), and the fair and convincing comparison with NeSy-MM (g5t4).

All reviewers rated their soundness concerns as minor or fixable. We addressed each of them with new experiments, summarised below.

1. The advantage of CGWM does not come only from the given knowledge. The neurosymbolic baseline NeSy-MM receives the same concepts, rules, constraints, labels and encoder architecture as CGWM, yet grounds only part of the concepts. As reviewer bhZG notes, this indicates that the training objective and optimisation strategy matter and that explicit concepts are not sufficient by themselves.
The new experiments further show that CGWM does not require the given knowledge to be correct:

  • Incorrect knowledge (bhZG, 5qgC). We introduced two variants of Two Rooms, Wall-Jump and Lava-Door, in which the rules given to CGWM contradict the true dynamics. A small residual predictor, learned from trajectories, corrects the symbolic predictions without rewriting any rule. With the residual predictor and the encoder trained on the original Two Rooms, CGWM solves 99.5% of the Wall-Jump instances and 97.1% of the Lava-Door instances.
  • No constraint loss (g5t4, bhZG). Without the constraint loss, CGWM can still ground all concepts, but much less reliably across seeds, indicating that the constraint loss helps rule out reasoning shortcuts.

2. With breadth-first search, CGWM solves every instance with a shortest plan (g5t4, bhZG, 5qgC). We added BFS on the learned world models to Table 1. BFS solves all instances in every setting. This confirms that the learned world models are accurate on every test instance, and that the remaining failures of the gradient-based planner come from the optimisation rather than from the world model. The revised paper presents CGWM as directly compatible with classical search, and differentiable planning as part of a single differentiable framework shared by learning, prediction and planning, which is what enables our study of fuzzy semantics across learning and planning.

3. The original LeWM planner does not close the gap (g5t4, 5qgC). CEM, the planner of the original LeWM, gives lower success than our gradient-based planner in most settings, both for the original LeWM and for the spatial variant used in the paper. Even after fine-tuning on Two Rooms (periodic door), LeWM remains far below CGWM. As reviewer bhZG notes, the protocol already favours LeWM, which receives frame-level rewards and re-plans every five actions.

4. Design choices.

  • Encoder (g5t4, 5qgC). We additionally implemented an encoder that also sees the neighbouring cells and it grounds near perfectly all concepts in both environments.
  • Stop-gradients (g5t4). Their placement follows a general criterion that does not depend on particular concepts or environments.
  • Planner (bhZG). Planning success is robust to the number of plans, optimisation steps, noise scale and horizon, and some variations even improve over the configuration used in the paper. We also report the planning runtimes.

Where to find the changes in the revised paper

ConcernRevised paper
Incorrect transition knowledgeSec 5 (lines 440-464), App B.3, App C.4.
No constraint lossSec 5 (lines 370-371), App C.1 (lines 1129-1133).
Classical planning (BFS)Sec 3.4 (lines 234-238), Table 1, Section 5 (lines 422-424).
LeWM with CEM and fine-tuningSec 4 (lines 325-326), Sec 5 (lines 411-412), App C.5.
Encoder with neighbouring cellsApp C.3.
Stop-gradient placementSec 3.3 (lines 214-215).
Planner sensitivity and runtimeApp C.6.
Scope of the claimsSec 7 (lines 526-528, 533-534).

We believe the revised paper addresses all the concerns raised, and we are happy to clarify any remaining points.

AuthorsOpenReview ↗

Summary of the discussion

We thank the reviewers again for engaging in the discussion, and we briefly summarise its outcome.

All three reviewers rated the soundness as having “minor or fixable concerns”. Their shared concerns are addressed with new experiments: LeWM with its original CEM planner does not close the gap, BFS on CGWM solves every instance, with incorrect knowledge the error stays on the affected concepts while a learned residual predictor corrects the dynamics, and without the constraints grounding all concepts remains possible.

Reviewer 5qgC agreed that the CEM results resolve the concern about the baseline, and that the main contribution is learning to ground concepts from weak supervision so that given rules can be used. The two remaining points, that the planning results come from the rules and that grounding requires largely correct knowledge, are the premise of this contribution. Without direct concept labels, learning correct concepts from incorrect knowledge is analogous to learning correct classes from incorrect labels. Reviewer bhZG stated that our responses addressed some of the points and kept the score taking the other reviews into account, without indicating which concerns remain open. We would welcome this indication, so that we can address it.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.