acceptodds
Under review as a conference paper at ICLR 2027

GADAR: Generalizing Across Domains via Action Ranking for Classical Planning

Abstract

Learned policies for classical planning act without search, but each is tied to the domain it was trained on: network modules per action schema, parameters per predicate, or vocabulary-sized encodings mean a trained policy cannot even be executed on a new domain’s predicates and action schemas. We present GADAR (Generalizing Across Domains via Action Ranking), a policy that takes the do- main description as input, π(s, G, D), and constructs grounded actions in domains never seen in training, without search. Symbols are described by the roles they play in the domain rather than by their names, in three layers that each extend the previous one: (i) a lifted domain layer embeds D in every state graph through structural signatures; (ii) a binding layer ties each ground action’s applicability to the lifted preconditions it instantiates; and (iii) a chain layer states which schema produces what another needs and how far each schema stands from the goal. The same network therefore runs on every domain, seen or unseen. We evaluate zero- shot generalization leave-one-domain-out over eight benchmark domains. A sin- gle GADAR policy trained on seven domains solves 58.3% of their test instances, one point below the 59.3% of dedicated single-domain GADAR models, so one policy serves many domains at no cost in coverage; a multi-task control that keeps symbol identity in its weights, while strong on some of the domains it trains on, transfers to none of the held-out ones. Zero-shot, GADAR executes on every held-out domain and exceeds random choice under the same executor on the large test instances of all eight: it solves 22.6% of them on average, against 3.3% for random choice and 0.6% for the same representation without the chain layer, and reaches 68% on held-out Visitall. Zero-shot coverage nonetheless remains well below in-domain coverage on most domains, and we analyze where it is weakest, including domains whose action schemas are structurally alike but play different roles.

Primary area:
applications to robotics, autonomy, planning
Keywords:
planning, learning for planning, relational learning, graph neural networks, sequential decision making, domain independent, size generalization, representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_9YTxOpenReview ↗
Rating
2
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces GADAR, a domain agnostic, search free classical planning strategy. Taking the current state, goal, and domain description, denoted as π(s, G, D), as input, the strategy directly constructs domain based actions in unseen domains. Its core contribution lies in a domain general graph representation featuring two key additions: binding edges, which explicitly link applicable domain descriptions to lifted precondition instances, and chain layers, which encode producer consumer relationships between schema instances and their distance to the goal. Across eight benchmark domains, a single multi domain GADAR model achieves in domain coverage comparable to domain specific versions and demonstrates zero shot execution in held out domains, significantly outperforming random selection and symbolic multi task control models.

Critical strengths

The paper introduces two novel and well-motivated representations of relationships: (i) binding edges that connect objects in the applicable base to lifted precondition occurrences, and (ii) chain layers atop these lifted occurrences that encode enabling/threat relations and goal-distance buckets. Furthermore, by incorporating the domain description into the policy input, it bridges a significant gap between domain-independent heuristics and per-domain policies. (I place particular importance on this aspect.)

Critical weaknesses

1: Experiments show that while zero-shot coverage is significantly higher than random coverage, it remains low on average—and very low in certain domains (e.g., 4% for logistics, 6% for grippers)—highlighting the limitations of relying solely on structural, renaming-invariant features to distinguish between structurally similar patterns. (This point is important to me.)

2: Regarding the lack of comparison with related work, this paper offers only a limited discussion of recent search-guided relational strategy learning methods (such as GSP).

Questions for authors

1: Could you compare the zero-shot strategy against domain-agnostic heuristic search methods (e.g., GOOSE with LLG) on the held-out domains to demonstrate the utility of the zero-shot strategy relative to search-based zero-shot baselines?

2: Could you conduct ablation studies by incrementally adding chain edges, grounding-effect edges (EeffE_{eff}), and goal features to the GADAR-BIND model to isolate their respective contributions? This would help clarify which component drives the significant performance gains observed in models like Visitall.

Reviewer_QEpzOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

GADAR is a GNN policy π(s,G,D)\pi(s,G,D) where ss is the state, GG the goal, and DD lifted domain descripton, that takes the PDDL domain as input to construct grounded actions in unseen domains without search. It extends GABAR's action-focused instance graph with three additions: i) a lifted domain layer, ii) binding edges that link objects of applicable groundings to the lifted precondition atoms they satisfy, iii) a "chain layer", i.e., enables/threatens edges, achieves-goal/deletes-true edges, a categorical schema goal distance and goal-relevance bits. Features are unaware of object identity, so the graph is permutation-invariant. The encoder and decoder are GABAR's, and training imitates LAMA-first plans. In leave-one-domain-out evaluation over GABAR's eight domains, GADAR reports only mediocre results, indicating that cross-domain generalization remains challenging.

Critical strengths

Generalisation across domains via a single learned policy or value heuristic is an important open problem. The ladder design, with a shared encoder, decoder and executor and graph-based representation, is a principled way to do so. Results are reported honestly, and the negative GADAR-BIND result is genuinely informative. However, several weaknesses (below) do not support the claims made in the paper.

Critical weaknesses

A major issue is that the experiments do not show which part matters and that results are strikingly weak.

GADAR-BIND performs worse than random, while GADAR without binding edges has not been evaluated. The only generalizing mode packages hand-crafted features that may well be the key information for the network on their own. The step from 0.6% performance to 22.6% adds enables/threatens together with hand-crafted lookahead features (achieves/deletes, goal distance, goal relevance). The authors state that these are not separately evaluated. The results don't convince that the network learned generalizing behaviour over domain structure, instead of ranking over hand-coded goal-lookahead features.

Overall, zero-shot generalization coverage is poor at 22.6 on average; near zero on Logistics and Gripper. The latter two are common expressivity hurdles for GNNs in planning, so a poorer outcome is not unexpected here. However, Gripper is the paper's own motivating example of a domain with structure that should be shared with Miconic, yet it reaches only 6% here too, 19% in Miconic. These are standard and basic domains with simple generalizing strategies. On its own, this is not a huge weakness, given that the cross-domain task is hard. Yet, in-domain coverage is not close to the state of the art either. GADAR 58.3% vs. 92.5% for published GABAR, and Logistics and Rovers stay below 10% even when trained on directly. Related work (e.g. Ståhlberg/Chen et. al.) showcases that SOTA for in-domain training lies at around 100% for many of these domains. The abstract claims to serve 'many domains at no cost in coverage', while averages swing from -37 in Spanner to +27 Visitall, against a single-domain baseline that is itself far below the state of the art. The concrete reasons for this weakness are not explored.

UNION, essentially GABAR with a widened vocabulary, reaches 18.6/9.6 on Logistics/Rovers vs. 79/82 published in the original paper. It was trained for 100 epochs vs. 500, contradicting the claimed identical training procedure. It is unclear to me whether the discrepancy stems from different test data or from systematic inaccuracies in implementation.

Finally, a strong or more meaningful cross-domain baseline is missing. UNION's 0% is by design. GOOSE and a greedy domain-independent heuristic policy are missing and directly valuable.

Questions for authors

Could change my assessment:

Please ablate enables/threatens, achieves/deletes, and goal distance/relevance separately, and evaluate GADAR without binding edges. If possible, please report a non-learned baseline that ranks actions by the hand-coded features alone. If it matches GADAR's coverage, the learned component likely contributes little. What is the achievement oracle's coverage on Visitall (reported as "–%") and on the other domains? Can you re-run GABAR on your split with your pipeline? Why was UNION trained for only 100 epochs? Matching training setups would be welcome.

Suggestions (not score relevant):

On the "first domain-agnostic policy" claim: an LLM prompted with the PDDL domain is also a search-free π(s,G,D)\pi(s,G,D). Evaluation on the IPC 2023 learning track for comparability with GOOSE.

Reviewer_HhM5OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes GADAR, a neural network architecture that extends GABAR (ICAPS25) to work domain-independently. Unlike existing domain-independent search policy learning systems, which primarily focus on returning an heuristic value, GABAR produces an action schema name and its parameters in each prediction. The output of GADAR has the same form, but in order to handle different domains, GADAR has a few more modules so that the output actions actually belongs to the input domain.

This is an incremental, empirical paper with weak results, missing key evaluations. In addition, the writing is a huge LLM slop, which is obvious from various clues: use of non-standard terminology which confuses technical terms with more general english words, non-precise wording, and distinctive/unnatural writing style.

  • line 16: also contain goal description G
  • line 16: constructs grounded actions -> plans
  • line 19: "state graph"
  • line 22: "what another needs" --- AI smell
  • line 22: "how far each schema stands from the goal" --- AI smell
  • line 28: "Zero-shot, GADAR ..." --- AI smell
  • line 29: "executor" --- AI smell?
  • line 52: "it is also what binds" --- AI smell
  • line 59: "construct grounded actions to search" --- clearly this part is talking about generating a plan / action sequence,
  • but this reads as if it is about action grounding from the lifted schema, which is a necessary step for planners that uses grounded representation (SAS / propositional). This mistake only happens in LLMs.
  • line 65: "produce or destroy" --- another pattern that only AI would write. This part talks about add/delete-effect.
  • line 156: action "a" is undefined.

Even with effort, the method is difficult to understand. This paper also heavily assumes readers to be already familiar with GABER, which makes it not self-contained.

For example, the most important part of the training is the training loss definition. It says the network is trained by cross-entropy loss, but does not specify how the output is represented; one-hot vector, logit, over what? How do you match the predicted arity and the ground-truth arity? This kind of subtle but important implementation detail becomes clear only after reading the GABER paper and is relatively complex; The arity k of action schema is derived from the ground-truth action schema (not predicted), and the GRU is unrolled exactly k times to avoid dimension mismatch.

Critical strengths

None

Critical weaknesses

It is unclear what is the particular strength in outputting the ground action directly, rather than heuristic values. Having a heuristic value allows discussing the theoretical characteristics, and an opportunity to combine it with various search algorithms. If the goal is to avoid the action schema grounding, which is indeed problematic in some grounding-heavy domains, then the approach must be evaluated in those domains, such as caldera and organic synthesis in IPC2018.

There is no comparison against heuristic-based (learned, or symbolic) planners.

Questions for authors

The main method section does not specify the training data. It is later mentioned that they originate from lama-first suboptimal plan. Does training data affect the performance, e.g., whether it is optimal or suboptimal?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.