pith. machine review for the scientific record. sign in

arxiv: 2604.16871 · v1 · submitted 2026-04-18 · 💻 cs.AI · cs.LG

Recognition: unknown

GRAIL: Autonomous Concept Grounding for Neuro-Symbolic Reinforcement Learning

Henri R\"o{\ss}ler, Hikaru Shindo, Kristian Kersting, Quentin Delfosse

Authors on Pith no claims yet

Pith reviewed 2026-05-10 07:16 UTC · model grok-4.3

classification 💻 cs.AI cs.LG
keywords neuro-symbolic reinforcement learningconcept groundingrelational conceptslarge language modelsweak supervisionAtari gamesautonomous adaptation
0
0 comments X

The pith

GRAIL enables neuro-symbolic RL agents to autonomously ground relational concepts by refining LLM-provided generic representations through environmental interaction.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces GRAIL as a way to let reinforcement learning agents define their own relational concepts instead of relying on human experts. It starts with broad ideas from large language models and sharpens them using feedback from the game environment. This helps agents deal with unclear rewards and mismatched understandings in complex tasks. Tests in Atari games like Kangaroo show the method works as well as or better than hand-designed concepts in simpler cases. It also highlights choices agents make between scoring points and achieving bigger goals.

Core claim

GRAIL autonomously grounds relational concepts through environmental interaction. It leverages large language models to provide generic concept representations as weak supervision, then refines them to capture environment-specific semantics. This addresses both sparse reward signals and concept misalignment prevalent in underdetermined environments.

What carries the argument

The refinement process in GRAIL, where LLM generic representations are updated based on environmental interactions to match specific semantics.

If this is right

  • Neuro-symbolic agents no longer need manual concept definitions for each environment.
  • Better performance in sparse reward settings through grounded concepts.
  • Agents can balance reward maximization with high-level goal completion.
  • Adaptability to different environments without expert intervention.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This could make neuro-symbolic RL more practical for real-world applications where concepts vary.
  • Future work might explore combining this with other forms of weak supervision beyond LLMs.
  • The trade-offs observed suggest new ways to design reward functions that incorporate concept alignment.

Load-bearing premise

Generic representations supplied by LLMs can be reliably refined through environmental interaction into accurate, stable, environment-specific concept semantics without introducing persistent misalignment or harming policy learning.

What would settle it

Observing that agents using GRAIL consistently fail to learn correct environment-specific meanings for concepts, leading to poor policy performance compared to manual baselines.

Figures

Figures reproduced from arXiv: 2604.16871 by Henri R\"o{\ss}ler, Hikaru Shindo, Kristian Kersting, Quentin Delfosse.

Figure 1
Figure 1. Figure 1: Concept grounding is environment-dependent in neuro-symbolic reinforcement learning. The spatial concept “left of” requires different interpretations across environments. In Kangaroo (left), the agent must verify that the player is both horizontally left of the ladder and vertically aligned on the same platform. In Seaquest (middle), “left of” is defined more flexibly based on horizontal positioning regard… view at source ↗
Figure 2
Figure 2. Figure 2: GRAIL learns environment-specific concept groundings. Given different environments, GRAIL au￾tonomously discovers distinct interpretations of “left” and “right.” In Kangaroo (left), the learned concepts activate along horizontal bands aligned with each platform, reflecting that “left of ladder” requires both horizontal offset and vertical alignment. In Skiing (right), activation extends diagonally above ea… view at source ↗
Figure 3
Figure 3. Figure 3: Valuation functions evaluate relationships between objects. These functions are typically hard-coded, limiting the applicability of neuro-symbolic reinforcement learning. GRAIL aims to learn these functions by aligning them with the correct concept directly from interactions with the environment. 2.2 Neuro-Symbolic Reinforcement Learning GRAIL specifically builds upon the ideas of neuro-symbolic reasoning … view at source ↗
Figure 4
Figure 4. Figure 4: GRAIL: Framework Overview. GRAIL extends the BlendRL framework [25], uniting neural and logic￾based policies within a neuro-symbolic RL agent. Here, the logic policy consists of weighted rules expressed over predicates, where each predicate is equipped with a differentiable valuation function capturing abstract state relations. Unlike BlendRL, which relies on hand-crafted predicate valuations, GRAIL introd… view at source ↗
Figure 5
Figure 5. Figure 5: GRAIL’s Policy Reasoning. A concept grounding module takes object-centric features Z extracted from an image X and computes object relations via differentiable valuation functions v p ψ . Those are then applied to a set of logical rules through forward reasoning to determine a logic policy. Likewise, a blending module utilizes the relational information to combine the logic policy with a neural policy that… view at source ↗
Figure 6
Figure 6. Figure 6: Logic programs used by (a) the logic actor and (b) the blending module in three Atari environments, generated by LLMs following [25]. Unlike previous studies, where spatial predicates are hand-crafted by human experts, GRAIL grounds these predicates as parameterized differentiable functions. In Skiing, the blending program always delegates to the logic agent. architecture. Given a binary spatial predicate … view at source ↗
Figure 7
Figure 7. Figure 7: GRAIL maximizes reward and aligns concepts semantically. Step 1: We generate proxy functions for each spatial predicate using LLMs. These functions are represented as normalized activation maps, typically produced as code snippets and visualized over the 2D state space. Step 2: The proxy functions are incorporated as an additional supervision signal during policy optimization. This auxiliary signal guides … view at source ↗
Figure 8
Figure 8. Figure 8: GRAIL achieves high-level goals reliably by grounding spatial concepts. Average number of goals achieved per episode on the simplified environment (Stage 1, enemies removed, logic policy only). In Kangaroo, a goal corresponds to reaching the child at the top platform; in Seaquest, a goal corresponds to successfully rescuing all six divers and surfacing. The red dashed line indicates the hard-coded BlendRL-… view at source ↗
Figure 9
Figure 9. Figure 9: GRAIL produces interpretable spatial concepts that capture subtle environmental details (Kangaroo). Heatmaps indicate the truth values produced by the valuation functions as the player’s position varies within the scene. Each white dot denotes the location of a ladder, relative to which truth values for the spatial predicates are evaluated. The results are visualized (from left to right) for: hand-crafted … view at source ↗
Figure 10
Figure 10. Figure 10 [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Return vs. goal completion in the full environment (Stage 2). Blue bars show average episodic return; pink bars show goals achieved per episode (reaching the child in Kangaroo; rescuing all divers in Seaquest). In Kangaroo, GRAIL (Claude) achieves the highest return but BlendRL’s hand-coded strategy yields far more goals. In Seaquest, only GRAIL completes any goals, while baselines that maximize return fa… view at source ↗
Figure 12
Figure 12. Figure 12: Stage 2: learned close by enemy concept in Seaquest. The blending module decides when to del￾egate control to the symbolic policy based on learned spatial predicates. Each heatmap shows the activation of close by enemy as a function of relative object position; white dots mark enemy positions. GRAIL produces broader, horizontally extended activations that capture lateral enemy movement, whereas hand-coded… view at source ↗
Figure 13
Figure 13. Figure 13: compares the concept alignment loss L CA and the number of goals achieved during Stage 1 training in Kangaroo for different values of the alignment coefficient cCA ∈ {0.3, 1.0} and the annealing factor γCA ∈ {0, 1}. Two observations stand out. First, performance consistently improves as the learned concepts diverge from the LLM proxy functions—rising L CA coincides with rising goals—indicating that the ag… view at source ↗
Figure 14
Figure 14. Figure 14: illustrates misaligned spatial concepts in Kangaroo. In the two leftmost examples, the agent incor￾rectly associates left of ladder with a ladder on a different platform. A similar cross-platform confusion arises for right of ladder (center). The two rightmost examples reveal a complementary failure mode: the agent’s on ladder activation is biased toward the top platform even when evaluated relative to la… view at source ↗
read the original abstract

Neuro-symbolic Reinforcement Learning (NeSy-RL) combines symbolic reasoning with gradient-based optimization to achieve interpretable and generalizable policies. Relational concepts, such as "left of" or "close by", serve as foundational building blocks that structure how agents perceive and act. However, conventional approaches require human experts to manually define these concepts, limiting adaptability since concept semantics vary across environments. We propose GRAIL (Grounding Relational Agents through Interactive Learning), a framework that autonomously grounds relational concepts through environmental interaction. GRAIL leverages large language models (LLMs) to provide generic concept representations as weak supervision, then refines them to capture environment-specific semantics. This approach addresses both sparse reward signals and concept misalignment prevalent in underdetermined environments. Experiments on the Atari games Kangaroo, Seaquest, and Skiing demonstrate that GRAIL matches or outperforms agents with manually crafted concepts in simplified settings, and reveals informative trade-offs between reward maximization and high-level goal completion in the full environment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 0 minor

Summary. The paper proposes GRAIL, a neuro-symbolic RL framework that uses LLMs to supply generic relational concept representations (e.g., 'left of', 'close by') as weak supervision, then refines these representations through environmental interaction to capture environment-specific semantics. It claims this addresses sparse rewards and concept misalignment in underdetermined settings, with experiments on Atari games (Kangaroo, Seaquest, Skiing) showing that GRAIL matches or outperforms agents using manually crafted concepts in simplified settings while revealing reward vs. goal-completion trade-offs in full environments.

Significance. If the refinement process and empirical results hold, GRAIL would reduce reliance on human experts for concept definition in NeSy-RL and improve adaptability across environments. The approach of leveraging LLM priors followed by interaction-based grounding is a potentially useful direction for handling concept misalignment, but the absence of quantitative results, baselines, implementation details, or stability analysis in the manuscript limits its assessed significance.

major comments (3)
  1. [Abstract] Abstract: the central empirical claim that 'experiments on the Atari games Kangaroo, Seaquest, and Skiing demonstrate that GRAIL matches or outperforms agents with manually crafted concepts' is unsupported, as the text provides no quantitative results, baselines, implementation details, error analysis, or performance metrics.
  2. [GRAIL framework description] Description of the refinement process: the claim that LLM generic representations can be reliably refined into accurate, stable, environment-specific semantics rests on an update rule that appears purely reward-driven; no anchoring loss, regularization term, semantic fidelity constraint, or convergence criterion is described that would prevent drift or persistent misalignment in sparse-reward Atari settings.
  3. [Experiments] Experimental validation section: without details on how concepts are represented (e.g., as embeddings or predicates), how the refinement update is implemented, or how semantic accuracy is measured post-refinement, it is impossible to assess whether the reported performance gains stem from improved concept grounding or from other factors.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback on our submission. The comments identify important areas for improvement in supporting our claims with details and evidence. We will perform a major revision to incorporate quantitative results, detailed methodological descriptions, and implementation specifics. We respond to each major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central empirical claim that 'experiments on the Atari games Kangaroo, Seaquest, and Skiing demonstrate that GRAIL matches or outperforms agents with manually crafted concepts' is unsupported, as the text provides no quantitative results, baselines, implementation details, error analysis, or performance metrics.

    Authors: We acknowledge that the abstract's empirical claim is not supported by quantitative data in the current manuscript text. The full paper describes the experiments but does not include specific metrics or baselines in the provided sections. To address this, we will revise the abstract to either qualify the claim or include key results (e.g., average rewards or success rates), and add a results table with comparisons to manual concepts, including error analysis where available. revision: yes

  2. Referee: [GRAIL framework description] Description of the refinement process: the claim that LLM generic representations can be reliably refined into accurate, stable, environment-specific semantics rests on an update rule that appears purely reward-driven; no anchoring loss, regularization term, semantic fidelity constraint, or convergence criterion is described that would prevent drift or persistent misalignment in sparse-reward Atari settings.

    Authors: The referee correctly notes that the refinement process description is incomplete regarding safeguards against drift. The current manuscript presents the refinement as interaction-based but does not detail the update mechanism beyond reward signals. In the revision, we will provide the precise update rule, introduce regularization terms for semantic fidelity if not already present, and specify convergence criteria. We will also analyze the risk of misalignment in sparse-reward settings and how the LLM weak supervision helps anchor the process. revision: yes

  3. Referee: [Experiments] Experimental validation section: without details on how concepts are represented (e.g., as embeddings or predicates), how the refinement update is implemented, or how semantic accuracy is measured post-refinement, it is impossible to assess whether the reported performance gains stem from improved concept grounding or from other factors.

    Authors: We agree that the experimental section lacks sufficient implementation details for full assessment. We will expand this section to explain that relational concepts are represented as differentiable embeddings in the neuro-symbolic policy, provide pseudocode and equations for the refinement update rule, and describe the post-refinement semantic accuracy evaluation using environment-specific proxies or human annotations. This will allow readers to determine the contribution of the grounding process. revision: yes

Circularity Check

0 steps flagged

No circularity: framework is a descriptive method validated by experiments, not a self-referential derivation

full rationale

The paper describes GRAIL as a framework that supplies generic LLM concept representations as weak supervision and then refines them via environmental interaction to produce environment-specific semantics. No equations, derivations, or first-principles predictions are presented in the abstract or described claims that reduce by construction to fitted inputs, self-citations, or renamed known results. The central claims rest on empirical validation in Atari environments rather than any tautological loop where outputs are presupposed by the inputs. This matches the reader's assessment of no evident circular dependence.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Ledger constructed from abstract only; full paper may contain additional parameters or assumptions not visible here.

axioms (1)
  • domain assumption LLMs can supply generic concept representations that serve as useful weak supervision for relational concepts in RL environments
    This premise is invoked as the starting point for the GRAIL pipeline in the abstract.

pith-pipeline@v0.9.0 · 5477 in / 1296 out tokens · 57546 ms · 2026-05-10T07:16:05.098419+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

84 extracted references · 13 canonical work pages · 2 internal anchors

  1. [1]

    Playing Atari with Deep Reinforcement Learning

    V olodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Playing atari with deep reinforcement learning.arXiv preprint arXiv:1312.5602, 2013

    work page internal anchor Pith review arXiv 2013

  2. [2]

    Proximal Policy Optimization Algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  3. [3]

    Agent57: Outperforming the atari human benchmark

    Adri `a Puigdom`enech Badia, Bilal Piot, Steven Kapturowski, Pablo Sprechmann, Alex Vitvitskyi, Zhaohan Daniel Guo, and Charles Blundell. Agent57: Outperforming the atari human benchmark. InProceedings of the Inter- national Conference on Machine Learning (ICML), 2020

    2020

  4. [4]

    Bhatt, D

    A. Bhatt, D. Palenicek, B. Belousov, M. Argus, A. Amiranashvili, T. Brox, and J. Peters. Crossq: Batch nor- malization in deep reinforcement learning for greater sample efficiency and simplicity. InProceedings of the International Conference on Learning Representations (ICLR), 2024

    2024

  5. [5]

    A role of environmental complexity on representation learning in deep rein- forcement learning agents.arXiv preprint arXiv:2407.03436, 2024

    Andrew Liu and Alla Borisyuk. A role of environmental complexity on representation learning in deep rein- forcement learning agents.arXiv preprint arXiv:2407.03436, 2024

    work page arXiv 2024

  6. [6]

    Deep reinforcement learning agents are not even close to human intel- ligence.arXiv preprint arXiv:2505.21731, 2025

    Quentin Delfosse, Jannis Bl ¨uml, Fabian Tatai, Th´eo Vincent, Bjarne Gregori, Elisabeth Dillies, Jan Peters, Con- stantin Rothkopf, and Kristian Kersting. Deep reinforcement learning agents are not even close to human intel- ligence.arXiv preprint arXiv:2505.21731, 2025

    work page arXiv 2025

  7. [7]

    Inter- pretable concept bottlenecks to align reinforcement learning agents

    Quentin Delfosse, Sebastian Sztwiertnia, Wolfgang Stammer, Mark Rothermel, and Kristian Kersting. Inter- pretable concept bottlenecks to align reinforcement learning agents. InProceedings of the Conference on Ad- vances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  8. [8]

    Neural logic reinforcement learning

    Zhengyao Jiang and Shan Luo. Neural logic reinforcement learning. InProceedings of the International Con- ference on Machine Learning (ICML), 2019

    2019

  9. [9]

    Neuro-symbolic reinforcement learning with first-order logic

    Daiki Kimura, Masaki Ono, Subhajit Chaudhury, Ryosuke Kohita, Akifumi Wachi, Don Joven Agravante, Michi- aki Tatsubori, Asim Munawar, and Alexander Gray. Neuro-symbolic reinforcement learning with first-order logic. InProceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2021

    2021

  10. [10]

    GALOIS: boosting deep reinforcement learning via generalizable logic synthesis

    Yushi Cao, Zhiming Li, Tianpei Yang, Hao Zhang, Yan Zheng, Yi Li, Jianye Hao, and Yang Liu. GALOIS: boosting deep reinforcement learning via generalizable logic synthesis. InProceedings of the Conference on Advances in Neural Information Processing Systems (NeurIPS), 2022

    2022

  11. [11]

    Interpretable and explainable logical policies via neurally guided symbolic abstraction

    Quentin Delfosse, Hikaru Shindo, Devendra Dhami, and Kristian Kersting. Interpretable and explainable logical policies via neurally guided symbolic abstraction. InProceedings of the Conference on Advances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  12. [12]

    Shao-Hua Sun, Te-Lin Wu, and Joseph J. Lim. Program guided agent. InProceedings of the International Conference on Learning Representations (ICLR), 2020

    2020

  13. [13]

    Programmati- cally interpretable reinforcement learning

    Abhinav Verma, Vijayaraghavan Murali, Rishabh Singh, Pushmeet Kohli, and Swarat Chaudhuri. Programmati- cally interpretable reinforcement learning. InProceedings of the International Conference on Machine Learning (ICML), 2018. 19

    2018

  14. [14]

    SDRL: interpretable and data-efficient deep rein- forcement learning leveraging symbolic planning

    Daoming Lyu, Fangkai Yang, Bo Liu, and Steven Gustafson. SDRL: interpretable and data-efficient deep rein- forcement learning leveraging symbolic planning. InProceedings of the AAAI Conference on Artificial Intelli- gence (AAAI), 2019

    2019

  15. [15]

    Com- bining reinforcement learning and constraint programming for combinatorial optimization

    Quentin Cappart, Thierry Moisan, Louis-Martin Rousseau, Isabeau Pr ´emont-Schwarz, and Andre A Cire. Com- bining reinforcement learning and constraint programming for combinatorial optimization. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2021

    2021

  16. [16]

    Interpretable and editable programmatic tree policies for reinforcement learning.arXiv preprint arXiv:2405.14956, 2024

    Hector Kohler, Quentin Delfosse, Riad Akrour, Kristian Kersting, and Philippe Preux. Interpretable and editable programmatic tree policies for reinforcement learning.arXiv preprint arXiv:2405.14956, 2024

    work page arXiv 2024

  17. [17]

    Bruner, Jacqueline J

    Jerome S. Bruner, Jacqueline J. Goodnow, and George A. Austin.A Study of Thinking. John Wiley and Sons, 1956

    1956

  18. [18]

    Natural categories.Cognitive psychology, 4(3):328–350, 1973

    Eleanor H Rosch. Natural categories.Cognitive psychology, 4(3):328–350, 1973

    1973

  19. [19]

    James Archer

    E. James Archer. The psychological nature of concepts. InAnalyses of Concept Learning, pages 37–49. Elsevier, 1966

    1966

  20. [20]

    The Neuro-Symbolic Concept Learner: Interpreting Scenes, Words, and Sentences From Natural Supervision

    Jiayuan Mao, Chuang Gan, Pushmeet Kohli, Joshua B. Tenenbaum, and Jiajun Wu. The neuro-symbolic concept learner: Interpreting scenes, words, and sentences from natural supervision.arXiv preprint arXiv:1904.12584, 2019

    work page Pith review arXiv 1904

  21. [21]

    What’s left? concept grounding with logic-enhanced foundation models

    Joy Hsu, Jiayuan Mao, Josh Tenenbaum, and Jiajun Wu. What’s left? concept grounding with logic-enhanced foundation models. InProceedings of the Conference on Advances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  22. [22]

    Tenenbaum, and Jiajun Wu

    Jiayuan Mao, Joshua B. Tenenbaum, and Jiajun Wu. Neuro-symbolic concepts.arXiv preprint arXiv:2505.06191, 2025

    work page arXiv 2025

  23. [23]

    Tenenbaum

    Tom Silver, Rohan Chitnis, Nishanth Kumar, Willie McClinton, Tom ´as Lozano-P ´erez, Leslie Pack Kaelbling, and Joshua B. Tenenbaum. Predicate invention for bilevel planning. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2023

    2023

  24. [24]

    George A. V ouros. Explainable deep reinforcement learning: State of the art and challenges.ACM Computing Surveys, 55(5):1–39, 2022

    2022

  25. [25]

    Blendrl: A framework for merging symbolic and neural policy learning

    Hikaru Shindo, Quentin Delfosse, Devendra Singh Dhami, and Kristian Kersting. Blendrl: A framework for merging symbolic and neural policy learning. InProceedings of the International Conference on Learning Representations (ICLR), 2025

    2025

  26. [26]

    Learning explanatory rules from noisy data.Journal of Artificial Intelligence Research, 61:1–64, 2018

    Richard Evans and Edward Grefenstette. Learning explanatory rules from noisy data.Journal of Artificial Intelligence Research, 61:1–64, 2018

    2018

  27. [27]

    Hikaru Shindo, Viktor Pfanschilling, Devendra Singh Dhami, and Kristian Kersting.αILP: thinking visual scenes as differentiable logic programs.Machine Learning (MLJ), 2023

    2023

  28. [28]

    Lloyd.Foundations of Logic Programming

    John W. Lloyd.Foundations of Logic Programming. Springer, Berlin, Heidelberg, 1984

    1984

  29. [29]

    MIT Press, 2001

    Raymond Reiter.Knowledge in Action: Logical Foundations for Specifying and Implementing Dynamical Sys- tems. MIT Press, 2001

    2001

  30. [30]

    Differentiable inductive logic programming for struc- tured examples

    Hikaru Shindo, Masaaki Nishino, and Akihiro Yamamoto. Differentiable inductive logic programming for struc- tured examples. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2021

    2021

  31. [31]

    Relational reinforcement learning.Machine Learning (MLJ), 2001

    Saso Dzeroski, Luc De Raedt, and Kurt Driessens. Relational reinforcement learning.Machine Learning (MLJ), 2001

    2001

  32. [32]

    Bellman goes relational

    Kristian Kersting, Martijn van Otterlo, and Luc DeRaedt. Bellman goes relational. InProceedings of the Inter- national Conference on Machine Learning (ICML), 2004

    2004

  33. [33]

    Non-parametric policy gradients: a unified treatment of propositional and relational domains

    Kristian Kersting and Kurt Driessens. Non-parametric policy gradients: a unified treatment of propositional and relational domains. InProceedings of the International Conference on Machine Learning (ICML), 2008

    2008

  34. [34]

    Exploration in relational domains for model-based rein- forcement learning.Journal of Machine Learning Research (JMLR), 2012

    Tobias Lang, Marc Toussaint, and Kristian Kersting. Exploration in relational domains for model-based rein- forcement learning.Journal of Machine Learning Research (JMLR), 2012

    2012

  35. [35]

    Deep explainable relational reinforcement learning: A neuro-symbolic approach

    Rishi Hazra and Luc De Raedt. Deep explainable relational reinforcement learning: A neuro-symbolic approach. InMachine Learning and Knowledge Discovery in Databases: Research Track - European Conference (ECML PKDD), 2023

    2023

  36. [36]

    Neurosymbolic reinforcement learning and planning: A survey.IEEE Transactions on Artificial Intelligence, 5(5):1939–1953, 2023

    Kamal Acharya, Waleed Raza, Carlos Dourado, Alvaro Velasquez, and Houbing Herbert Song. Neurosymbolic reinforcement learning and planning: A survey.IEEE Transactions on Artificial Intelligence, 5(5):1939–1953, 2023. 20

    1939

  37. [37]

    Human-allied relational reinforcement learning

    Fateme Golivand Darvishvand, Hikaru Shindo, Sahil Sidheekh, Kristian Kersting, and Sriraam Natarajan. Human-allied relational reinforcement learning. In25th Annual Conference on Advances in Cognitive Systems (ACS), 2025

    2025

  38. [38]

    Learning differentiable logic programs for abstract visual reasoning.Machine Learning (MLJ), 2024

    Hikaru Shindo, Viktor Pfanschilling, Devendra Singh Dhami, and Kristian Kersting. Learning differentiable logic programs for abstract visual reasoning.Machine Learning (MLJ), 2024

    2024

  39. [39]

    Concept bottleneck models

    Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. InProceedings of the International Conference on Machine Learning (ICML), 2020

    2020

  40. [40]

    Concept embedding models: Beyond the accuracy-explainability trade-off

    Mateo Espinosa Zarlenga, Pietro Barbiero, Gabriele Ciravegna, Giuseppe Marra, Francesco Giannini, Michelan- gelo Diligenti, Zohreh Shams, Frederic Precioso, Stefano Melacci, Adrian Weller, Pietro Li´o, and Mateja Jamnik. Concept embedding models: Beyond the accuracy-explainability trade-off. InProceedings of the Conference on Advances in Neural Informatio...

    2022

  41. [41]

    Right for the right concept: Revising neuro- symbolic concepts by interacting with their explanations

    Wolfgang Stammer, Patrick Schramowski, and Kristian Kersting. Right for the right concept: Revising neuro- symbolic concepts by interacting with their explanations. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 2021

    2021

  42. [42]

    Object-centric concept-bottlenecks

    David Steinmann, Wolfgang Stammer, Antonia W¨ust, and Kristian Kersting. Object-centric concept-bottlenecks. InProceedings of the Conference on Advances in Neural Information Processing Systems (NeurIPS), 2025

    2025

  43. [43]

    Neuro-symbolic predicate in- vention: Learning relational concepts from visual scenes.Neurosymbolic Artificial Intelligence Journal (NAIJ), 2025

    Jingyuan Sha, Hikaru Shindo, Kristian Kersting, and Devendra Singh Dhami. Neuro-symbolic predicate in- vention: Learning relational concepts from visual scenes.Neurosymbolic Artificial Intelligence Journal (NAIJ), 2025

    2025

  44. [44]

    Gestalt vision: A dataset for evaluating gestalt principles in visual perception

    Jingyuan Sha, Hikaru Shindo, Kristian Kersting, and Devendra Singh Dhami. Gestalt vision: A dataset for evaluating gestalt principles in visual perception. In19th International Conference on Neurosymbolic Learning and Reasoning (NeSy), 2025

    2025

  45. [45]

    Synthesizing visual concepts as vision-language programs

    Antonia W ¨ust, Wolfgang Stammer, Hikaru Shindo, Lukas Helff, Devendra Singh Dhami, and Kristian Kersting. Synthesizing visual concepts as vision-language programs. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026

    2026

  46. [46]

    Deisam: Segment anything with deictic prompting

    Hikaru Shindo, Manuel Brack, Gopika Sudhakaran, Devendra Singh Dhami, Patrick Schramowski, and Kristian Kersting. Deisam: Segment anything with deictic prompting. InProceedings of the Conference on Advances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  47. [47]

    V-lol: A diagnostic dataset for visual logical learning.arXiv preprint arXiv:2306.07743, 2023

    Lukas Helff, Wolfgang Stammer, Hikaru Shindo, Devendra Singh Dhami, and Kristian Kersting. V-lol: A diagnostic dataset for visual logical learning.arXiv preprint arXiv:2306.07743, 2023

    work page arXiv 2023

  48. [48]

    James F. Allen. Maintaining knowledge about temporal intervals.Communications of the ACM, 26(11):832–843, 1983

    1983

  49. [49]

    Object-centric learning with slot attention

    Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. InProceedings of the Conference on Advances in Neural Information Processing Systems (NeurIPS), 2020

    2020

  50. [50]

    SPACE: unsupervised object-oriented scene representation via spatial attention and decomposition

    Zhixuan Lin, Yi-Fu Wu, Skand Vishwanath Peri, Weihao Sun, Gautam Singh, Fei Deng, Jindong Jiang, and Sungjin Ahn. SPACE: unsupervised object-oriented scene representation via spatial attention and decomposition. InProceedings of the International Conference on Learning Representations (ICLR), 2020

    2020

  51. [51]

    Conditional object-centric learning from video

    Thomas Kipf, Gamaleldin Fathy Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional object-centric learning from video. In Proceedings of the International Conference on Learning Representations (ICLR), 2022

    2022

  52. [52]

    Boosting object representation learning via motion and object continuity

    Quentin Delfosse, Wolfgang Stammer, Thomas Rothenbacher, Dwarak Vittal, and Kristian Kersting. Boosting object representation learning via motion and object continuity. InMachine Learning and Knowledge Discovery in Databases: Research Track - European Conference (ECML PKDD), 2023

    2023

  53. [53]

    Entity-centric reinforcement learning for object manipulation from pixels

    Dan Haramati, Tal Daniel, and Aviv Tamar. Entity-centric reinforcement learning for object manipulation from pixels. InProceedings of the International Conference on Learning Representations (ICLR), 2024

    2024

  54. [54]

    Slot-based object-centric reinforcement learning algorithm

    Chao Chen, Fei Wang, and Xinyao Wang. Slot-based object-centric reinforcement learning algorithm. InInter- national Conference on CYBER Technology in Automation, Control, and Intelligent Systems, 2024

    2024

  55. [55]

    Interpretable end-to-end neurosymbolic reinforcement learning agents.arXiv preprint arXiv:2410.14371, 2024

    Nils Grandien, Quentin Delfosse, and Kristian Kersting. Interpretable end-to-end neurosymbolic reinforcement learning agents.arXiv preprint arXiv:2410.14371, 2024

    work page arXiv 2024

  56. [56]

    Sold: Slot object-centric latent dy- namics models for relational manipulation learning from pixels

    Malte Mosbach, Jan Niklas Ewertz, Angel Villar-Corrales, and Sven Behnke. Sold: Slot object-centric latent dy- namics models for relational manipulation learning from pixels. InProceedings of the International Conference on Machine Learning (ICML), 2025. 21

    2025

  57. [57]

    Better decisions through the right causal world model.arXiv preprint arXiv:2504.07257, 2025

    Elisabeth Dillies, Quentin Delfosse, Jannis Bl ¨uml, Raban Emunds, Florian Peter Busch, and Kristian Kersting. Better decisions through the right causal world model.arXiv preprint arXiv:2504.07257, 2025

    work page arXiv 2025

  58. [58]

    Object-centric world models for causality-aware reinforcement learn- ing

    Yosuke Nishimoto and Takashi Matsubara. Object-centric world models for causality-aware reinforcement learn- ing. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2026

    2026

  59. [59]

    Deep reinforcement learning via object-centric attention.arXiv preprint arXiv:2504.03024, 2025

    Jannis Bl ¨uml, Cedric Derstroff, Bjarne Gregori, Elisabeth Dillies, Quentin Delfosse, and Kristian Kersting. Deep reinforcement learning via object-centric attention.arXiv preprint arXiv:2504.03024, 2025

    work page arXiv 2025

  60. [60]

    Learning interactive world model for object-centric reinforcement learning

    Fan Feng, Phillip Lippe, and Sara Magliacane. Learning interactive world model for object-centric reinforcement learning. InProceedings of the Conference on Advances in Neural Information Processing Systems (NeurIPS), 2025

    2025

  61. [61]

    Prafulla Dhariwal, Christopher Hesse, Oleg Klimov, Alex Nichol, Matthias Plappert, Alec Radford, John Schulman, Szymon Sidor, Yuhuai Wu, and Peter Zhokhov

    Quentin Delfosse, Jannis Bl ¨uml, Bjarne Gregori, Sebastian Sztwiertnia, and Kristian Kersting. Ocatari: Object- centric atari 2600 reinforcement learning environments.arXiv preprint arXiv:2306.08649, 2023

    work page arXiv 2023

  62. [62]

    End-to-end neuro-symbolic reinforcement learning with textual explanations

    Lirui Luo, Guoxi Zhang, Hongming Xu, Yaodong Yang, Cong Fang, and Qing Li. End-to-end neuro-symbolic reinforcement learning with textual explanations. InProceedings of the International Conference on Machine Learning (ICML), 2024

    2024

  63. [63]

    Objects matter: Object-centric world models improve reinforcement learning in visually complex environments

    Weipu Zhang, Adam Jelley, Trevor McInroe, and Amos Storkey. Objects matter: Object-centric world models improve reinforcement learning in visually complex environments. InReinforcement Learning and Video Games Workshop@ RLC, 2025

    2025

  64. [64]

    Rusu, Joel Veness, Marc G

    V olodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex Graves, Martin A. Riedmiller, Andreas Fidjeland, Georg Ostrovski, Stig Petersen, Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wierstra, Shane Legg, and Demis Hassabis. Human- level control through deep reinforcement...

    2015

  65. [65]

    You only look once: Unified, real-time object detection

    Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016

    2016

  66. [66]

    Fast segment anything,

    Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything.arXiv preprint arXiv:2306.12156, 2023

    work page arXiv 2023

  67. [67]

    Self-supervised visual reinforcement learning with object-centric representations

    Andrii Zadaianchuk, Maximilian Seitzer, and Georg Martius. Self-supervised visual reinforcement learning with object-centric representations. InProceedings of the International Conference on Learning Representations (ICLR), 2021

    2021

  68. [68]

    Semantic tracklets: An object- centric representation for visual multi-agent reinforcement learning

    Iou-Jen Liu, Zhongzheng Ren, Raymond A Yeh, and Alexander G Schwing. Semantic tracklets: An object- centric representation for visual multi-agent reinforcement learning. InProceedings of the IEEE/RSJ Interna- tional Conference on Intelligent Robots and Systems (IROS), 2021

    2021

  69. [69]

    An investigation into pre-training object-centric rep- resentations for reinforcement learning

    Jaesik Yoon, Yi-Fu Wu, Heechul Bae, and Sungjin Ahn. An investigation into pre-training object-centric rep- resentations for reinforcement learning. InProceedings of the International Conference on Machine Learning (ICML), 2023

    2023

  70. [70]

    Pix2code: Learning to compose neural visual concepts as programs

    Antonia W ¨ust, Wolfgang Stammer, Quentin Delfosse, Devendra Singh Dhami, and Kristian Kersting. Pix2code: Learning to compose neural visual concepts as programs. InProceedings of Conference on Uncertainty in Artificial Intelligence (UAI), 2024

    2024

  71. [71]

    Neural concept binder

    Wolfgang Stammer, Antonia W ¨ust, David Steinmann, and Kristian Kersting. Neural concept binder. InProceed- ings of the Conference on Advances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  72. [72]

    Rainbow: Combining improvements in deep reinforcement learning

    Matteo Hessel, Joseph Modayil, Hado Van Hasselt, Tom Schaul, Georg Ostrovski, Will Dabney, Dan Horgan, Bilal Piot, Mohammad Azar, and David Silver. Rainbow: Combining improvements in deep reinforcement learning. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2018

    2018

  73. [73]

    Decision transformer: Reinforcement learning via sequence modeling

    Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. InProceed- ings of the Conference on Advances in Neural Information Processing Systems (NeurIPS), 2021

    2021

  74. [74]

    Stabilizing transformers for reinforcement learn- ing

    Emilio Parisotto, Francis Song, Jack Rae, Razvan Pascanu, Caglar Gulcehre, Siddhant Jayakumar, Max Jader- berg, Raphael Lopez Kaufman, Aidan Clark, Seb Noury, et al. Stabilizing transformers for reinforcement learn- ing. InProceedings of the International Conference on Machine Learning (ICML), 2020

    2020

  75. [75]

    Interactive disentanglement: Learning concepts by interacting with their prototype representations

    Wolfgang Stammer, Marius Memmel, Patrick Schramowski, and Kristian Kersting. Interactive disentanglement: Learning concepts by interacting with their prototype representations. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), 2022. 22

    2022

  76. [76]

    Human-in-the- loop or ai-in-the-loop? automate or collaborate? InProceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2025

    Sriraam Natarajan, Saurabh Mathur, Sahil Sidheekh, Wolfgang Stammer, and Kristian Kersting. Human-in-the- loop or ai-in-the-loop? automate or collaborate? InProceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2025

    2025

  77. [77]

    Claude 4 sonnet.https://claude.ai/, 2025

    Anthropic. Claude 4 sonnet.https://claude.ai/, 2025. [Large language model]

    2025

  78. [78]

    Chatgpt-4o.https://openai.com/research/, 2025

    OpenAI. Chatgpt-4o.https://openai.com/research/, 2025. [Large language model]

    2025

  79. [79]

    Bellemare, Yavar Naddaf, Joel Veness, and Michael Bowling

    Marc G. Bellemare, Yavar Naddaf, Joel Veness, and Michael Bowling. The arcade learning environment: An evaluation platform for general agents.Journal of Artificial Intelligence Research (JAIR), 2013

    2013

  80. [80]

    Hackatari: Atari learning environments for robust and continual reinforcement learning

    Quentin Delfosse, Jannis Bl ¨uml, Bjarne Gregori, and Kristian Kersting. Hackatari: Atari learning environments for robust and continual reinforcement learning. InWorking Notes of the RLC 2024 Workshop on Interpretable Policies in Reinforcement Learning, 2024

    2024

Showing first 80 references.