Open research questions in AI-based Problem Solving and Planning

What the literature leaves open

• Crucially, even when reflection substantially increases the number of fixed_by edges, these relations remain sparse relative to overall graph size, indicating that EXG-based methods rely on amplifying a limited number of well-aligned, high-impact fixes rather than on frequent trial-and-error corrections.

  EXG: Self-Evolving Agents with Experience Graphs · 2026

• Future research could focus on integrating these diverse domains into a more unified RLVR framework and evaluating performance across a broader range of model scales. , 70B parameters) remains to be further explored.

  Knowledge-to-Verification: Exploring RLVR for LLMs in Knowledge-Intensive Domains · 2026

• Component-Isolation Ablations. While we have isolated the validator’s contribution via a controlled experiment (loop active, validator dis- abled; §5.3), controlled comparisons of lazy vs. full schema injection across model sizes—particularly on SGD where the 26-service schema amplifies context load—and always-on history inclusion vs. on-demand retrieval via τH remain as future work. Extended Dialogue Management. A natural ex- tension of ReacTOD is to enrich the agent’s tool repertoire beyond Intent Classification and Slot Resolution to support broader dialogue manage- ment capabilities. In particular, we plan to intro- duce tools for handling conversational control flows that arise frequently in real-world deployments: for instance, a wait or clarification-pending tool to gracefully manage turns where the system must defer resolution until additional user information is collected, and a repeat or confirmation tool to allow the agent to re-surface prior questions or con- firm ambiguous slot values with the user. These additions would move ReacTOD closer to a fully agentic dialogue manager, capable of handling the full spectrum of conversational acts beyond state tracking.

  ReacTOD: Bounded Neuro-Symbolic Agentic NLU for Zero-Shot Dialogue State Tracking · 2026

• , Transformer-based routers, reinforcement- learning-based routing, or token-level routing) could offer different trade-offs and are not explored here.

  MMoA: An AI-Agent framework with recurrence for Memoried Mixure-of-Agent · 2026

• Satellite- terrestrial integrated edge computing networks: Architecture, challenges, and open issues.

  Beyond Scaling: Agents Are Heading to the Edge · 2026

• When training data is limited, parallel sampling compensates by exploring varied reasoning paths at inference time.

  Generative Recursive Reasoning · 2026

• Abstract The lack of standardized interfaces remains one of the main barriers to the adoption of automated planning in robotic systems.

  UPF4ROS2 for integrating planner independent task planning into ROS 2 · 2026 · DOI

• Furthermore, we present examples of architectures and discuss open problems associated with artificial intelligence (AI)-enabled solutions in relation to SoS interoperability requirements.

  AI Concepts for System of Systems Dynamic Interoperability · 2024 · DOI

• The primary gap addressed by this work is the lack of fully functional, domain-specific applications of agentic AI. Prior research has addressed autonomous reasoning frameworks, multi-agent orchestration strategies, and safety mechanisms, but few studies have implemented these principles in consumer-facing applications.

  A multi model orchestrated agent for live flight and hotel itinerary generation · 2026 · DOI

• The paper identifies several limitations, including the fact that the dataset is intentionally weakly supervised and that the model is trained on a single workstation GPU. The authors note that student-scale research is often compute-limited rather than idea-limited.

  Scaling Laws and Architectural Advances of Hierarchical JEPA (H-JEPA) Model for Planning, Control and Robotics in Physical Systems · 2026 · DOI

• The paper suggests several future research directions, including incorporating language-conditioned abstract sub-goals into H-JEPA and incorporating uncertainty into the planning process. The authors suggest that H-JEPA can be used as a bridge between latent planning and instruction following.

  Scaling Laws and Architectural Advances of Hierarchical JEPA (H-JEPA) Model for Planning, Control and Robotics in Physical Systems · 2026 · DOI

• The need for transparent and interpretable decision-making in MAPF. The lack of a unified semantic schema for MAPF.

  maPO: An Ontology for Multi-Agent Path Finding and Its Usage for Explaining Planner Behaviour · 2026 · DOI

• The gap between the expertise of planners and the rigid models embedded in existing systems. The lack of decision support tools that can keep pace with changing battlefield conditions.

  A Customizable Tactical Engine for Mission Planning with Probabilistic Inference · 2026 · DOI

• Traditional prompt engineering collapses concerns into a single unstructured buffer, making provenance, trust, topology, replayability, and access control difficult to enforce. The need for a governed knowledge unit that moves AI systems beyond the prompt-as-string paradigm.

  AIKernel Phase-1 Paper 01: ROM Format and Knowledge Snapshot Model · 2026 · DOI

• This model has several limitations. First, ROM‑level equivalence is not full semantic equivalence. It is equivalence under a declared canonical‑ ization contract. Natural‑language meaning remains broader than any deterministic canonical form. Second, canonicalization intentionally discards non‑semantic representation entropy. Prompt behaviors that depend on microscopic whitespace or formatting perturbations are not preserved. This is intentional: AIKernel treats such behaviors as unstable prompt manipulation rather than governable architecture. Third, the security model depends on correct implementation of canonicalization, hashing, signature val‑ idation, relation resolution, and trust‑store management. Bugs in those components can undermine the governance boundary. Fourth, bounded topology may reject knowledge graphs that are useful but too large or too dynamic. This trade‑off is deliberate. AIKernel prioritizes deterministic governance over unbounded context expansion. 11. Relationship to Other Phase‑1 Papers Paper 01 provides the knowledge substrate for the rest of Phase‑1. • Paper 02 governs the storage and transport layer that persists and retrieves ROM documents. • Paper 03 uses trusted knowledge snapshots as preconditions for admissibility decisions. • Paper 04 uses stable root goals and governed context as inputs to trajectory monitoring. • Papers 05 and 06 rely on ROM‑derived contracts during execution and delegation. 7 • Paper 07 validates the model through AIKernel.NET implementation. • Paper 08 integrates the Phase‑1 architecture as a unified AIOS governance model.

  AIKernel Phase-1 Paper 01: ROM Format and Knowledge Snapshot Model · 2026 · DOI

• The lack of constitutional governance infrastructure in long-term AI-assisted operational systems. The need for topology-aware operational cognition in intelligent systems.

  Persistent Operational Memory Through Knowledge Engine Architecture · 2026 · DOI

• The lack of precision and structure in natural language mission descriptions. The need to capture temporal and logical consistency in mission descriptions. The challenge of evaluating the performance of models in NL-to-FOL translation.

  A Pilot Benchmark for NL-to-FOL Translation in Planetary Exploration · 2026

• Future work should explore quantitative evaluation methods, including structural similarity, logical equivalence, and execution-based validation. The benchmark can be used to investigate the challenges of NL-to-FOL translation in planetary exploration. The development of more advanced models and prompting strategies is needed.

  A Pilot Benchmark for NL-to-FOL Translation in Planetary Exploration · 2026

• Investigating how to generate interaction data that reliably drives continual improvement. Exploring the use of scalable environments as a training substrate for general agents. Developing new methods for environment scaling and stateful learning mechanisms.

  Scalable Environments Drive Generalizable Agents · 2026

• Current scaling practices focus on collecting more experience or broader task sets under fixed benchmarks. There is a lack of systematic exposure to environments with meaningfully different executable rule-sets. Generalizable agents require environment scaling to adapt to diverse tasks and unseen environments.

  Scalable Environments Drive Generalizable Agents · 2026

• The adaptive replanning capabilities of LLMs under spatio-temporal dynamics are challenging to evaluate. The performance of LLMs degrades as task difficulty increases. The strongest frontier model achieves limited performance.

  STT-Arena: A More Realistic Environment for Tool-Using with Spatio-Temporal Dynamics · 2026

• Although STT-Arena provides a rigorous and diverse benchmark, several limitations remain. The authors do not specify the limitations in detail.

  STT-Arena: A More Realistic Environment for Tool-Using with Spatio-Temporal Dynamics · 2026

• The human preference study is constrained by time and budget, resulting in a small number of annotators and a limited sample size. Applying MA$^{2}$P to a new domain requires a warm-up phase. Most of the automatic metrics rely on an LLM evaluator.

  MA$^{2}$P: A Meta-Cognitive Autonomous Intelligent Agents Framework for Complex Persuasion · 2026

• To apply MA$^{2}$P to new domains and evaluate its effectiveness. To develop more advanced meta-cognitive configurators that can select more effective meta-strategies. To explore the use of MA$^{2}$P in other applications, such as education and behavior change.

  MA$^{2}$P: A Meta-Cognitive Autonomous Intelligent Agents Framework for Complex Persuasion · 2026

• Future work includes testing the framework on broader game styles and lower-level interfaces. Future work includes longer deployments and the use of audio cues and richer multimodal feedback.

  SPIKE: An Adaptive Dual Controller Framework for Cost-Efficient Long-Horizon Game Agents · 2026

Most-cited papers in AI-based Problem Solving and Planning

• Code as Policies: Language Model Programs for Embodied Control · 2023 · 638 citations
• ProgPrompt: Generating Situated Robot Task Plans using Large Language Models · 2023 · 510 citations
• Using cognitive psychology to understand GPT-3 · Proceedings of the National Academy of Sciences · 2023 · 443 citations
• Artificial cognition for social human–robot interaction: An implementation · Artificial Intelligence · 2016 · 316 citations
• Accelerating Innovation With Generative AI: AI-Augmented Digital Prototyping and Innovation Methods · IEEE Engineering Management Review · 2023 · 252 citations
• The increasing cost tree search for optimal multi-agent pathfinding · Artificial Intelligence · 2012 · 195 citations
• Towards a science of integrated AI and Robotics · Artificial Intelligence · 2017 · 145 citations
• If you don't want to be late, enumerate: Unpacking reduces the planning fallacy · Journal of Experimental Social Psychology · 2003 · 139 citations
• Representations for robot knowledge in the KnowRob framework · Artificial Intelligence · 2015 · 129 citations
• A probabilistic plan recognition algorithm based on plan tree grammars · Artificial Intelligence · 2009 · 109 citations

Most recent work

• The Lever Generalizes -- and It Brakes: A Late, Bidirectional Action-Commitment Lever Across Agent Decisions and Architectures · Zenodo (CERN European Organization for Nuclear Research) · 2026
• Carl-OS: A Modular Architecture for Intelligent Systems · Zenodo (CERN European Organization for Nuclear Research) · 2026
• How Well Can AI Do Strategy? Empirical Benchmarking Using Strategy Simulations · Strategy Science · 2026
• Efficient Learning of Predictive Maps for Flexible Planning · bioRxiv · 2026
• An overview of artificial intelligence approaches for automating evidence synthesis · Public Health · 2026
• The lure of automated feedback · Theory Into Practice · 2026
• Reducing Cost of LLM Agents with Trajectory Reduction · Proceedings of the ACM on software engineering. · 2026
• S2CR: A self-supervised self-consistency reasoning framework coupled to retrieval-augmented generation · Information Processing & Management · 2026
• Enhancing trust and decision-making in AI-driven construction: Applying cognitive fit theory to interface clarity and alignment · Acta Psychologica · 2026
• Using system characteristics for reasoning about complex chemical phenomena · International Journal of Science Education · 2026

Related topics in Computer Science