Science AI Journal
Computer Science · Research topic

Open research questions in AI-based Problem Solving and Planning

36 unresolved questions extracted from the limitations and future-work sections of 526 AI-based Problem Solving and Planning papers in our library. Each links back to the study that raised it.

What the literature leaves open

  • Third, our gap analysis identi(cid:28)es three persistently sparse taxonomy cells, six open problems, and a three-horizon research roadmap spanning concrete short-term actions (uni(cid:28)ed benchmarks), medium-term investments (compositional memory, formal trade-o(cid:27) models), and transformative long-term goals (multi-year stability, privacy-compliant forgetting, cross-modal transfer).

    Agent Memory Research in 2026: A Data-Driven Survey and Extended Taxonomy (v1.1.0 - Revised) · 2026 · DOI

  • Cross-domain transfer remains unexplored—can knowledge from research questions generalize to related tasks? Knowledge graph visualization approaches [26, 27] could enable users to inspect and refine accumulated domain understanding, providing transparency into how their corrections shape system behavior.

    Context-Mediated Domain Adaptation in Multi-Agent Sensemaking Systems EICS003 · 2026 · DOI

  • Large language models (LLMs) have driven rapid progress in electronic design automation (EDA), yet their application to radio-frequency (RF) circuit design remains limited by the scarcity of domain-specific datasets and standardized benchmarks.

    RF-Agent: A Practical Framework for Building Language Agents for RFIC Design · 2026

  • These results show that truthfulness and instruction filtering alone are insufficient: robust agentic RAG also requires defenses against salience-relevance decoupling.

    Salience Induction against Multi-Hop RAG Agents: Threat and Defense · 2026

  • Yet it remains unclear whether general-purpose coding agents, even with domain-specific EDA skills, can reliably execute an end-to-end RTL-to-GDS flow encompassing synthesis, physical implementation, and engineering change order (ECO) optimization.

    Can AI Agents Really Complete RTL-to-GDS? Lessons from Benchmarking Tool-Interactive EDA Workflows · 2026

  • We introduce SLEUTH, which makes this state explicit and actionable through a structured epistemic working memory: the agent maintains Confirmed Facts grounded to sources, Active Hypotheses ranked by evidence, and Open Questions that directly drive its next action.

    Track, Rank, Crack: Epistemic Working Memory Scales Multi-Hop Reasoning in Language Agents · 2026

  • Omni-modal evidence-seeking QA requires agents to answer questions whose evidence is sparsely distributed across videos, audio, images, web pages, and computation results.

    Omni-Decision: A Progressive Evidence-State Agent System for Omni-Modal QA · 2026

  • Within this framework, clinical environment scaling, the integration of tools, data, and clinical gyms, is identified as the most actionable yet underexplored direction for agents operating in PACS, EHR, and FHIR ecosystems.

    The Path to Self-Evolving Clinical Systems: Scaling Medical Agents from Assistance to Autonomy · 2026

  • A risk-constrained mechanism suppresses corrections under unreliable LLM evidence or abrupt scenario shifts and falls back to the original frozen model when evidence is insufficient.

    Open-Ended Scenario Reasoning for Specialist Model Adaptation · 2026

  • Agent Architecture Search (AAS) automates such design for text-domain agents, but has not been systematically evaluated on perceptual embodied agents through simulator rollouts.

    Automating the Design of Embodied AgentArchitectures · 2026

  • Large Language Models (LLMs) can infer latent user states from static profiles, but this reasoning becomes unreliable when profiles are sparse, and applying an LLM to billions of users is prohibitively expensive.

    ScaleToT: Generalizing Structured LLM Reasoning for Billion-Scale Low-Activity User Modeling · 2026

  • These results suggest that GUI and CLI expose different execution bottlenecks: GUI agents are limited by reliable grounded interaction over long-horizon workflows, whereas CLI agents are limited by the coverage and scalability of their skill interfaces.

    GUI vs. CLI: Execution Bottlenecks in Screen-Only and Skill-Mediated Computer-Use Agents · 2026

  • Three limitations deserve explicit statement. First, the aligned-set evaluation (F1 = 0.920) uses violation plans generated with constraint-specific directives that align DOI: 10.20517/scierxiv202606.0425.v1 14 https://www.scierxiv.com/ with monitored constraints; NOT PEER-REVIEWED this validates monitor implementation correctness but overstates deployment performance relative to the natural set (F1 = 0.582). Second, the extraction quality of the TC-IR pipeline (precision, recall, and type accuracy of constraint extraction from wiki text) was not formally evaluated against human- annotated ground truth; this is an important gap for future work. Third, experiments cover only one domain (medical planning) at pilot scale (137 plans); results should not be extrapolated to other domains without separate validation. CONCLUSIONS WikiPlanner demonstrates that temporal constraints can be automatically extracted from domain wikis and compiled into parallel runtime monitors for validating LLM agent plans. The two-stage TC-IR pipeline isolates LLM extraction uncertainty from deterministic compilation, enabling transparent, source-localised violation reporting at sub-millisecond latency. On a controlled medical planning benchmark, WikiPlanner- Full achieves F1 = 0.920 with linear latency scaling, and exhibits a super-additive interaction between wiki-derived constraints and sequential monitoring (+ 0.132 beyond additive). On natural plans without vocabulary alignment, performance drops to F1 = 0.582, identifying action grounding as the primary deployment challenge. The paper also corrects a formal error in an earlier version of the MutualExclusion LTLf template and consolidates the Response and MandatoryFollowup classes at the formula level while preserving their clinical semantic distinction in TC-IR. Four failure mode classes are identified: vocabulary alignment, schema insufficiency (no bidirectional temporal windows), temporal granularity mismatch, and missing constraint provenance. These constitute a concrete research agenda for future iterations of the wiki-to-monitor pipeline. Future directions include: formal evaluation of TC-IR extraction quality against human- annotated wiki corpora; extension of the TC-IR schema to support bidirectional temporal windows and exception conditions; robust semantic action normalisation using embedding retrieval; multi-domain evaluation including industrial process control; and iterative multi-pass repair for ordering violations. DECLARATIONS DOI: 10.20517/scierxiv202606.0425.v1 15 https://www.scierxiv.com/ NOT PEER-REVIEWED Authors’ contributions The author contributed solely to the article.

    WikiPlanner: Wiki-Derived Temporal Monitors for Safe LLM Agent Planning · 2026 · DOI

  • Agent skills -- structured, reusable knowledge artifacts that augment LLM agent capabilities -- have been rapidly adopted in industry, yet their cross-domain impact and use across commercial and open-source models remain under-studied, and no reusable methodology exists for evaluating an individual skill.

    A Framework for Evaluating Agentic Skills at Scale · 2026

  • While multi-agent systems (MAS) offer collective reasoning and topology-aware orchestration, capabilities naturally suited for graph-structured tasks, their application to dynamic graphs remains unexplored.

    AdaSTORM: Scaling LLM Reasoning on Dynamic Graphs via Adaptive Spatio-Temporal Multi-Agent Collaboration · 2026

  • MetaForge factorizes agentic behavior into four coupled stages: Decide (judging whether tool use is warranted), Retrieve (selecting suitable tools), Adapt (grounding tool parameters in task context), and Forge (synthesizing new skills online and recycling them into the tool library for reuse), forming a closed judge-retrieve-adapt-forge-recycle loop.

    MetaForge: A Self-Evolving Multimodal Agent that Retrieves, Adapts, and Forges Tools On Demand · 2026

  • This is insufficient for practical RTL design, where useful Verilog must be correct, synthesizable, timing-conscious, and friendly to downstream hardware objectives.

    Verilog-Evolve: Feedback-Driven and Skill-Evolving Verilog Generation · 2026

  • Existing generation methods based on large language models (LLMs) often fall short when handling complex CAD models characterized by long operation sequences, diverse operation types, and strong geometric constraints, primarily because reasoning chains break and effective error-correction mechanisms are lacking.

    Memory-Augmented Reinforcement Learning Agent for CAD Generation · 2026

  • We presented a state-centric formulation of generalized planning in which models learn to predict successor states rather than action sequences. When combined with size- and permutation-invariant relational embeddings, this approach enables compact transition models (∼1–2M parameters, no data augmentation) to achieve strong extrapolation perfor- mance in locally factored domains, matching or exceeding Transformer baselines (∼25–220M parameters) that rely on orders of magnitude more data and parameters. Empirically, our results show that explicit transition modeling provides a stronger inductive bias for extrapolation than architectural scale alone, though limitations remain in domains with hier- archical and long-range dependencies. The neuro-symbolic decoding interface further improves robustness by enforc- ing symbolic validity at every planning step. Future work will target hierarchical and long-range dependency domains, where one-step state prediction fails under strict extrapola- tion. We will extend the framework to multi-step or abstract transitions while preserving symbolic verification. 701 Hochreiter, S.; and Schmidhuber, J. 1997. Long short-term memory. Neural computation, 9(8): 1735–1780. Howey, R.; Long, D.; and Fox, M. 2004. VAL: Automatic plan validation, continuous effects and mixed initiative plan- ning using PDDL. In 16th IEEE International Conference on Tools with Artificial Intelligence, 294–301. IEEE. Kambhampati, S.; Valmeekam, K.; Guan, L.; Verma, M.; Stechly, K.; Bhambri, S.; Saldyt, L. P.; and Murthy, A. B. 2024. Position: LLMs can’t plan, but can help planning in LLM-modulo frameworks. In Forty-first International Con- ference on Machine Learning. Liu, B.; Jiang, Y.; Zhang, X.; Liu, Q.; Zhang, S.; Biswas, J.; and Stone, P. 2023. Llm+ p: Empowering large language models with optimal planning proficiency. arXiv preprint arXiv:2304.11477. Pallagani, V.; Muppasani, B.; Murugesan, K.; Rossi, F.; Horesh, L.; Srivastava, B.; Fabiano, F.; and Loreggia, A. 2022. Plansformer: Generating symbolic plans using trans- formers. arXiv preprint arXiv:2212.08681. Rivlin, O.; Hazan, T.; and Karpas, E. 2020. Generalized planning with deep reinforcement learning. arXiv preprint arXiv:2005.02305. Rossetti, N.; Tummolo, M.; Gerevini, A. E.; Olivato, M.; Putelli, L.; and Serina, I. 2024a. Enhancing GPT-based plan- In Interna- ning policies by model-based plan validation. tional Conference on Neural-Symbolic Learning and Rea- soning, 328–337. Springer. Rossetti, N.; Tummolo, M.; Gerevini, A. E.; Putelli, L.; Se- rina, I.; Chiari, M.; and Olivato, M. 2024b. Learning general policies for planning through GPT models. In Proceedings of the International Conference on Automated Planning and Scheduling, volume 34, 500–508. Shen, W.; Trevizan, F.; and Thi´ebaux, S. 2020. Learning domain-independent planning heuristics with hypergraph networks. In Proc. of the International Conference on Auto- mated Planning and Scheduling, volume 30, 574–584. St˚ahlberg, S.; Bonet, B.; and Geffner, H. 2022. Learning generalized policies without supervision using gnns. arXiv preprint arXiv:2205.06002. Toyer, S.; Trevizan, F.; Thi´ebaux, S.; and Xie, L. 2018. Ac- tion schema networks: Generalised policies with deep learn- ing. In Proc. of the AAAI Conference on Artificial Intelli- gence, volume 32.

    On Sample-Efficient Generalized Planning via Learned Transition Models · 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

  • Future work will address these limitations through independent domain-expert evaluation, retrieval- only and classical anomaly-detection baselines, stronger graph-reasoning variants, richer temporal representations, and external benchmarks when suitable knowledge resources can be constructed.

    Enhancing LLM inference with human expert knowledge: a case study on multi-agent robotics fault diagnosis and prediction · 2026 · DOI

  • Although abstraction is a popular approach to explana- tion, until now there has been limited evidence that humans utilise abstraction to explain plan quality differences.

    How Humans Explain the Difference in the Quality of Plans – A User Study · 2026 · DOI

  • However, progress is limited by the lack of high-coverage, long-horizon interaction trajectories collected from element-rich and rapidly evolving apps.

    SEE: Structure-aware Exploring \& Exploiting for Long-horizon GUI Agent Trajectory Synthesis · 2026

  • However, this text-based paradigm rarely incorporates internal mechanistic signals, leaving how retrieved memory is actually utilized during task execution underexplored.

    Mechanistic Attention Guidance for Agent Memory Refinement · 2026

  • In contrast, real-world applications often involve multilingual inputs and outputs within a unified workflow, yet the interaction between multilinguality and agentic execution remains underexplored.

    PolyWorkBench: Benchmarking Multilingual Long-Horizon LLM Agents · 2026

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

Most recent work

Find a gap in your own AI-based Problem Solving and Planning sub-topic

This page shows what the AI-based Problem Solving and Planning literature already flags as unresolved. To narrow it to your specific question, run the guided finder — it searches the gap library on demand and checks candidates against 250M+ OpenAlex works.

Open the Research Gap Finder →

Related topics in Computer Science

Online Learning and Analytics
363 open questions
AI in Service Interactions
278 open questions
Teaching and Learning Programming
239 open questions
Digital literacy in education
230 open questions
Mobile Learning in Education
183 open questions
Virtual Reality Applications and Impacts
178 open questions
Computational Drug Discovery Methods
126 open questions
Blockchain Technology Applications and Security
121 open questions
36 open questions have been extracted from the limitations and future-work passages of 526 AI-based Problem Solving and Planning papers in our library. Each one below links back to the study that raised it, so you can read the original claim in context.

Tools for your next paper

Gaps & Literature SearchFind open questions in any field — 120,000+ open gaps mined from our own library.Pre-CheckIs your paper ready? A Tier 1–5 acceptance probability in seconds.Journal FinderA ranked shortlist of target journals from a 17,500-venue index.Duplicate Publication CheckerCheck prior-publication & salami-slicing risk across 8 sources.Citation GeneratorFreeA DOI or title → APA, MLA, Vancouver, Chicago, BibTeX, RIS. No signup.Graphical Abstract MakerFreeTurn your findings into a colorblind-safe graphical abstract.

Compare the category — Honest roundups of the AI research tools, ours listed alongside the alternatives.

Command palette

Jump anywhere, run any action.