Loading Now

EXPSPACE and Beyond: Navigating the Extreme Frontiers of Computational Complexity in AI and Beyond

Latest 16 papers on computational complexity: Nov. 22, 2026

The world of AI and computational systems is constantly pushing the boundaries of what’s possible, but with great power comes great complexity. Understanding the inherent difficulty of problems is crucial for developing efficient algorithms and practical applications. Recent research has delved into the farthest reaches of computational complexity, exploring everything from the strategic intricacies of ancient board games to the nuanced dynamics of modern AI systems and advanced networking.

The Big Idea(s) & Core Innovations

These papers highlight a compelling trend: the need for new theoretical frameworks to truly understand and tackle computationally challenging problems. A groundbreaking insight comes from Markus Bläser and Felix Czerner (Saarland University) in their paper, Go under Chinese Rules with a Succinct Game History is Complete for Exponential Space. They reveal that a seemingly simple no-repetition rule, when coupled with a succinct game history, can elevate Go’s complexity to EXPSPACE-complete. This demonstrates how compact representations can significantly escalate the computational burden, pushing the limits of what computers can solve within practical timeframes.

In a similar vein of complexity classification, Craig Tennenhouse and Kanae Yoshiwatari tackle the longstanding open problem of Undirected Vertex Geography with a height constraint in The computational complexity of Undirected Vertex Geography with maximum height 3. They prove that UIR3 is PSPACE-complete, even for highly restricted graphs, by cleverly simulating directed edges with “diode-like” gadgets. This result provides a sharp demarcation in the game’s complexity, bridging the gap between polynomial-time and PSPACE-complete variants.

Meanwhile, Simon Mackenzie and Mashbat Suzuki (UNSW Sydney), in their work Nearly Settling the Complexity of Connected ε-Envy-Free Cake Cutting, uncover a critical transition in fair division problems. They show that while up to five agents allow for polylogarithmic query solutions, seven or more agents necessitate polynomial queries, proving UEOPL-hardness and PPAD-completeness. This highlights how slight increases in the number of agents can dramatically alter the computational feasibility of achieving fairness. Similarly, Amir Masoumzadeh, Paliath Narendran, and Padmavathi Iyer explore Expressiveness and Convertibility of Rule-Based Access Control Policies, showing that converting between certain access control policy models can be co-NP-complete, requiring global structural checks. This reveals inherent complexity in managing and migrating security policies, even with seemingly equivalent expressive power.

Not all complexity is purely theoretical; practical applications in AI/ML also present their own unique challenges. Yu-Hung Wu et al. (National Taiwan University, NVIDIA AI Technology Center), in LLM Complexity: Towards a Computational Complexity View of Large Language Model Inference, introduce “LLM Complexity,” a model-aware metric for Large Language Model (LLM) inference. Their key insight is that classical time complexity poorly predicts LLM token usage. Instead, output length and maintaining internal state are stronger drivers of inference cost. This new metric offers a more nuanced understanding of LLM difficulty, impacting curriculum learning and model evaluation.

In the realm of optimization, Kalpesh K. Patel et al. (Indian Institute of Technology Roorkee, University of Southampton) present SFiD: Self-Supervised Fitness-Driven Optimization for Next-Generation Networking. Their Self-Supervised Fitness-Driven Diffusion (SFiD) framework uses the optimization objective itself as supervisory feedback, generating its own training data. This eliminates the need for labeled expert datasets, significantly reducing inference time for problems like RIS phase-shift optimization, demonstrating how self-supervision can circumvent the computational burden of labeled data.

Under the Hood: Models, Datasets, & Benchmarks

To address these complex problems, researchers are developing and leveraging innovative models, datasets, and benchmarks:

  • LLM Complexity: Introduced by Yu-Hung Wu et al., this metric is computed as one minus normalized inference utility over multiple token budgets. It was evaluated on 12 algorithmic tasks, 10 problem sizes, and five open-weight LLMs from three model families, showing stable cross-model task rankings. The authors leveraged models like Qwen3-1.7B and Nemotron-253B and extended analysis to DeepMath and RuleArena benchmarks.
  • SFiD (Self-Supervised Fitness-Driven Diffusion): Proposed by Kalpesh K. Patel et al., SFiD includes continuous (SFiD-C) and discrete (SFiD-D, SFiD-D BNN) variants. These models were validated on De Jong and Ackley benchmark functions and applied to RIS phase-shift optimization in cell-free massive MIMO systems.
  • RegNC-Mamba3: From Zhikang Li et al. (Xidian University) in RegNC-Mamba3: Shared Structural Feature Learning Using Non-Causal State Space Principles for Multi-Modal Image Registration, this detector-free network uses a non-causal State Space Principles (NC-SSP) formulation with learnable multi-orientation Log-Gabor filters. It was evaluated on VIS-SAR (OSDataset) and VIS-IR (DroneVehicle) datasets, outperforming existing baselines in accuracy and efficiency.
  • optPASP (Optimal Probabilistic Answer Set Programming): Damiano Azzolini et al. (University of Ferrara, University of Calabria) in Optimal Probabilistic Answer Set Programming: Complexity and Implementations extend probabilistic Answer Set Programming with weak constraints. Implementations are based on Second-Level Algebraic Model Counting and Answer Set Programming with Quantifiers, applied to real-world-inspired problems like optimal reviewer assignment and patient scheduling.
  • Q-CFR (Quantum-Assisted Counterfactual Regret Minimization): Proposed by G. Maresch et al. (TU Wien, Bull GmbH, DataTeach.ai) in Quantum-Assisted Counterfactual Regret Minimization for Bounded-Payoff Imperfect-Information Games, this theoretical framework uses quantum amplitude estimation for bounded counterfactual value computation. It aims for future experiments on games like Kuhn or Leduc poker.
  • Formal Verification: Simon Mackenzie and Mashbat Suzuki provided a Lean 4 formalization for their cake-cutting query theorems and computational complexity constructions, available at https://github.com/SimonWMackenzie/connected-envy-free-lean.

Impact & The Road Ahead

These advancements have profound implications across AI, game theory, and network design. The redefinition of LLM complexity by Yu-Hung Wu et al. enables more effective curriculum learning and better resource allocation for LLM deployment. The self-supervised optimization in SFiD, as demonstrated by Kalpesh K. Patel et al., offers a paradigm shift for next-generation networking, significantly reducing inference time and removing the bottleneck of labeled data for complex optimization problems. Similarly, Zhikang Li et al.’s RegNC-Mamba3 shows a pathway to significantly more efficient and accurate multi-modal image registration, a critical component for diverse computer vision applications.

From a theoretical standpoint, the EXPSPACE-completeness of Go with succinct histories by Markus Bläser and Felix Czerner reminds us that complexity can hide in subtle rule changes, pushing the boundaries of what is computationally solvable. The insights into fair division (e.g., cake cutting) and access control policy conversions highlight fundamental limits on managing and migrating complex systems, urging developers to design with these inherent complexities in mind. The work by Hannaneh Akrami et al. on On the Theoretical Foundations of Data Exchange Economies provides a critical framework for understanding reciprocal and stable data sharing without payments, vital for the burgeoning data economy. Furthermore, Kodai Kanno et al.’s Contract-Based Decomposition of Temporal Logic Specifications for Networked Systems under Arbitrary Partitions offers a novel approach to decentralized control, making network partitioning a design variable for balancing computational load and robustness.

The road ahead will undoubtedly involve further exploration of these extreme complexity frontiers. Researchers will continue to seek new algorithmic breakthroughs, leveraging insights from quantum computing and advanced optimization techniques to tackle problems that seem insurmountable today. The ongoing work on graph problems by Jan Diester et al. (Karlsruhe Institute of Technology, Germany) in Feedback Vertex Set, Max Cut, and Edge Coloring on the Strong Product of a Graph and a Path and traffic engineering by Matthias Bentert and Stefan Schmid in The Complexity of Traffic Engineering in Segment Routing Networks illustrates how even well-understood problems can reveal surprising complexity in specialized graph structures or network configurations. This relentless pursuit of understanding and overcoming computational barriers will define the next generation of AI and intelligent systems.

Share this content:

mailbox%403x EXPSPACE and Beyond: Navigating the Extreme Frontiers of Computational Complexity in AI and Beyond
Hi there 👋

Get a roundup of the latest AI paper digests in a quick, clean weekly email.

Spread the love

Discover more from SciPapermill

Subscribe to get the latest posts sent to your email.

Post Comment

Discover more from SciPapermill

Subscribe now to keep reading and get access to the full archive.

Continue reading