Sokoban
PulseAugur coverage of Sokoban — every cluster mentioning Sokoban across labs, papers, and developer communities, ranked by signal.
4 day(s) with sentiment data
-
New 'Follow the Winners' algorithm enhances LLM reinforcement learning
Researchers have introduced "Follow the Winners" (FTW), a novel critic-free reinforcement fine-tuning algorithm designed for agentic large language models. Unlike existing GRPO-style methods that rely on impractical rep…
-
New planner tackles complex Multi-Agent Sokoban challenges
Researchers have developed a new planner called Sokoban-LaCAM that can efficiently solve instances of the Multi-Agent Sokoban game, which involves agents pushing boxes to target locations. This approach leverages recent…
-
Recurrent AI models achieve better performance with fewer layers
A new research paper explores the optimal allocation of computational resources in AI models for streaming tasks. The study, which varied within-step depth, expert width, and the number of parallel experts, found that t…
-
New L-ICL technique enhances LLM planning accuracy
Researchers have developed a new technique called Localized In-Context Learning (L-ICL) to improve the planning capabilities of large language models (LLMs). This method involves iteratively augmenting instructions with…
-
New AI method discovers reusable code primitives for game content generation
Researchers have developed a new method called Continual Abstraction Discovery (CAD) to improve the generation of procedural content for video games. This technique leverages large language models to evolve Python progr…
-
Diffusion model predicts Sokoban puzzle solvability without explicit training
Researchers have developed a transformer-based diffusion model that can predict the solvability of Sokoban puzzles with 77.4% accuracy, despite being trained solely on tile completion without explicit solvability labels…
-
Classic 1980s puzzle game Sokoban gets an AI upgrade
Sokoban, a classic 1980s puzzle game, has been reimagined with an AI component that assists players in moving boxes. The game, characterized by its pixelated graphics, offers a challenging yet engaging experience for pu…
-
Sokoban AI Solver uses A* search for optimal puzzle solutions
A developer has created an AI solver for the classic Sokoban puzzle game, which involves pushing boxes onto designated goals. The solver, implemented in JavaScript and C++, utilizes an A* search algorithm with optimizat…
-
Answer Set Programming enhanced with new semantics and LLM-driven optimization · 2 sources tracked
Two new research papers explore advancements in Answer Set Programming (ASP). The first paper introduces a unified logical framework, Bound-Founded Semantics, to characterize various semantics for ASP extensions with li…
-
New HDR framework boosts multi-step reasoning in video models
Researchers have introduced HDR (Hierarchical Denoising for Visual Reasoning), a novel framework designed to enhance multi-step reasoning capabilities in video foundation models. HDR employs a hierarchical latent struct…
-
New "cake" representation and PRP method generate diverse game levels
Researchers have introduced a new domain-independent "cake" representation for game levels over time, designed to implicitly encode dynamic information. This representation is used with a novel level generation approach…
-
AI model for Sokoban game uses 'path channels' for planning
Researchers have partially reverse-engineered a convolutional recurrent neural network (RNN) used for the game Sokoban. They discovered that the network stores future moves, or plans, as activations within specific "pat…
-
New WA* framework achieves zero-shot generalization in AI planning
Researchers have developed a novel self-improving planning framework called WA* that combines a value heuristic represented by a Relational Graph Neural Network with Q-learning. This approach guides search and uses the …
-
Generative models achieve high-quality plan generation via self-improvement
Researchers have developed a self-improvement technique for generative models to produce high-quality plans more efficiently. This method involves fine-tuning an initial model with improved plans generated through a com…