Tag: #agent-planning7 results
NeSyFS is a neuro-symbolic framework designed to help LLM agents operate under partial observability. It combines knowledge-graph belief states, reactive fast thinking, uncertainty-aware slow planning, and reflection-based switching, outperforming prior methods on ALFWorld, WebShop, and ScienceWorld.
Environment Maps provide a persistent, structured graph representation that consolidates screen recordings and execution traces to mitigate errors in long-horizon agents. The framework includes Contexts, Actions, Workflows, and Tacit Knowledge. On WebArena benchmark, it achieves 28.2% success, nearly doubling baselines.
PseudoAct is a new framework that uses LLM-generated pseudocode to create structured plans for agents, encoding control flows like loops, conditionals, and parallelism. It overcomes ReAct's limitations in long-horizon tasks by reducing redundant actions and ensuring coherent execution. Experiments show 20.93% success gain on FEVER and new SOTA on HotpotQA.
HyperAgent introduces a Tool–Schema Hypergraph that models tool dependencies through required inputs and produced outputs. Its schema-aware planning and state-conditioned execution improve task completion on AppWorld while reducing redundant API calls, LLM interactions, and token usage.
ToolTree is a novel Monte Carlo tree search-inspired paradigm for LLM agent tool planning. It uses dual-feedback LLM evaluation and bidirectional pruning to explore tool trajectories adaptively. Achieves ~10% average gain on 4 benchmarks for open/closed-set tasks.
This arXiv paper explores memory strategies for spatial navigation in non-stationary environments with uncertain sensing in a foraging task. It compares simple to sophisticated agents, finding hybrid architectures with episodic memories and on-the-fly planning most efficient for exploration, search, and path optimization. Advanced agents substantially outperform minimal-memory ones as task difficulty increases, provided uncertainty is manageable.
C-JEPA extends masked joint embedding prediction to object-centric representations with object-level masking, inducing latent interventions for interaction reasoning. It boosts counterfactual VQA by 20% and enables efficient agent planning using 1% of latent features. Code is on GitHub.