partial-observability

该论文系统研究了LLM规划智能体在部分可观测环境中的计划承诺长度问题，比较了激进式与保守式策略在安全性与效率之间的权衡，并提出可调的混合策略。

This paper proposes complex-valued unitary recurrent networks (uRNNs) as drop-in replacements for recurrent PPO architectures, using phase-preserving dynamics inspired by quantum state measurement to address partial observability. The authors report up to 3x the reward of baselines on memory-improvable tasks such as rocksample and Craftax, with code released publicly.

This paper proposes a risk-averse online POMDP planning method that applies CVaR to the immediate cost over the belief at each step, keeping the standard expected cumulative return objective so existing expectation-based planners can be reused. The authors prove finite-time guarantees for policy evaluation, sparse sampling, and a bound on the gap between the particle belief MDP surrogate and the original POMDP, recovering standard planning in the risk-neutral limit.

The paper introduces HIDE, a benchmark for evaluating manipulation memory under partial observability, and SEEK, a framework combining three memory mechanisms that improves task success in simulation and real-world robotic manipulation experiments.

This paper characterizes the minimal recurrent behavioral memory required for imitating an expert under partial observability, using information-theoretic measures and experimental validation.

This paper proposes SIHA, a method for inferring interaction structures in multi-agent systems when agents are hidden, combining structure-agnostic initialization with iterative refinement to improve structural inference and trajectory reconstruction.

The paper proposes the Belief-State Engine, an inference module that maintains a Bayesian posterior over hidden states to augment LLM agents for principled planning under partial observability, demonstrating improved task performance and decision consistency.

This paper introduces Counterfactual Latent World Models (CLWM) to address counterfactual collapse in world models, improving planning success in embodied reasoning under partial observability through contrastive objectives.

This paper introduces Belief-Based World Models (BB-WMs) to enhance LLM agents' decision-making under partial observability by providing direct access to beliefs about uncertain states, showing improved task performance.

This paper proposes a t-step lookahead threshold policy for Whittle index in partially observable restless bandits, proving geometric convergence to the exact index and enhancing numerical accuracy.

NeSyFS is a neuro-symbolic framework for LLM agents under partial observability that uses a knowledge graph to represent belief state, combines fast/slow thinking with uncertainty-aware planning, and reflection, showing gains on ALFWorld, Webshop, and ScienceWorld.

This paper introduces a neuro-symbolic meta-policy for temporal knowledge-graph memory in partially observable reinforcement learning, combining RDF-based graph representations with symbolic memory management heuristics to achieve inspectable and adaptive control.

This paper studies stochastic linear bandits where the agent only observes a random subset of action coordinates, proving that sublinear regret is possible when actions have low intrinsic dimension, and proposes the TOFU-POV algorithm with theoretical guarantees.

This paper formalizes deliberative collaboration for LLM agents under partial observability, introduces a scalable benchmark across multiple domains, and systematically evaluates representative LLMs, finding that complex tasks remain challenging while deliberation can enable error correction.

This paper proposes ASK+, a method for uncertainty-gated assistance from small language models (SLMs) to reinforcement learning agents in partially observable environments (POMDPs). By providing trajectory-aware context and structured chain-of-thought reasoning, ASK+ significantly improves success rates over baselines, demonstrating that prompt design and selective gating dominate model scale.

DeepMind's TacticAI, a system for analyzing football tactics, is published in Nature Communications, with potential applications in robotics and computer games due to its handling of partial observability and multimodal data.

This paper studies restless bandits with binary latent states and imperfect binary feedback, developing a partial conservation laws (PCL)-based framework for establishing indexability and computing the Whittle index, with applications to opportunistic spectrum access.

This paper models cancer treatment as a belief-space planning problem using active inference, deriving an expected free-energy objective that unifies goal-directed control and information acquisition under measurement budgets. The framework is validated on real clinical data from the AACR Project GENIE, demonstrating simultaneous patient categorization and high treatment efficacy.

This paper introduces a framework for two-sided matching with temporally extended feedback, formulating it as a partially observable Markov game with costly screening, noisy observations, and evolving latent profiles. The authors present Learn2Match, a multi-agent reinforcement learning benchmark, and show that independent PPO outperforms bandit baselines in social welfare but incurs higher information-friction loss.

SpatialWorld is a unified benchmark for evaluating interactive spatial reasoning in multimodal agents across diverse real-world tasks, revealing that even the strongest models achieve low task success rates.