graph-based belief revision

Existing belief representation formalisms—such as propositional sets or probability distributions—fail to capture the internal structure of beliefs, conflate external credibility with internal coherence, and cannot adequately model fragmented or contradictory cognitive states. Method: We propose a directed weighted graph-based belief system model: nodes represent individual beliefs, and directed edges encode cognitive relations (e.g., support, contradiction); we introduce a dual-dimensional quantification scheme—“credibility” (reflecting reliability of external sources) and “confidence” (measuring strength of internal structural support)—thereby decoupling these orthogonal dimensions. Contribution/Results: This framework transcends limitations of classical logic, probabilistic, and argumentation-based models by enabling rigorous representation of inconsistent and fragmented beliefs. Leveraging graph-theoretic tools—including weighted directed graphs, node-weight functions, and connectivity analysis—it supports fine-grained relational expression and provides formally grounded, static analyses of cognitive coherence, structural conflict, and representational boundaries.

How can cognitive states and their dynamic evolution be uniformly and computationally represented, encompassing concepts, events, perceptual inputs, and task context, while supporting belief updating and interaction with working memory? This work proposes a typed weighted graph–based framework for cognitive state representation, in which nodes may embed nested subgraphs and edges encode six semantic relation types. The framework distinguishes between a long-term belief graph and a capacity-limited working memory graph, and formally defines their interaction mechanisms. It establishes the first unified cognitive state ontology, subsuming major cognitive architectures—including ACT-R, Soar, and Global Workspace Theory—as special cases. Furthermore, it enables formal diagnosis and analysis of cognitive phenomena such as conflict, coherence, and self-processing through a toolkit of graph-theoretic operators.

This work addresses the lack of a unified architecture and formal semantic foundation in existing AI agent memory systems, which struggle to jointly support cognitive memory and versionable asset management. The authors propose Kumiho, a graph-native cognitive memory architecture that establishes a formal correspondence between AGM belief revision theory and attributed graph operations, yielding a unified memory model supporting immutable revisions, typed dependencies, and URI-based addressing. The architecture introduces three key innovations: prospective indexing, event extraction, and client-side LLM reranking. It employs a dual-storage design using Redis and Neo4j, integrating full-text and vector retrieval to enable decoupled hybrid retrieval and reasoning. Evaluated on the LoCoMo benchmark, Kumiho achieves an F1 score of 0.565 and 93.3% judgment accuracy on LoCoMo-Plus, significantly outperforming baselines such as Gemini 2.5 Pro, while allowing seamless large language model substitution without pipeline modifications.

Belief systems often exhibit global inconsistency yet support reliable local reasoning. This paper addresses the challenge of enabling sound classical logical inference over globally inconsistent symbolic knowledge graphs. Method: We introduce the notion of “reasoning regions”—high-confidence, structurally balanced subgraphs extracted from directed signed weighted graphs. Our approach decouples source credibility from structural confidence, employs a contractive confidence propagation algorithm augmented with parity-based structural balance detection, and incorporates shock-robust local updates. Greedy repair and Jaccard-based deduplication further yield compact, interpretable region atlases. Contribution/Results: The framework achieves near-linear time complexity and demonstrates strong robustness against perturbations on synthetic benchmarks. It establishes, for the first time, a computationally tractable and dynamically evolvable foundation for inconsistency-tolerant reasoning—providing both theoretical grounding and practical algorithms for identifying locally coherent fragments within globally contradictory belief systems.

This work addresses the susceptibility of large language models to noise and structural misalignment in knowledge graph reasoning, which often leads to rigid and error-prone inference. To mitigate this, the authors propose CoG, a training-free dual-process framework inspired by the dual-system theory in cognitive science—combining intuitive and analytical reasoning. CoG employs relational blueprints to impose interpretable soft constraints that guide the reasoning process and activates an evidence-driven backtracking mechanism upon detecting failure, enabling iterative refinement. This approach allows for controllable optimization of the reasoning trajectory and achieves state-of-the-art performance across three standard benchmarks, demonstrating significant improvements in both accuracy and efficiency over existing methods.

This work addresses the challenge of effectively managing contextual belief states in large language models during long-horizon interactions by formally defining, for the first time, the Contextual Belief Management (CBM) task. To enable rigorous evaluation, the authors introduce BeliefTrack, a closed-world benchmark that supports precise assessment of belief consistency in tasks such as rule discovery and circuit diagnosis. Methodologically, they propose an integrated approach combining symbolic verifiers, reinforcement learning–based reward mechanisms, and representation-level belief guidance to explicitly regulate the model’s belief states. Experimental results demonstrate that the reinforcement learning strategy reduces belief management failure rates by 70.9% on average, while representation-level intervention achieves a 46.1% reduction in failure rates across two tasks, substantially enhancing the reliability of belief dynamics in language models.

This study investigates how humans update their beliefs when confronted with qualitative AI recommendations derived from an unknown data-generating process. Drawing on a controlled behavioral experiment comprising 60,252 prior–posterior belief pairs, the authors systematically compare the performance of three prominent belief-updating models through integrated modeling and statistical analysis. The findings reveal asymmetric belief updating under extreme priors and significantly attenuated adjustments under moderate priors. Building on these patterns, the study articulates four testable properties of belief updating and identifies three archetypal behavioral profiles. Model applicability is validated at both individual and aggregate levels, offering crucial empirical foundations for AI-assisted decision-making.

This work addresses the limitations of existing graph-based memory approaches, which employ flat structures that introduce excessive irrelevant context during retrieval and struggle to efficiently capture dynamic relationships among memory units. To overcome these challenges, we propose HiGram, a hierarchical graph memory framework that organizes nodes from coarse to fine granularity, leverages MicroGraph-based path-level evidence localization, and incorporates a collaborative rewriting mechanism. This design enables precise memory updates and effective multi-hop reasoning. HiGram substantially improves memory relevance, update efficiency, and dependency consistency, outperforming current baselines in long-term conversational question answering and conflict-aware memory tasks across key metrics—including answer quality, evidence accuracy, and token efficiency.

Current agent memory systems treat writes as immediate truths, rendering them vulnerable to contamination, obsolescence, or incompleteness, which can trigger irreversible errors. This work proposes MemTX—the first transactional belief-commit protocol—that employs snapshot isolation to stage writes and validates them based on evidence, authority, provenance, and validity, committing irreversible actions only when beliefs are deemed trustworthy. MemTX further supports cascading rollback and repair mechanisms. We introduce attribute-based testing and bounded exhaustive verification—covering 5.5 million states—to formally prove the protocol’s correctness. Experimental results demonstrate that MemTX consistently outperforms eight baselines across five backbone models from three model families, achieving significant gains on four backbones, matching performance on one, and crucially inducing zero downstream harm.

Existing large language model (LLM)-driven multi-agent systems often lack explicit modeling of conflicting relationships during aggregation, leading to error propagation and unreliable reasoning. To address this limitation, this work proposes SIGMA, a novel framework that introduces signed graphs into multi-agent reasoning for the first time. SIGMA explicitly captures trust, conflict, and neutral relationships among agents by constructing a signed relational graph and incorporates a conflict-aware message-passing mechanism alongside a structure- and conflict-aware weighted aggregation strategy. Extensive experiments demonstrate that SIGMA significantly outperforms current state-of-the-art methods across six benchmark datasets, achieving higher accuracy and greater robustness against conflicting information under diverse LLM backbones and multi-agent configurations.