State-Centric Retrieval

• State-centric retrieval is a paradigm that maps queries and data to explicit or latent states, enabling robust reasoning and precise ranking.
• It unifies symbolic, neural, and embodied approaches by employing logical reconstruction, matrix-valued embeddings, and spatio-temporal tuples to handle diverse retrieval tasks.
• Empirical evaluations show enhanced efficiency and performance in causal reasoning and memory management, despite challenges in scalability and default assumption handling.

State-centric retrieval is a retrieval paradigm in which the fundamental units of reasoning, matching, and ranking are explicit or latent “states” summarizing the informational content of a record, sequence, or memory trace. Unlike traditional information retrieval, which operates over surface forms or shallow embeddings, state-centric systems explicitly reconstruct, manipulate, or score states—whether system states, world states, or compact neural representations—according to downstream inference needs. This paradigm is instantiated across symbolic (action-centric), neural (vector-state-based), and embodied (spatio-temporal) settings, unifying retrieval and reasoning across diverse domains.

1. Formal Definitions and Conceptual Foundations

State-centric retrieval shifts the retrieval objective from mere content similarity or term overlap to reasoning about the states induced or represented by documents or memory records. The core principle is to map queries and data to a state space—either explicitly as logical world states, or as latent vectors in neural models—and to select or compose those document-states that most effectively reduce uncertainty or fulfill a query’s goals.

Symbolic settings utilize explicit state variables—e.g., fluents and actions in action languages—to reconstruct world states after narratives of events, enabling queries about the indirect or causal consequences of those events (Balduccini et al., 2019). In neural settings, states are high-dimensional vector summaries of content (e.g., RWKV model states, SSM states), encoding the essential information required for future inference or answer generation (Hou et al., 10 Jan 2026, Becker et al., 13 Jun 2025). In embodied robotics, states are tuples encapsulating observed attributes, locations, and time, unifying spatial, temporal, and attribute information for retrieval and action (Chen et al., 18 Nov 2025).

A state-centric retrieval query thus asks: “Given a representation of the current or desired state, which subset, mixture, or sequence of memory records (or documents) must be retrieved so that model uncertainty about the target answer or goal is minimized?”

2. Symbolic State-Centric Retrieval: Action-Centered Formalization

The symbolic instantiation, typified by “action-centered” (or “state-centric”) IR, was systematically formalized using action languages such as 𝒜ℒ_IR (Balduccini et al., 2019). In this framework:

• States are sets of extended literals {f, ¬f, u(f)}, where u(f) denotes “unknown” status of fluent f, closed under state constraints.
• Documents are narratives: sequences of actions, initial states (with defaults), and dynamic/static laws.
• Queries ask about the truth of a “fluent” after the narrative—e.g., whether a particular condition holds.

The retrieval process proceeds as follows:

1. Logical Reconstruction: Compile the narrative and query into an Answer Set Programming (ASP) model, encoding actions, state transitions, non-determinism, state constraints, and default reasoning.
2. State Simulation: Generate answer sets representing possible event histories and resulting states.
3. Matching and Ranking: For each source, search for minimal assumption (defaults and non-det splits) completions that entail (or preclude) the query at the relevant timepoint, assigning a “semantic score” proportional to the modeling cost (minimal set of contextual assumptions).
4. Principled Scoring: Only those documents whose implied states entail the desired post-condition (and not simply due to arbitrary forced assumptions) are matched and ranked.

This approach enables robust handling of indirect causes, inertia, defaults, and uncertainty, surpassing surface-level retrieval, as demonstrated in controlled experiments (Balduccini et al., 2019). However, scalability is constrained by the complexity of symbolic state expansion and the absence of large public benchmarks for real-world narrative event-to-state IR.

3. Neural State-Centric Retrieval: Vector-State Models and Efficient RAG

Neural state-centric retrieval redefines records as collections of learned, reusable latent states, enabling efficient retrieval and reranking in large-scale retrieval-augmented generation (RAG) (Hou et al., 10 Jan 2026, Becker et al., 13 Jun 2025).

EmbeddingRWKV: Unified State Representation and Retrieval

Empirically, training a single backbone RWKV (recurrent, matrix-valued) model to produce both dense retrieval embeddings and reusable, per-layer document states, results in:

• Each layer $\ell$ outputs a matrix state $S_t^{(\ell)}\in\mathbb R^{d\times d}$ that can be cached per document at indexing time.
• Reranking is performed by initializing the model’s hidden state with cached $S_d^{(\ell)}$ and processing only the query tokens, decoupling computational cost from document length.

Uniform layer selection techniques show that using just a fraction ($\sim 25\%$) of layers for state caching preserves >98% of full-model retrieval quality. This architecture yields 5.4×–44.8× speedups in reranking throughput and drastically reduces memory requirements compared to Transformer-based two-stage RAG (Hou et al., 10 Jan 2026).

Retrieval In-Context Optimization (RICO): Model-Aware Gradient-Based Selection

RICO frames retrieval as a test-time optimization over state mixtures (Becker et al., 13 Jun 2025):

• Each document is preprocessed into a fixed vector state $h_{d_i}$.
• For a given query $q$, assign weights $\alpha_i$ to each document, forming a mixture state $\bar{h}(\alpha) = \sum_i \alpha_i h_{d_i}$.
• The retrieval loss $\mathcal L(\alpha)$ is defined as the negative log-likelihood of the query under the model conditioned on this mixture plus the query state.
• Gradients of $\mathcal L$ with respect to $\alpha$ guide an optimization (e.g., SGD, AdamW) that determines which documents most reduce answer uncertainty.

RICO demonstrates that ranking documents by $h_{d_i}^\top(-\nabla_{\bar h}\mathcal L)$ is a theoretically sound and empirically effective stand-in for full leave-one-out evaluation, requiring no retriever fine-tuning and interpolating between continuous state mixing and discrete top-k selection. Experiments confirm that this model-aware retrieval often matches or exceeds the generation F1 and nDCG@10 metrics of strong baselines (BM25, E5) (Becker et al., 13 Jun 2025).

4. Embodied and Spatio-Temporal State-Centric Retrieval

In robotics, state-centric retrieval underpins frameworks where autonomous systems must retrieve or localize entities based on dynamically evolving spatial, temporal, and attribute-based memory (Chen et al., 18 Nov 2025). The STAR ("SpatioTemporal Active Retrieval") framework exemplifies this approach:

• State Tuples: Each memory entry is a tuple $s = (x, t, \alpha)$: position in robot-frame ($x$), timestamp ($t$), and attribute embedding ($\alpha$), typically produced via a vision-LLM.
• Query Parsing: Natural language instructions are parsed into attribute, spatial, and temporal constraints.
• Joint Retrieval: The system retrieves all states whose components jointly satisfy parsed predicates, enabling resolution of queries such as “the red mug that was on the table yesterday.”
• Unified Memory-Action Loops: STAR treats memory queries as actions within a policy loop, co-mingling retrieval and manipulation/navigational actions as choices for the agent.
• Reinforcement Learning (optional): Policies can be optimized to maximize retrieval success and minimize steps, formalized by policy-gradient methods or memory-specific losses.

Empirical evaluation on STARBench and real-world deployment demonstrates that STAR substantially outperforms scene-graph and recall-only baselines, especially in interactive, attribute-rich, and spatio-temporal tasks (Chen et al., 18 Nov 2025).

5. Comparative Analysis and Efficiency Gains

State-centric retrieval methods introduce substantial efficiency and effectiveness improvements across domains:

Paradigm	Core State Formalism	Main Efficiency Gain
Action-centered (ASP) (Balduccini et al., 2019)	Symbolic literals/states	State-driven reasoning, precise rank
EmbeddingRWKV (Hou et al., 10 Jan 2026)	Matrix-valued neural states	Decoupling reranking from doc length
RICO (Becker et al., 13 Jun 2025)	State mixture, model gradients	Model-aware retrieval, no tuning
STAR (Chen et al., 18 Nov 2025)	(pos, time, attr) tuples	Unified retrieval/action/policy

Neural and hybrid approaches (RWKV, RICO) allow offline state caching and rapid online ranking with minimal re-computation, while symbolic methods offer exact causal inference albeit with scalability challenges. The embodied approach tightly integrates retrieval with physical action selection, enabling rapid adaptation to open-world, attribute-rich searches.

6. Limitations, Open Directions, and Practical Implications

Significant challenges and future opportunities for state-centric retrieval include:

• Symbolic systems: Scalability remains bounded by branching over defaults and non-determinism; robust NLP-to-logic translation for real texts is required (Balduccini et al., 2019).
• Neural models: The optimal set of layers/states to cache for maximal efficiency remains an open research problem, as does the question of information redundancy and compression across depth (Hou et al., 10 Jan 2026).
• Gradient-based retrieval: The use of richer, potentially supervised proxy losses could further align retrieval with answer quality; adaptation of mixture weights during multi-hop reasoning remains an active area (Becker et al., 13 Jun 2025).
• Embodied agents: As recorded memories grow, strategies for staleness management, subroutine abstraction, and co-training of perception and memory modules are needed (Chen et al., 18 Nov 2025).
• Broader applications: Extensions include fact-checking systems, RL value models, memory-augmented verifiers, and scalable state-driven knowledge access.

State-centric retrieval thus provides a principled, domain-agnostic, and highly efficient bridge between memory, inference, and action, supporting complex querying and reasoning in symbolic, neural, and embodied AI systems.