Pneuma-Seeker: Dynamic Intent Formalization

• Pneuma-Seeker is a paradigm that transforms evolving user intent into actionable data representations and control commands using stateful, agentic architectures.
• It employs specialized LLM-driven agents—including a Conductor, IR system, and Materializer—to iteratively update a shared state based on user feedback and retrieved evidence.
• Demonstrated successes include high convergence rates in data-centric workflows and reduced control effort in autonomous interception, validating its practical impact.

Pneuma-Seeker denotes a family of advanced computational and physical systems unified by the principle of incrementally surfacing, formalizing, and fulfilling dynamic user intent in domains ranging from data-centric workflows to autonomous guidance. The Pneuma-Seeker paradigm leverages stateful, agentic architectures—whether in LLM-driven software workflows for data discovery or in the sensor-actuator-seeker loops of time-constrained interception hardware. In its core instantiation as described in the Pneuma Project, Pneuma-Seeker operationalizes the user’s evolving information need as an explicit, manipulable shared state, advancing both interaction with structured data and physically situated control (Balaka et al., 7 Jan 2026, R et al., 24 Jun 2025).

1. Foundational Concepts and System Overview

Pneuma-Seeker is a platform that iteratively transforms poorly specified or evolving information needs into concrete, actionable representations—specifically, a relational schema $T$ and an associated query sequence $Q$. At any step, the system state is a pair $(T, Q)$, expressing the current formalization of user need as a set of tables and queries. User feedback and retrieved evidence induce a sequence of updates to this state, with the objective of materializing fit-for-purpose artifacts, such as data tables or control trajectories, precisely tailored to the latent target $(T^*, Q^*)$, which remains unknown but is converged upon through interaction (Balaka et al., 7 Jan 2026).

In autonomous guidance contexts, such as seeker-equipped interceptors, Pneuma-Seeker encompasses not only information retrieval but also the feedback-driven closure of spatial position, orientation, and control surface commands, subject to field-of-view (FOV) and physical actuation constraints (R et al., 24 Jun 2025).

2. Core Architectural Mechanisms

Pneuma-Seeker’s software instantiation is based on three architectural mechanisms:

• Context Specialization: The workflow is decomposed into specialized LLM-powered agents:
  • Conductor: orchestration, dialog management, action selection.
  • IR System: retrieval of candidate tables or documents.
  • Materializer: integration, code and query synthesis, repair.
  • Each agent receives only the information necessary to its function, substantially reducing hallucinations and prompt entropy (Balaka et al., 7 Jan 2026).
• Conductor-Style Planning: The Conductor, an LLM-driven planning agent, iteratively:
  1. Performs internal reasoning,
  2. Initiates tool calls (retrievers, materializers),
  3. Modifies the shared state $(T, Q)$,
  4. Issues user-facing messages. Execution is bounded by a planning horizon (empirically $i=5$ actions), after which user interaction resumes, supporting incremental, human-in-the-loop convergence.

• Convergence Mechanism: Central to Pneuma-Seeker is the iterative update of the shared state. At each step,

$$(T^{t+1}, Q^{t+1}) = \mathcal{U}\big((T^t, Q^t), f^t, IR^t\big)$$

where $f^t$ is user feedback and $IR^t$ is retrieved evidence. Convergence is empirically established when either the user declares the output sufficient or a task-specific distance metric $d((T^t, Q^t), (T^*, Q^*))$ is minimized.

3. Integrated Techniques in Data-Centric Pneuma-Seeker

The Pneuma-Seeker workflow tightly interweaves multiple established and novel techniques:

• Retrieval-Augmented Generation (RAG): Utilizes a hybrid index—HNSW vector store and BM25 inverted index—to efficiently return relevant database tables and documents, fusing structured schema, sample rows, and domain metadata into LLM prompts.
• Agentic Framework: Agents are dynamically orchestrated; tool outputs and query results are passed among components in a flexible, non-pipelined topology.
• Structured Data Preparation: The Materializer emits SQL (for DuckDB DDL/DML) or Python (Pandas, NumPy) scripts as dictated by required transformations. Error-driven code repair closes the loop: executor exceptions are re-injected into the prompt for correction.
• Emergent Documentation: Each $(T, Q)$ update and associated dialog forms an institutional knowledge trace, accruing accessible organizational memory.

A schematic table of architectural agents and core roles appears below:

Agent	Input Context	Core Role
Conductor	$(T, Q)$, feedback	Planning, state update
IR System	Retrieval request	Table/document retrieval
Materializer	$T$, $Q$, evidence	Integration, code synthesis

4. Quantitative and Qualitative Evaluation

Pneuma-Seeker’s performance is measured using LLM-driven user simulation (LLM Sim) on data integration benchmarks. Evaluation metrics include:

• Convergence Percentage: Fraction of tasks where $(T, Q)$ matches the latent need within 15 turns.
• Median Turns to Convergence
• Accuracy: Proportion of final answers numerically matching established benchmarks.

Reported results on KramaBench’s Archaeology and Environment domains indicate Pneuma-Seeker achieves convergence on $\sim$83% (Archaeology) and $\sim$75% (Environment), outperforming LlamaIndex and static retrievers by wide margins. The system’s accuracy of 41.7% (Archaeology) and 55.0% (Environment) substantially exceeds DS-Guru and LlamaIndex baselines (Balaka et al., 7 Jan 2026). Latency per response is $\sim$70 s and cost is \$0.27 input/\$0.01 output per 250K tokens on O4-mini.

Qualitative findings highlight Pneuma-Seeker's ability to identify schema gaps, dispatch clarifying queries, and guide the user through ambiguous or incomplete requirements—a capability not evidenced in static pipelines.

5. Pneuma-Seeker in Autonomy and Control

In autonomous interception applications, Pneuma-Seeker’s principles manifest as precise, Lyapunov-stabilized, time-constrained guidance for seeker-equipped platforms. The dynamic system state $x = (r, \theta, \psi, \theta_M, \psi_M)$—relative range and angles—evolves according to nonlinear kinematics with control via lateral acceleration components $a_{My}, a_{Mz}$, subject to input-affine saturation: $$\dot a_{Mi} = [1 - (a_{Mi}/a_{i,\max})^n ]\,b_i - \rho a_{Mi}$$ where $b_i$ is command input, $a_{i,\max}$ actuator limit, $n \geq 2$, $\rho > 0$ (R et al., 24 Jun 2025).

Guidance enforces a strict seeker FOV bound $|\sigma| \leq \sigma_{\max}$ by the construction of a virtual lead angle command $\sigma_d$, allocating reference headings and stabilizing acceleration-tracking errors via established backstepping procedures. Empirical simulation shows zero miss at planned impact time, no violation of actuator bounds, and a $\sim$50% reduction in control effort compared to classical schemes not modeling saturation.

For hardware realization, seeker FOV and actuator dynamics are mapped from sensor and aerodynamic look-up tables, and the control law is directly embedded in the flight loop, enabling robust time-of-flight-exact interception under realistic constraints.

6. Strengths, Limitations, and Future Directions

Strengths:

• LLM-driven orchestration of agents enables rapid accommodation of new retrieval and integration tools.
• Context specialization substantially minimizes LLM hallucination and ensures prompt relevance.
• Shared schema-query state provides an effective communication bridge between user intent and formalized data/workflow products.
• Institutional knowledge accumulates as documentation, expediting future queries and knowledge transfer.

Limitations:

• Latency and computational cost of LLM-driven orchestration currently preclude real-time deployment.
• System evaluation predominantly utilizes simulated (LLM-based) users; human subject studies remain necessary for validation.
• Operator action space is constrained by available tool repertoire.

Future Work:

• Integrate user feedback into retrieval mechanisms to reduce irrelevant candidate tables.
• Parallelize Conductor planning (e.g., Plan-over-Graph) for increased throughput.
• Expand Materializer with declarative semantic operator sets and interactive code debugging.
• Explore fuzzy convergence metrics for queries and active learning strategies for dynamic clarification.
• User-centered UI refinements for $(T, Q)$ visualization.

A plausible implication is that extensions of the Pneuma-Seeker architecture could further advance autonomous, intent-aligned workflows across domains where intent acquisition, context-constrained planning, and controlled convergence to actionable representations are mission-critical.

---

Key Sources:

• "The Pneuma Project: Reifying Information Needs as Relational Schemas to Automate Discovery, Guide Preparation, and Align Data with Intent" (Balaka et al., 7 Jan 2026)
• "Time-Constrained Interception of Seeker-Equipped Interceptors with Bounded Input" (R et al., 24 Jun 2025)