InFerActive: Interactive Inference Frameworks

• InFerActive is a framework integrating active inference, interactive evaluation, and human-guided exploration underpinned by probabilistic models.
• It employs Bayesian free-energy minimization and adaptive visualization techniques to balance exploration and exploitation in complex systems.
• The framework enables scalable, real-time evaluation across domains like LLM testing, human-computer interaction, and edge computing while ensuring valid post-selection inference.

InFerActive refers to a set of frameworks, systems, and algorithmic paradigms at the intersection of interactive inference, active inference, and human-guided exploration in artificial and natural intelligence. The term encompasses a family of theoretical and practical methodologies, including probabilistically grounded agent models for perception and action (active inference), systems for interactive model evaluation and visualization, and computational workflows that intertwine exploratory tasks with rigorous statistical inference. Across diverse domains—including human-computer interaction, LLM evaluation, edge computing, and scientific data analysis—InFerActive approaches are unified by explicit modeling of uncertainty, closed-loop feedback between user/system and model, and the integration of interactive decision cycles with formal inference objectives.

1. Theoretical Foundations: Unifying Active and Interactive Inference

At its mathematical core, InFerActive draws from the free-energy principle and variational Bayesian inference, integrating these with action-perception cycles and selective inference methodologies. In active inference models, the agent maintains a generative model $p(o, s) = p(o \mid s)\,p(s)$, where $s$ are latent hidden states and $o$ are observations. The agent minimizes variational free energy

$$F[q, o] = D_{KL}[q(s) \,\|\, p(s \mid o)] - \log p(o)$$

over approximate posteriors $q(s)$, thereby performing efficient Bayesian filtering and planning under uncertainty (McGregor et al., 2015, Sedlak et al., 2023, Watson et al., 2020). Action selection further involves computing the expected free energy

$$G(a) = \mathbb{E}_{q(o',s' \mid a)}[ -\ln p(s', o') ] + H[q(s' \mid a)]$$

which accounts for both pragmatic value (reward/goal attainment) and epistemic value (information gain), realizing an intrinsic form of exploration alongside exploitation.

In inferactive data analysis, inference is performed conditional on the exploratory/interactive path taken by the analyst, enabling selective statistical guarantees that remain valid post-selection. This is formalized via conditioning on the data-analysis generative DAG (DAG–DAG) and the use of selective samplers to construct exact or approximate post-selection laws (Bi et al., 2017). Explicit randomization in selection queries increases power and tractability of inference after adaptively chosen modeling steps.

2. Interactive Evaluation and Visualization Systems

InFerActive systems generalize static evaluation protocols in favor of interactive, probabilistically controlled exploration of high-dimensional outcome spaces. "InFerActive: Towards Scalable Human Evaluation of LLMs through Interactive Inference" (Hwangbo et al., 11 Dec 2025) implements this philosophy for LLM evaluation by surfacing the full autoregressive sampling tree: every possible partial or complete output is a node, with edges weighted by conditional token probabilities. The system exposes:

• Probabilistic filtering: Top-N at each tree depth, probability mass pruning, label-based subtree collapse/expansion.
• Adaptive visualization: Big-token merging (concatenation of deterministic token chains), flex-tree layouts, Sankey-style link opacity for display proportional to probability.
• Interactive controls: Node-wise expansion, pinning, marking as Good/Bad with label inheritance, real-time calculation of cumulative probabilities, and detail-on-demand for any subtree.
• Evaluation workflow: Human annotators guide exploration, marking regions of interest or concern, while coverage and KL-divergence metrics quantify the efficiency and exhaustiveness of the interactive protocol.

Technical comparisons show order-of-magnitude greater coverage and discovery rates versus sample-based static evaluation—enabling more comprehensive, scalable, and reproducible judgment of stochastic model behaviors.

3. Active Inference Agents for Human and Machine Systems

Event-driven and neuromorphic implementations of InFerActive include:

• Human-Computer Interaction: "Interactive Inference: A Neuromorphic Theory of Human-Computer Interaction" (Vertegaal et al., 9 Feb 2025) posits the user as a Bayesian agent minimizing free energy between goal and progress distributions at each interaction step. This model predicts reaction and movement times as logarithmic functions of signal-to-noise ratio, unifying Hick's Law, Fitts' Law, and the power-law of practice within a single quantitative framework. Empirical validation in car-following tasks aligns measured driver information processing with the predicted logarithmic law.
• Edge Computing and Stream Processing: On-device controllers for stream processing instantiate the action-perception loop of active inference: the agent updates beliefs over latent states (e.g., system utilization, delay metrics) via variational inference, encodes SLOs as goal priors, and selects actions by minimizing expected free energy over future paths (Sedlak et al., 2023, Sedlak et al., 2024). Transparency and troubleshooting follow from the explicit causal models driving updates and from the agent's ability to backtrace SLO violations.

The following table summarizes selected domains and InFerActive realizations:

Domain	InFerActive Realization (Paper)	Key Features
LLM evaluation	Interactive tree exploration (Hwangbo et al., 11 Dec 2025)	Probabilistic tree visualization, adaptive filtering
Human-computer interaction	Bayesian user/event model (Vertegaal et al., 9 Feb 2025)	SNR-driven free-energy minimization, behavioral laws
Edge streaming / IoT control	On-device (A)CI agents (Sedlak et al., 2023, Sedlak et al., 2024)	Real-time free-energy loop, SLO-aware optimization
Selective data analysis	DAG–DAG/conditional inference (Bi et al., 2017)	Randomized selection, selective pivots/intervals
Image segmentation/annotation	Feature amortization, lightweight decoding	Real-time CPU pipelining, attention-based updating

4. Methodological Innovations: Algorithms and Architectural Patterns

Algorithmic kernels underlying InFerActive share several characteristics:

• Explicit generative and goal models: Agents/models characterize latent, observation, and action/control variables, and encode goal priors as explicit preferred regions or outcome distributions.
• Iterative inference and planning: Free-energy or expected free-energy functionals provide the unified objective for updating belief states (perception) and selecting control/configuration actions (planning/optimization) (McGregor et al., 2015, Watson et al., 2020, Sedlak et al., 2024).
• Probabilistic interfaces: Systems maintain and expose probability distributions (full autoregressive output trees, progress/goal overlaps, belief/posterior densities) for both algorithmic optimization and interactive exploration by users.
• Selective interaction-aware inference: In data analysis, all selection/adaptation events are explicitly conditioned upon in subsequent inference, using conditional pivots, randomized queries for power enhancement, and specialized samplers for feasible computation (Bi et al., 2017).

Notably, in UI-heavy applications (e.g., tree-based LLM or image segmentation pipelines), heavy initial computational tasks (e.g., global ViT encoding, full prompt*model tree expansion) are amortized, with subsequent low-latency interactive updates anchored in precomputed or cached features (Huang et al., 2023, Hwangbo et al., 11 Dec 2025).

5. Empirical Evaluations and Case Studies

InFerActive approaches have undergone empirical validation across diverse setups:

• In LLM evaluation (Hwangbo et al., 11 Dec 2025), computational coverage studies demonstrate >20× efficiency over random sampling for probability mass coverage; user studies show improved error discovery and evaluative confidence, especially in high-branching tree regions.
• Neuromorphic HCI studies (Vertegaal et al., 9 Feb 2025) confirm the predicted logarithmic user processing capacity and quadratic error scaling in a car-following simulation, with model fits $r^2 > 0.85$.
• Edge streaming/IoT case studies (Sedlak et al., 2023, Sedlak et al., 2024) show rapid convergence (5–30 iterations) to optimal resource configurations under dynamically shifting workloads, outperforming greedy or static policies in convergence rate and SLO adherence.
• Interactive segmentation pipelines (Huang et al., 2023) achieve sub–250 ms per click latency, supporting high-quality, low-compute pixel-level annotation on CPU-only hardware.

6. Limitations, Challenges, and Future Directions

Principal challenges for InFerActive systems include:

• Scalability: Exponentially large inference or evaluation trees require principled pruning, probabilistic filtering, and/or abstracted visualization to avoid overwhelming human or computational resources (Hwangbo et al., 11 Dec 2025).
• Automation of evaluation: While human users drive marking and exploration in current systems, integration with automated grading (e.g., via LLMs-as-judges, uncertainty propagation, or semantic metrics) is a key future direction.
• Robust online learning: In agentic settings, adaptive updating of the generative model remains sensitive to structural mis-specification or model drift; identification of failure modes for re-training or causal model revision is required (Sedlak et al., 2024, Sedlak et al., 2023).
• Broader domain-adaptation: Extending InFerActive frameworks to medical, legal, or scientific discovery tasks involves nontrivial challenges in domain-specific goal encoding, semantic grouping, and appropriate user interface abstractions.

Empirical and methodological advances suggest that InFerActive frameworks offer a unifying, probabilistically rigorous, and human-centered approach to interactive intelligence and decision-making across computational and statistical frontiers.