PACT: Diverse Computational Frameworks

• PACT is a multi-faceted framework integrating ML, computational imaging, and security, designed to reduce computational burden while preserving accuracy in diverse applications.
• It employs techniques like token pruning and clustering in vision-language models, achieving up to 71% token reduction with minor accuracy loss and scalable throughput gains.
• PACT also extends to quantization, active learning controllers, photoacoustic CT, and privacy-preserving contact tracing, offering practical benefits in efficiency, memory savings, and adaptive learning.

PACT refers to a wide range of methods, frameworks, and technologies spanning machine learning, computational imaging, systems, agent-based programming, and privacy/security. For an arXiv-oriented, technical audience, the most prominent usages of "PACT" map to algorithmic techniques in token reduction for vision-LLMs, activated quantization for deep networks, advanced frameworks in photoacoustic computed tomography, active learning controllers for security, model merging, contract-theoretic pricing for AI services, incentive-aware agentic choreography, and several others. Below follows a detailed, multi-domain, encyclopedia article.

1. Pruning and Clustering-Based Token Reduction for Visual LLMs

PACT ("Pruning and Clustering-Based Token Reduction for Faster Visual LLMs") is an inference-time framework designed to reduce the computational—and memory—burden of processing visual inputs in large multimodal transformer models (Dhouib et al., 11 Apr 2025). Current Visual LLMs (VLMs) process hundreds or thousands of "visual tokens" (e.g., patch embeddings), many of which are either redundant or uninformative for model outputs. This inflates self-attention cost and GPU consumption.

PACT executes two tightly integrated steps at an early transformer layer:

• Pruning via EUTI: EUTI computes a light importance score per token using global query-key dot-products and hidden-state norms, retaining only the tokens above a percentile threshold.
• Merging via DBDPC: The Distance-Bounded Density Peak Clustering algorithm forms dense token clusters, enforcing a strict distance cutoff so all cluster members are similar and spatially proximate. Pruned tokens near cluster centers are optionally recovered.

These reductions are constructed to avoid materializing expensive attention score maps—maintaining compatibility with FlashAttention—and are insensitive to positional bias by selecting when in the transformer stack to reduce. The overall process preserves model accuracy (≤1% drop for ≤50% token reduction; ≤1.4% for up to 71.3% reduction), yields 2–3× throughput gains, and approximately 30% memory savings, outperforming prior token-pruning/merging schemes.

Model / Metric	No Reduction	PACT	Best Prior (FastV/ToME)
Visual Token Reduction	0%	71%	50–40%
LLM Throughput Ratio	100%	225%	165–137%
Accuracy Drop @70%	—	1.4%	4.4%+

Pruning operated solely on visual tokens (not text-conditioned), and DBDPC clustering presents quadratic overhead at the reduction layer, albeit with parallel GPU acceleration. Future directions include learning-based cluster heads, text-aware pruning, end-to-end differentiable pipeline integration, and adaptation to architectures delivering more diverse token representations in early layers (Dhouib et al., 11 Apr 2025).

2. Parameterized Clipping Activation for Quantized Neural Networks

PACT ("Parameterized Clipping Activation" for quantization) introduces a trainable, per-layer clipping threshold for neural network activations, enabling effective quantization to ultra-low precision (2–4 bits) with negligible accuracy loss (Choi et al., 2018). Each activation passes through a ReLU-style clamp with learned upper bound α (trainable), followed by uniform quantization within [0, α], e.g., $y = q(clip(x, 0, \alpha))$. The gradient with respect to α is managed via a straight-through estimator with an L₂ penalty regularizer to prevent α explosion.

This method enables:

• Mixed precision quantization—arbitrary weights/activation precision on most popular network families.
• Linear scaling of quantization step size with the learned α.
• Drop-in compatibility with standard training regimes (SGD variants, batch normalization, etc.).

On ImageNet with ResNet-18, PACT 4-bit weight/activation models lose <0.3% accuracy versus 32-bit baseline. Hardware benefits arise from memory, MAC-area, and data-movement reductions, yielding improvements of up to 2–4× in TOPS/Watt for custom accelerators.

3. Pareto-Aware Active Learning Controller for Security Operations

PACT in the context of Security Operations Centers (SOC) denotes a Pareto-aware controller for triggered active learning (Ndichu et al., 21 May 2026). Operating atop a frozen XGBoost-Focal screener, it employs:

• An adaptive, unsupervised ADWIN-based score-shift trigger to identify model drift in low-prevalence alert streams.
• A hybrid query policy (boundary-uncertainty + high-score sampling) to efficiently select samples for analyst labeling post-trigger.
• A warm-start update (10 trees per trigger) for label-efficient adaptation.

Evaluated on public low-prevalence datasets (AIT-ADS, BOTSv1), PACT reduces benign-normalized FP burden by 21–43% compared to a static baseline and uses up to 5.2× fewer analyst queries than periodic uniform updating, with only modest recall loss. Pure threshold tightening achieves lower FP but with unacceptably poor recall, especially under bursty attack windows.

4. PACT in Photoacoustic Computed Tomography (PACT Imaging)

PACT refers to "Photoacoustic Computed Tomography," a hybrid imaging modality integrating optical absorption contrast with ultrasonic spatial resolution (Fatima et al., 2019, Poudel et al., 2019, Cam et al., 2023, Cao et al., 2023). Key methodological and system engineering aspects include:

• System Cost & Design: System cost $C_{total} = C_{opt} + C_{det} + C_{DAQ} + C_{SW}$ with design tradeoffs between optical source, detector count, DAQ channels, and open source reconstruction. Solutions range from simple LED-based systems (<$5k, shallow imaging) to complex, high-speed DAQ-backed arrays (>$80k, deep imaging) (Fatima et al., 2019).
• Physics & Inverse Problem: Initial acoustic pressure $p_0(r)$ is related via the acoustic wave equation to measured transient signals at the detectors; analytic inversion is feasible only under idealized conditions. For practical, heterogeneous media or incomplete data, iterative reconstructions are prevalent (Poudel et al., 2019).
• Dynamic and 3D Imaging: Spatiotemporal low-rank regularized reconstructions (e.g., LRME-STIR) enable 4D PACT even with under-sampled, rotating-gantry hardware by exploiting temporal coherence in dynamic scans (Cam et al., 2023). Densely packed arrays with 3072 channels enable single-shot 3D imaging through tissue and skull, supporting real-time, in vivo neuroimaging (Cao et al., 2023).

5. Contract-Theoretic Pricing and Agentic Language Ecosystems

PACT encompasses frameworks for complex, multi-agent, and economic settings including:

• Contract-Theoretic Pricing for AI Services: PACT models multi-dimensional Quality of Service (QoS)—objective (latency) and subjective (satisfaction)—for agentic AI service provisioning (Yang et al., 27 May 2025). It explicitly incorporates computational, hardware, and liability costs in the provider's optimization, with contract menus designed to be incentive compatible and individually rational under user-type asymmetry. Users self-select their desired QoS–price point, maximizing provider profit under information asymmetry by solving a classic screening problem.
• Choreographic Language for Agentic Ecosystems: Pact extends deadlock-free choreographic programming with agent choices, preferences, and exogenous variables, translating protocols into formal extensive-form games (Gopinathan et al., 4 May 2026). Behavioral strategies and equilibria for self-interested, partially informed agents are computed via a bounded-rational, recursive solver, providing incentive-aware, correct-by-construction interaction code for open multi-agent LLM systems.

6. Diverse Applications: Model Merging, Robotics, Tool Use, and Compliance

PACT additionally denotes frameworks and algorithms such as:

• Preserving Anchored Cores in Task-vectors: In model merging, PACT corrects the failure of task-vector methods to account for task-critical knowledge that remains embedded ("Load-Bearing Wall" dimensions) in pre-trained weights by explicitly shielding these subspaces before merging task vectors, using SVD/QR geometry, yielding state-of-the-art multi-task model performance (Shi et al., 17 Jun 2026).
• Perception-Action Causal Transformer (robotics): PACT pre-trains a causal Transformer on interleaved sequences of high-dimensional sensory states and actions. Downstream heads for navigation, localization, or mapping are fine-tuned efficiently, outperforming task-from-scratch learning (Bonatti et al., 2022).
• Privileged Trace Co-Training for Tool-Use Agents: In multi-turn tool-use, PACT leverages expert traces only as optimization hints—component-aware SFT and trace-conditioned RL surrogates—while keeping inference prompt-only, thus balancing dense supervision with generalization (Du et al., 15 Jun 2026).
• Human-Robot Collaboration: PACT formalizes ask-or-act clarification, combining RL-learned policies with cross-day history for continual, proactive task assistance and introduces "clarification utility" as a fundamental assessment metric (He et al., 23 May 2026).
• Enterprise Privacy Artifact Linking: The Privacy Artifact ConnecTor builds dense, cross-type embedding graphs for privacy compliance, using contrastive learning and fine-tuned DRAGON encoders; LLMs traverse this graph to dramatically improve artifact discovery and compliance query match (Fang et al., 23 Jul 2025).
• Physical Safety Alignment in Diffusion Policies: PACT post-trains robotic diffusion policies to satisfy physical constraints via constraint-tilted teacher distillation, theoretically bounded curriculum, and dense supervision, achieving substantial reductions in safety violations and task failures (Wu et al., 7 Jun 2026).
• Peak-Aware Cross-Attention Graph Transformer: PACT enables efficient, accurate storm-surge prediction by combining graph-based spatial coding, station-specific cross-attention, and tail-focused loss strategies, substantially improving both average and extreme-event prediction (Liu et al., 9 May 2026).
• Part-Decomposed Articulated Object Generation: PAct (note capitalization) achieves rapid, single-image generation of part-level articulated 3D assets by learning explicit part-level latent codes and part-aware diffusion (Liu et al., 16 Feb 2026).

7. Privacy and Contact Tracing

PACT in mobile contact tracing refers to a third-party–free, privacy-preserving architecture for proximity-based epidemic tracing (Chan et al., 2020). This system emphasizes:

• Ephemeral, unlinkable, pseudorandom Bluetooth IDs, stored locally and revealed only upon user-consented positive testing.
• No global mapping from IDs to real identities, minimal metadata exposure, and robust formal privacy guarantees (anonymity, unlinkability, bounded de-anonymization risk).
• Support for privacy-sensitive exposure notifications, contact interviews, and narrowcast health advisories, all without centralized collection of raw user data.

• PACT: Pruning and Clustering-Based Token Reduction (Dhouib et al., 11 Apr 2025)
• PACT: Parameterized Clipping Activation for Quantization (Choi et al., 2018)
• PACT: Pareto-Aware Controller for Alert Fatigue (Ndichu et al., 21 May 2026)
• PACT: Review and Advances in Photoacoustic Computed Tomography (Fatima et al., 2019, Poudel et al., 2019, Cam et al., 2023, Cao et al., 2023)
• PACT: Contract-Theoretic Pricing (Yang et al., 27 May 2025)
• Pact: Choreographic Language for Agents (Gopinathan et al., 4 May 2026)
• PACT: Anchored Core Model Merging (Shi et al., 17 Jun 2026)
• PACT: Perception-Action Causal Transformer (Bonatti et al., 2022)
• PACT: Privileged Trace Co-Training (Du et al., 15 Jun 2026)
• PACT: Human-Robot Continual Assistance (He et al., 23 May 2026)
• PACT: Privacy Artifact ConnecTor (Fang et al., 23 Jul 2025)
• PACT: Physical Safety Alignment (Wu et al., 7 Jun 2026)
• PACT: Peak-Aware Cross-Attention Graph Transformers (Liu et al., 9 May 2026)
• PAct: Part-Decomposed Single-View Articulation (Liu et al., 16 Feb 2026)
• PACT: Privacy-Sensitive Contact Tracing Protocol (Chan et al., 2020)

The term "PACT" thus subsumes a broad spectrum of algorithmic, infrastructural, and methodological contributions, all characterized by formal formulations and empirical validation in settings ranging from multimodal AI and robotics to healthcare, security, privacy, and scientific modeling.