Preacher: Multifaceted Research Applications

• Preacher is a polysemous term defining a paper-to-video system that automates video abstracts by decomposing, summarizing, and synthesizing research papers into coherent visual content.
• In exercise science, preacher curls are monitored using wrist-mounted IMUs to detect near-failure states, enabling real-time coaching with precise measurements like RiR ≤ 2.
• Preacher also refers to mediation analysis techniques that integrate classical regression with counterfactual graphical models to estimate direct, indirect, and path-specific effects.

Searching arXiv for the cited papers to ground the article in current records. The term Preacher is used in multiple technical senses. In machine learning, it denotes Preacher, a paper-to-video agentic system that takes a complete research paper and automatically produces a video abstract (Liu et al., 13 Aug 2025). In resistance-training sensing, it appears in the compound expression preacher curls, which serve as the target exercise in a wrist-IMU system for real-time detection of near-failure states defined by $\text{RiR} \leq 2$ (King et al., 5 Dec 2025). In methodology for psychology and related fields, it also appears in discussion of Preacher-style mediation practice, which later work situates inside a broader counterfactual and graphical framework for direct, indirect, and path-specific effects (Shpitser, 2012).

1. Terminological scope

The three principal research uses are distinct in domain, object, and methodology.

Sense	Technical referent	Source
Preacher	paper-to-video agentic system for generating a video abstract from a full paper	(Liu et al., 13 Aug 2025)
preacher curls	exercise context for wrist-IMU near-failure coaching in hypertrophy training	(King et al., 5 Dec 2025)
Preacher-style mediation	regression/SEM-based mediation practice later reframed via counterfactuals and graphical models	(Shpitser, 2012)

This polysemy is not merely lexical. Each usage belongs to a different technical regime: multimodal generative systems, wearable sensing for resistance exercise, and causal inference for mediation analysis. As a result, interpretation depends entirely on disciplinary context.

2. Preacher as a paper-to-video agentic system

"Preacher: Paper-to-Video Agentic System" defines the paper-to-video task as the conversion of a research paper into a structured video abstract, with the aim of distilling key concepts, methods, and conclusions into an accessible, well-organized format (Liu et al., 13 Aug 2025). The system is positioned as the first paper-to-video agentic system. The motivating problem is explicit: scientific papers are long, multimodal, and domain-specific, whereas useful video abstracts must be concise, visually structured, semantically faithful, and often stylistically varied. Manual production is described as expensive because it requires both subject-matter understanding and professional video-making skills.

The paper situates Preacher against several limitations of state-of-the-art video generation models: limited context windows, rigid video duration constraints, limited stylistic diversity, and an inability to represent domain-specific knowledge. Preacher is introduced as an end-to-end automation framework intended to address precisely those constraints. Its central design principle is a top-down approach to decompose, summarize, and reformulate the paper, followed by bottom-up video generation, which synthesizes diverse video segments into a coherent abstract.

The significance of this formulation is methodological rather than merely presentational. The system treats the source document as a structured scientific object rather than as a short prompt. This suggests that the contribution is as much about long-context planning and semantic control as about raw video synthesis.

3. Formal framing, planning, and cross-modal alignment

In the formal framing, the input is a complete paper $P$, including text, equations, figures, and tables, and the target is a video abstract $V$. The paper defines a video as a frame sequence

$V=\{F_1,F_2,\ldots,F_T\},$

thereby making the task a mapping from a multimodal scientific document to an ordered visual-temporal representation (Liu et al., 13 Aug 2025).

To align cross-modal representations, the system defines key scenes and introduces a Progressive Chain of Thought (P-CoT) for granular, iterative planning. The top-down stage performs decomposition, summarization, and reformulation; the bottom-up stage then synthesizes diverse video segments into a coherent whole. The paper states that Preacher successfully generates high-quality video abstracts across five research fields, and attributes this to a design that demonstrates expertise beyond current video generation models.

Two points are technically important. First, the workflow is explicitly hierarchical: planning precedes synthesis. Second, the planning mechanism is iterative and scene-centric rather than monolithic. A plausible implication is that the system is designed to preserve discourse structure from scientific writing while translating it into temporal audiovisual structure. The paper also states that code will be released at https://github.com/GenVerse/Paper2Video.

4. Preacher curls as a sensing and coaching problem

In exercise science and edge AI, preacher appears in preacher curls, the sole exercise examined in "Rep Smarter, Not Harder: AI Hypertrophy Coaching with Wearable Sensors and Edge Neural Networks" (King et al., 5 Dec 2025). The paper studies using a single wrist-mounted IMU to monitor preacher curls and coach users in real time as they approach failure. The practical objective is hypertrophy-oriented: help lifters end sets close to failure, but not necessarily at total failure.

The central training variable is Repetitions in Reserve (RiR). The paper defines momentary muscular failure as the point in a set where the lifter can no longer complete another repetition with proper form, and RiR as the number of additional reps the lifter could still complete before that failure point. For the monitoring task, near-failure is operationalized exactly as

$\text{RiR} \leq 2.$

The system therefore does not estimate exact continuous RiR; it determines whether the user is in the final roughly $0$–$2$ reps before failure.

The dataset is purpose-built around preacher curls performed to failure. It contains 13 participants total, 68 sets, and 631 reps. Data were collected with a Raspberry Pi 5 and an Adafruit MPU-6050 6-axis IMU, using only a single 6-axis wrist IMU with 3-axis accelerometer and 3-axis gyroscope signals. The IMU was taped to the participant’s left wrist. Raw data were recorded at approximately 100 Hz and later interpolated to exactly 100 Hz. Preprocessing also included smoothing with a moving average filter with a window: 15 data points, equivalent to 150 ms, and manual annotation correction of rep-end markers.

The label structure is dual. The segmentation task is per-sample, with an 80 ms neighborhood around each end-of-rep marker labeled positive. The near-failure task is per-window, with a window labeled positive if more than 50% of its samples lie in a region where $\text{RiR} \le 2$. This distinction is central: the system is rep-aware, but the final coaching output is window-based.

5. Two-stage pipeline, real-time evaluation, and deployment

The core technical contribution is a two-stage edge-deployable pipeline (King et al., 5 Dec 2025). Stage 1 is repetition segmentation. It takes a window of 6-channel IMU data and outputs a binary prediction for each input time point indicating whether that point corresponds to the end of a rep. Each 256-sample input window therefore yields 256 output confidences. The architecture is a 1D convolutional ResNet with strided 1D convolutional layers, batch normalization, ReLU, residual connections, and a final global average pooling and linear classification stage. The final NAS-selected segmentation model contains 2,971,648 parameters. Training uses a hybrid BCE–MSE objective: $$Loss = \frac{1}{N} \sum_{i=1}^{N} \left( \alpha\cdot\text{BCE}(y_i, \hat y_i) + (1-\alpha)\frac{1}{256} \sum_{j=1}^{256} (\hat y_{ij} - y_{ij})^2 \right)$$ with

$\alpha = 0.8.$

The segmentation outputs are further processed into a time-in-rep feature: predictions are smoothed, converted into positive regions, reduced to rep-end midpoints, and used to estimate time spent in each rep inside the window. Stage 2 is near-failure classification. Its positive class is defined by

$P$0

The classifier fuses four information sources: a frozen segmentation backbone yielding a 512-dimensional vector, a 256-dimensional time-in-rep vector, a skip-path convolution over raw IMU contributing a 64-dimensional vector, and historical context modeled by a one-directional LSTM. The fused feature size is

$P$1

Final classifier hyperparameters are LSTM hidden size: 256, number of LSTM layers: 4, and a classification head: linear layer from 256 to 1. The total parameter count is 5,291,841, of which 2,320,193 are trainable.

Real-time operation is defined by window size $P$2, stride $P$3, and temporal context $P$4. At 100 Hz, a 256-sample window corresponds to 2.56 s and a 64-sample stride corresponds to 640 ms, yielding an update rate of approximately

$P$5

reported in the abstract as 1.6 Hz inference rate. Under simulated real-time evaluation, the segmentation model achieved F1 score: 0.830, Precision: 0.800, Recall: 0.869, and Accuracy: 92.7% / 92.8%. The near-failure classifier achieved F1 score: 0.824, Precision: 0.860, Recall: 0.856, and Accuracy: 86.6%.

Deployment results were reported on both embedded and mobile hardware. The final deployed model size is 20 MB. On Raspberry Pi 5, average inference latency was 112.01 ms average, rising to 130.03 ms for full 32-window inputs; profiling attributed 45.81% of inference CPU time to the convolution operator and 26.63% to the RNN operator. On iPhone 16, average inference latency was 23.5 ms average, rising to 35.19 ms for full 32-window inputs, with encoder: 14.05 ms, time-in-rep function: 1.51 ms, and LSTM: 7.94 ms. Compared with the 640 ms update interval, both deployments are fast enough for live feedback, and the paper states that the watch/phone system can provide haptic feedback when near-failure is detected.

Several limitations are explicit. The paper reports no formal ablation table or baseline comparison. Data collection focused solely on preacher curls, included only 13 participants, and may not capture the full variety of lifting styles and fatigue responses. It also states that future validation is needed for other exercises such as squats and bench press. A common misreading would be to treat the system as a general RiR estimator; the paper, by contrast, defines the task more narrowly as binary detection of the near-failure zone.

6. Preacher in mediation-analysis methodology

A third use of the term appears in methodological discourse around mediation analysis. "Counterfactual Graphical Models for Longitudinal Mediation Analysis with Unobserved Confounding" notes that about 60% of papers published in leading journals in social psychology contain at least one mediation test and situates the standard difference method and product method inside a more general counterfactual framework for causality (Shpitser, 2012). The paper is directly relevant to the tradition associated with Baron & Kenny, MacKinnon, and Preacher-style regression/SEM-based mediation.

For the standard single-mediator setup with treatment $P$6, mediator $P$7, and outcome $P$8, the familiar linear models are

$P$9

$V$0

and

$V$1

In this formulation, the indirect effect is either $V$2 or $V$3. The paper’s main conceptual move is to define the relevant causal estimands counterfactually: $V$4

$V$5

and

$V$6

These correspond respectively to the total effect, natural direct effect, and natural indirect effect. Under the special case where $V$7 is continuous, the structural functions are linear with no interactions, and the noise terms are Gaussian, these counterfactual definitions reduce to the familiar regression-coefficient formulas.

The paper’s emphasis is that mediation analysis requires strong assumptions, many of which are hidden in routine regression practice. These include ignorability conditions such as $V$8 and the cross-world assumption

$V$9

Under the stated assumptions, the direct effect is identified by the mediation formula

$V=\{F_1,F_2,\ldots,F_T\},$0

The paper argues that classical regression approaches are therefore not a general causal theory of mediation, but rather a special parametric corner of a broader counterfactual framework.

Its novel theoretical contribution concerns longitudinal settings with unobserved confounders and path-specific effects. Using causal diagrams and ADMGs, the paper introduces the recanting district criterion and states the theorem that the $V=\{F_1,F_2,\ldots,F_T\},$1-specific effect of $V=\{F_1,F_2,\ldots,F_T\},$2 on $V=\{F_1,F_2,\ldots,F_T\},$3 is expressible as a functional of interventional densities if and only if there does not exists a recanting district for this effect. It further states that, when there is no recanting district, the corresponding counterfactual is expressible in terms of the observed data $V=\{F_1,F_2,\ldots,F_T\},$4 if and only if the total effect $V=\{F_1,F_2,\ldots,F_T\},$5 is expressible in terms of $V=\{F_1,F_2,\ldots,F_T\},$6.

This framework is important for interpreting what is often called Preacher-style mediation. It complements improvements in estimation and inference for indirect effects, but it also warns that better standard errors or bootstrap confidence intervals do not solve identification problems. The paper explicitly discusses failures of the product and difference methods under binary outcomes, interaction terms in the outcome, non-linearities, unobserved confounding, and more complex mediator structures. It also highlights the null paradox and the fact that cross-world assumptions are untestable from observed data alone. In that sense, the methodological meaning of Preacher is not a single estimator or macro, but a broader conversation about how regression-based mediation relates to causal estimands, graphical structure, and identification theory.