1<!DOCTYPE html> 2<html lang="en"> 3<head> 4 <meta charset="UTF-8" /> 5 <meta name="google-site-verification" content="f1SOW2rlBn2OZ2wEit5apbfDKtJXCdaw0CN22K308r4" /> 6 <meta name="viewport" content="width=device-width, initial-scale=1.0" /> 7 <title>EDM-ARS & LSAR — Educational Data Mining Research Tools</title> 8 <meta name="description" content="EDM-ARS automates end-to-end educational data mining research â prediction, causal inference and psychometrics â from raw data to a formatted manuscript. LSAR provides automated peer review across 8 dimensions for 12 learning science venues, with thresholds calibrated on published papers." /> 9 <link rel="preconnect" href="https://fonts.googleapis.com" /> 10 <link rel="preconnect" href="https://fonts.gstatic.com" crossorigin /> 11 <link href="https://fonts.googleapis.com/css2?family=Oxanium:wght@300;400;600;700;800&family=Source+Serif+4:ital,wght@0,300;0,400;0,600;1,300&family=JetBrains+Mono:wght@400;500&display=swap" rel="stylesheet" /> 12 <link rel="stylesheet" href="style.css" /> 13</head> 14<body> 15 16 <!-- ==================== NAV ==================== --> 17 <nav> 18 <a class="nav-logo" href="#hero">EDM-ARS</a> 19 <ul class="nav-links"> 20 <li><a href="#about">Overview</a></li> 21 <li><a href="#pipeline">Architecture</a></li> 22 <li><a href="#capabilities">Features</a></li> 23 <li><a href="#demo">Demo Stats</a></li> 24 <li><a href="#lsar">LSAR</a></li> 25 <li><a href="#roadmap">Roadmap</a></li> 26 <li><a href="#papers">Papers</a></li> 27 </ul> 28 </nav> 29 30 <!-- ==================== HERO ==================== --> 31 <section id="hero"> 32 <canvas id="mars-canvas" aria-hidden="true"></canvas> 33 <div class="hero-vignette"></div> 34 <div class="hero-content"> 35 <p class="hero-badge">ðª EDM-ARS · v5 · Open Source</p> 36 <h1>Educational Data Mining<br><span>Automated</span> Research System</h1> 37 <p class="tagline">A multi-agent LLM pipeline that turns a dataset and a research prompt into a complete, reviewer-ready academic paper â across prediction, causal inference and psychometrics â then scores it against thresholds calibrated on papers those venues actually published.</p> 38 <div class="hero-buttons"> 39 <a class="btn btn-primary" href="#pipeline">System Architecture</a> 40 <a class="btn btn-outline" href="#capabilities">Key Features</a> 41 <a class="btn btn-outline" href="https://github.com/cgpan/edm-ars-public" target="_blank" rel="noopener">Source Code</a> 42 <a class="btn btn-outline" href="#lsar">Meet LSAR</a> 43 </div> 44 45 <div class="hero-next"> 46 <span class="hero-next-flag"><span class="hero-next-dot" aria-hidden="true"></span>Coming Soon</span> 47 <span class="hero-next-body"><strong>QM-ARS</strong> — the Quantitative Methods Automated Research System. The next generation of EDM-ARS, widening the pipeline to broader quantitative methods and research topics.</span> 48 </div> 49 </div> 50 <div class="hero-stats"> 51 <div class="hero-stat"> 52 <div class="hero-stat-val">5</div> 53 <div class="hero-stat-label">Study Types Supported</div> 54 </div> 55 <div class="hero-stat"> 56 <div class="hero-stat-val">12</div> 57 <div class="hero-stat-label">Review Venues (LSAR)</div> 58 </div> 59 <div class="hero-stat"> 60 <div class="hero-stat-val">30â60 min</div> 61 <div class="hero-stat-label">End-to-End Runtime</div> 62 </div> 63 <div class="hero-stat"> 64 <div class="hero-stat-val">$0.53</div> 65 <div class="hero-stat-label">Measured Cost Per Paper</div> 66 </div> 67 </div> 68 </section> 69 70 <!-- ==================== ABOUT ==================== --> 71 <section class="section reveal" id="about"> 72 <p class="section-label">Overview</p> 73 <h3 class="section-title">What It Does</h3> 74 <p class="section-body"> 75 Inspired by <a href="https://analemma.ai/blog/introducing-fars/" target="_blank" rel="noopener noreferrer">FARS</a>, EDM-ARS is an open-source, domain-specific multi-agent LLM pipeline that automates the complete workflow of educational data mining research. Given a curated dataset and a research prompt, it formulates a research question, engineers the analytic sample, runs the analysis, retrieves real citations via the <strong>Semantic Scholar API</strong> with three-layer verification (exact title match, Jaccard similarity, CrossRef cross-validation), and produces a complete <strong>LaTeX manuscript</strong> in ACM <em>sigconf</em> or APA-7 journal format â with a built-in Critic agent that enforces methodological rigor through automated peer review and targeted revision loops. 76 </p> 77 <p class="section-body"> 78 v5 covers <strong>five study types</strong>: predictive modeling, three causal-inference designs (selection-on-observables, individualized treatment regimes, difference-in-differences), and psychometrics (IRT, measurement invariance, DIF, cognitive diagnosis). Ten estimation methods are certified against synthetic data with known ground truth before they are allowed near a real dataset. Four datasets are curated: <strong>HSLS:09</strong>, <strong>ELS:2002</strong>, <strong>
78ASSISTments 2009–10</strong>, and a constructed ELS×HSLS cross-cohort panel. 79 </p> 80 81 <div class="about-grid"> 82 <div class="about-item"> 83 <div class="about-item-icon">ð</div> 84 <h4>Domain-Specific Design</h4> 85 <p>Each dataset carries a <strong>three-tier variable registry</strong>: hand-curated Tier 1 variables with educational annotations, auto-generated Tier 2 variables, and Tier 3 exclusions (weights, IDs) enforced programmatically. Survey-design realities â school clustering, sentinel missing codes, legitimate skips, suppressed identifiers â are encoded per dataset rather than left for the model to infer.</p> 86 </div> 87 <div class="about-item"> 88 <div class="about-item-icon">ð§©</div> 89 <h4>Skill-Based Architecture</h4> 90 <p><strong>70 composable skills</strong> across four layers â task type, dataset, methodology and writing â are matched at runtime and injected into each agent's prompt. Adding a capability means adding a skill, not enlarging a prompt. Skills marked <em>mandatory</em> bypass the size cap, because a rule whose violation corrupts output silently must always reach the agent.</p> 91 </div> 92 <div class="about-item"> 93 <div class="about-item-icon">ð¬</div> 94 <h4>Certified Before Live Use</h4> 95 <p>Every estimator faces a <strong>synthetic-data gate</strong> before it touches real data: a data-generating process with a known true effect, and the method must recover it. Difference-in-differences was held back until it hit 0.006 bias and detected pre-trend violations at a rate of 1.0.</p> 96 </div> 97 <div class="about-item"> 98 <div class="about-item-icon">ð</div> 99 <h4>Deterministic Honesty Guards</h4> 100 <p>Prompting a model not to invent numbers does not stop it. Every figure in the manuscript is <strong>reconciled against the analysis artifacts</strong>, unverifiable claims are flagged, and a paper whose numbers cannot be traced is blocked from passing the gate â regardless of how well it reads.</p> 101 </div> 102 </div> 103 </section> 104 105 <hr class="divider" /> 106 107 <!-- ==================== PIPELINE ==================== --> 108 <section id="pipeline"> 109 <div class="inner reveal"> 110 <p class="section-label">Architecture</p> 111 <h3 class="section-title">Six-Agent Pipeline</h3> 112 <p class="section-body">The dynamic graph below demonstrates our architecture and its information flow. It replays a real EDM-ARS run.</p> 113 114 <div class="rp" id="replay"> 115 116 <div class="rp-transport"> 117 <div class="rp-controls"> 118 <button class="btn btn-primary rp-btn" id="rp-replay" type="button"> 119 <svg viewBox="0 0 12 12" aria-hidden="true"><path d="M9.6 3.1A4.3 4.3 0 1 0 10.3 6.9" fill="none" stroke="currentColor" stroke-width="1.5" stroke-linecap="round"/><path d="M10.6 0.9v3.4H7.2z" fill="currentColor"/></svg> 120 <span>Replay run</span> 121 </button> 122 <button class="btn btn-outline rp-btn" id="rp-toggle" type="button" disabled> 123 <svg viewBox="0 0 12 12" aria-hidden="true"><path d="M2.5 1.2v9.6L10.6 6z" fill="currentColor"/></svg> 124 <span>Resume</span> 125 </button> 126 <button class="btn btn-outline rp-btn" id="rp-end" type="button"> 127 <svg viewBox="0 0 12 12" aria-hidden="true"><path d="M1.4 1.2v9.6L7.7 6zM8.6 1.2h2.2v9.6H8.6z" fill="currentColor"/></svg> 128 <span>Jump to end</span> 129 </button> 130 </div> 131 <div class="rp-speed"> 132 <label class="rp-speed-lab" for="rp-speed">Speed</label> 133 <input type="range" id="rp-speed" min="0" max="4" step="1" value="1" /> 134 <output id="rp-speed-out" for="rp-speed">30à · â 1 min</output> 135 </div> 136 <div class="rp-status"> 137 <span class="rp-pill" id="rp-pill" data-kind="done">Completed</span> 138 <span class="rp-clock" title="Wall-clock time of the run, UTC"><span id="rp-clock">18:22:41</span><span class="rp-elapsed" id="rp-elapsed">+29:33</span></span> 139 </div> 140 </div> 141 142 <div class="rp-panel"> 143 <header>
144 <span class="rp-eyebrow">Agents, orchestrator and the resources they reach</span> 145 <span class="rp-eyebrow rp-dim" id="rp-stage">run complete</span> 146 </header> 147 <div class="rp-graph-scroll"> 148 <svg id="rp-graph" viewBox="0 0 966 384" role="img" 149 aria-label="Diagram of the EDM-ARS pipeline: five agents above a central orchestrator rail, with the external resources they reach above them and orchestrator-only steps below."></svg> 150 </div> 151 <p class="rp-pan-hint">Swipe sideways to see the whole diagram.</p> 152 <div class="rp-caption"> 153 <div class="rp-cap-hd"><span class="rp-cap-art" id="rp-cap-art">run complete</span><span class="rp-cap-route" id="rp-cap-route"></span></div> 154 <p class="rp-cap-note" id="rp-cap-note">Press Replay run to watch the 29-minute run compressed into about a minute, or drag the trace below to scrub.</p> 155 </div> 156 </div> 157 158 <div class="rp-panel"> 159 <header> 160 <span class="rp-eyebrow">Wall-clock trace · 2026-08-29 17:53:08 â 18:22:41 UTC</span> 161 <span class="rp-eyebrow rp-dim">click or drag to scrub</span> 162 </header> 163 <div class="rp-tl" id="rp-tl" role="slider" tabindex="0" 164 aria-label="Replay position" aria-valuemin="0" aria-valuemax="1773" aria-valuenow="1773"> 165 <svg id="rp-timeline" aria-hidden="true"></svg> 166 </div> 167 <div class="rp-legend"> 168 <span class="rp-key"><i class="k-work"></i>agent working</span> 169 <span class="rp-key"><i class="k-revise"></i>Critic → REVISE</span> 170 <span class="rp-key"><i class="k-pass"></i>Critic → PASS</span> 171 <span class="rp-key"><i class="k-mech"></i>LaTeX compile</span> 172 <span class="rp-key"><i class="k-tick"></i>LLM call returned · 21 calls, bar height = tokens</span> 173 </div> 174 <p class="rp-foot">Hand-offs recorded in the same second are staggered by a few seconds so the packets can be told apart; every clock time shown is as recorded.</p> 175 </div> 176 177 </div> 178 </div> 179 </section> 180 181 <hr class="divider" /> 182 183 <!-- ==================== CAPABILITIES ==================== --> 184 <section class="section reveal" id="capabilities"> 185 <p class="section-label">Features</p> 186 <h3 class="section-title">Key Capabilities</h3> 187 188 <div class="caps-grid"> 189 <div class="cap-card"> 190 <div class="cap-icon">ð¤</div> 191 <p class="cap-stat">6 Specialized Agents</p> 192 <p class="cap-desc">Coordinated by a state-machine orchestrator. Each agent has its own system prompt, temperature and model routing, configurable per stage. The reference configuration runs the whole pipeline on <strong>DeepSeek-v4-pro</strong>; Anthropic, OpenAI and MiniMax backends are supported through the same interface.</p> 193 </div> 194 <div class="cap-card"> 195 <div class="cap-icon">ð</div> 196 <p class="cap-stat">End-to-End Automation</p> 197 <p class="cap-desc">From a raw CSV and a research prompt to a compiled ACM LaTeX paper â with real citations, methodology validation, and SHAP explainability figures.</p> 198 </div> 199 <div class="cap-card"> 200 <div class="cap-icon">ð¡</div> 201 <p class="cap-stat">Self-Healing Pipeline</p> 202 <p class="cap-desc">Contract validation at every stage boundary. Auto-patching for classifiable errors (SHAP failure, dtype mismatch, missing column) before falling back to LLM repair.</p> 203 </div> 204 <div class="cap-card"> 205 <div class="cap-icon">ð</div> 206 <p class="cap-stat">Live Academic Citations</p> 207 <p class="cap-desc">The ProblemFormulator queries the Semantic Scholar API with exponential-backoff retry logic to retrieve and validate real, current citations.</p> 208 </div> 209 <div class="cap-card"> 210 <div class="cap-icon">âï¸</div> 211 <p class="cap-stat">6-Model Battery</p> 212 <p class="cap-desc">Logistic Regression, Random Forest, XGBoost, ElasticNet, MLP, and a Stacking Ensemble are trained, compared, and reported with SHAP explainability. A <strong>Model Quality Gate</strong> (AUC ≥ 0.60 for classification, R² ≥ 0.05 for regression) runs before SHAP, so a model that has learned nothing is not given an interpretation.</p> 213 </div> 214 <div class="cap-card"> 215 <div class="cap-icon">ð³</div> 216 <p class="cap-stat">Docker Sandboxing</p> 217 <p class="cap-desc">LLM-generated analysis code executes inside a Docker sandbox (network-disabled). Gracefully falls back to subprocess when Docker is unavailable.</p> 218 </div> 219 <div class="cap-card"> 220 <div class="cap-icon">ð¯</div> 221 <p class="cap-stat">Causal Inference <span class="cap-new">New</span></p> 222 <p class="cap-desc">Selection-on-observables (matching, IPW, regression adjustment, AIPW/TMLE), c
222ausal forests for heterogeneous effects, individualized treatment regimes, and difference-in-differences. Estimands are declared before estimation; balance, overlap and unmeasured-confounding sensitivity are reported as a matter of course.</p> 223 </div> 224 <div class="cap-card"> 225 <div class="cap-icon">ð</div> 226 <p class="cap-stat">Psychometrics <span class="cap-new">New</span></p> 227 <p class="cap-desc">Item response theory, graded response models, measurement invariance testing, differential item functioning, and cognitive diagnosis (DINA / G-DINA) â executed in R through a bridge, with degeneracy reported rather than hidden.</p> 228 </div> 229 <div class="cap-card"> 230 <div class="cap-icon">ð§¾</div> 231 <p class="cap-stat">Numeric Reconciliation <span class="cap-new">New</span></p> 232 <p class="cap-desc">Every number in the manuscript is matched back to the analysis outputs that produced it. An audit once found a paper that scored well while reporting fabricated values for fields the analysis left empty; that class of failure is now blocked deterministically rather than discouraged by prompt.</p> 233 </div> 234 <div class="cap-card"> 235 <div class="cap-icon">ð</div> 236 <p class="cap-stat">Calibrated Review Gate <span class="cap-new">New</span></p> 237 <p class="cap-desc">Papers are scored by LSAR against the 25th percentile of papers the target venue actually published â not an arbitrary threshold. Borderline verdicts trigger median-of-three sampling, because a single review carries real measurement noise.</p> 238 </div> 239 <div class="cap-card"> 240 <div class="cap-icon">ðµ</div> 241 <p class="cap-stat">Measured Token Accounting <span class="cap-new">New</span></p> 242 <p class="cap-desc">Every run records prompt, completion and cached-input tokens per call. Rates live in configuration, so a price change re-prices past runs without re-running them, and an unpriced model yields <code>null</code> rather than a silent zero.</p> 243 </div> 244 </div> 245 </section> 246 247 <!-- ==================== DEMO STATS ==================== --> 248 <section id="demo"> 249 <div class="inner reveal"> 250 <p class="section-label">Measured Results</p> 251 <h3 class="section-title">What It Actually Produces</h3> 252 <p class="section-body">Fifteen complete papers have been generated and scored through the calibrated review gate. Every study type has at least one paper that clears its venue's threshold. Scores span <strong>3.7 to 7.5</strong> â the low end is reported here for the same reason it is reported in the papers themselves.</p> 253 254 <div class="demo-grid"> 255 <div class="demo-stat-card"> 256 <div class="demo-stat-val">15</div> 257 <div class="demo-stat-label">Papers Generated & Scored</div> 258 </div> 259 <div class="demo-stat-card"> 260 <div class="demo-stat-val">7.5</div> 261 <div class="demo-stat-label">Best LSAR Score (of 10)</div> 262 </div> 263 <div class="demo-stat-card"> 264 <div class="demo-stat-val">30â60 min</div> 265 <div class="demo-stat-label">Runtime Per Paper</div> 266 </div> 267 <div class="demo-stat-card"> 268 <div class="demo-stat-val">$0.53</div> 269 <div class="demo-stat-label">Measured Cost Per Paper</div> 270 </div> 271 </div> 272 273 <p class="section-body" style="margin-top:28px;"> 274 That cost is measured, not estimated: one fully instrumented run made <strong>84 LLM calls</strong>. <strong>81% of input tokens were prompt-cache hits</strong>, billed roughly ten times lower, which is the dominant lever and one that summed token counts cannot see. An earlier figure on this page quoted several dollars per paper; it came from a hardcoded rate constant rather than a measurement, and it was wrong by nearly an order of magnitude. 275 </p> 276 </div> 277 </section> 278 279 <!-- ==================== LSAR ==================== --> 280 <section id="lsar"> 281 <div class="inner reveal"> 282 <p class="section-label">Companion Tool · v0.4</p> 283 <h3 class="section-title">LSAR — Learning Science Auto-Reviewer</h3> 284 <p class="section-body">Inspired by <a href="https://paperreview.ai/tech-overview" target="_blank" rel="noopener noreferrer">PaperReview.ai</a>, LSAR is an automated, agentic paper reviewer for learning science venues. A six-stage pipeline ingests a PDF, identifies the venue, retrieves related work, and produces a structured review across <strong>8 quality dimensions</strong> for <strong>12 venues</strong> â four conferences and eight journals.</p> 285 286 <div class="lsar-dims-grid"> 287 <div class="lsar-dim-card"> 288 <div class="lsar-dim-icon">ð¯</div> 289 <strong>Relevance</strong> 290 <p>Alignment with venue scope and topic fit</p> 291 </div> 292 <div class="lsar-dim-card"> 293 <div class="lsar-dim-icon">ð¡</div> 294 <strong>Novelty</strong> 295 <p>Originality and contribution beyond prior work</p> 296 </div> 297 <div class="lsar-dim-card"> 298 <div class="lsar-dim-icon">ð</div> 299 <strong>Theoretical Grounding</strong> 300 <p>Strength of conceptual framework and motivation</p> 301 </div> 302 <div class="lsar-dim-card"> 303 <div class="lsar-dim-icon">ð§</div> 304 <strong>Methodological Rigor</strong> 305 <p>Soundness of research design and analysis</p> 306 </div> 307 <div class="lsar-dim-card"> 308 <div class="lsar-dim-icon">ð</div> 309 <strong>Empirical Support</strong> 310 <p>Quality and interpretation of evidence</p> 311 </div> 312 <div class="lsar-dim-card"> 313 <div class="lsar-dim-icon">ð</div> 314 <strong>Significance</strong> 315 <p>
315Impact on the field and practical implications</p> 316 </div> 317 <div class="lsar-dim-card"> 318 <div class="lsar-dim-icon">âï¸</div> 319 <strong>Ethics & Fairness</strong> 320 <p>Ethical considerations and bias awareness</p> 321 </div> 322 <div class="lsar-dim-card"> 323 <div class="lsar-dim-icon">âï¸</div> 324 <strong>Communication</strong> 325 <p>Clarity, structure, and readability of writing</p> 326 </div> 327 </div> 328 329 <div class="lsar-venues"> 330 <span class="lsar-venue-tag">AIED</span> 331 <span class="lsar-venue-tag">EDM</span> 332 <span class="lsar-venue-tag">L@S</span> 333 <span class="lsar-venue-tag">LAK</span> 334 <span class="lsar-venue-tag">AERA Open</span> 335 <span class="lsar-venue-tag">JEDM</span> 336 <span class="lsar-venue-tag">JLA</span> 337 <span class="lsar-venue-tag">JEBS</span> 338 <span class="lsar-venue-tag">JEM</span> 339 <span class="lsar-venue-tag">JREE</span> 340 <span class="lsar-venue-tag">MBR</span> 341 <span class="lsar-venue-tag">Psychometrika</span> 342 </div> 343 344 <div class="lsar-loop-callout" style="margin-top:32px;"> 345 <div class="lsar-loop-icon">ð</div> 346 <p><strong>Calibrated thresholds.</strong> A raw review score means little on its own â nothing anchors “6.5” to any real standard. LSAR reviews papers a venue <em>actually published</em> and takes the 25th percentile of their scores. A submission below that sits below the weakest quartile of what the venue accepted. Current gates: <strong>EDM 6.3</strong> (15 anchors), <strong>AERA Open 6.6</strong> (11), <strong>JEDM 5.3</strong> (10), <strong>JLA 5.4</strong> (9).</p> 347 </div> 348 349 <div class="lsar-loop-callout" style="margin-top:16px;"> 350 <div class="lsar-loop-icon">â ï¸</div> 351 <p><strong>And what those numbers cannot do.</strong> Measured test–retest deviation is <strong>1.9 points</strong>, so a single review is a noisy instrument and borderline cases are sampled three times and taken at the median. Anchor corpora run 9–15 papers per venue: these are working thresholds, not population parameters. A gate is a triage signal for authors, never an acceptance decision.</p> 352 </div> 353 354 <div class="lsar-features"> 355 <div class="lsar-feature"> 356 <div class="lsar-feature-icon">ð</div> 357 <p><strong>Venue Detection</strong> â automatically classifies papers for their intended conference or allows manual specification</p> 358 </div> 359 <div class="lsar-feature"> 360 <div class="lsar-feature-icon">ð¦</div> 361 <p><strong>Batch Processing</strong> â review multiple papers in a single run from directories or CSV manifests</p> 362 </div> 363 <div class="lsar-feature"> 364 <div class="lsar-feature-icon">ð</div> 365 <p><strong>Multiple Formats</strong> â generates reviews in Markdown, JSON, and PDF reports with modular pipeline stages</p> 366 </div> 367 <div class="lsar-feature"> 368 <div class="lsar-feature-icon">ð©º</div> 369 <p><strong>Review Health Checks</strong> â a truncated review is rejected rather than scored. 21% of reviews were once cut off between Strengths and Weaknesses, keeping the praise and losing the criticism, and scored <strong>+0.83 too high</strong> as a result. Truncation is now a hard error at three layers.</p> 370 </div> 371 </div> 372 373 <div class="lsar-loop-callout"> 374 <div class="lsar-loop-icon">ð</div> 375 <p><strong>EDM-ARS</strong> generates a paper → <strong>LSAR</strong> reviews it across 8 dimensions → feedback drives targeted revisions â a closed-loop generate-then-review cycle for automated research refinement.</p> 376 </div> 377 378 <div class="lsar-buttons"> 379 <a class="btn btn-primary" href="https://github.com/cgpan/LSAR-public" target="_blank" rel="noopener">View LSAR on GitHub</a> 380 </div> 381 </div> 382 </section> 383 384 <hr class="divider" /> 385 386 <!-- ==================== PAPERS ==================== --> 387 <section class="section reveal" id="papers"> 388 <p class="section-label">Publications</p> 389 <h3 class="section-title">Papers & Reports</h3> 390 <p class="section-body">Technical reports and demo papers generated by EDM-ARS. Demo papers are full submission-formatted manuscripts produced by the pipeline on HSLS:09 prediction tasks — APA-7 journal format for the JEDM submissions, ACM <em>sigconf</em> for the EDM submissions.</p> 391 392 <div class="papers-grid"> 393 394 <div class="paper-card"> 395 <div class="paper-meta">
396 <span class="paper-badge paper-badge-report">Technical Report</span> 397 <span class="paper-date">March 2026</span> 398 </div> 399 <h4 class="paper-title">EDM-ARS: A Domain-Specific Multi-Agent System for Automated Educational Data Mining Research</h4> 400 <p class="paper-authors">Chenguang Pan, Zhou Zhang, Weixuan Xiao, Chengyuan Yao</p> 401 <a class="paper-link" href="papers/edmars_tech_report_v2.pdf" target="_blank" rel="noopener">PDF →</a> 402 </div> 403 404 <div class="paper-card"> 405 <div class="paper-meta"> 406 <span class="paper-badge paper-badge-demo">Demo Paper</span> 407 <span class="paper-badge paper-badge-review">LSAR Reviewed</span> 408 <span class="paper-date">September 2026</span> 409 </div> 410 <h4 class="paper-title">Engagement and Belonging Add Predictive Signal for High School Dropout Episodes Beyond Achievement and SES: Evidence from a Nationally Representative U.S. Cohort</h4> 411 <p class="paper-authors">EDM-ARS (Automated) · Target Venue: JEDM</p> 412 <div class="paper-links"> 413 <a class="paper-link" href="papers/demo_paper_03.pdf" target="_blank" rel="noopener">Paper PDF →</a> 414 <a class="paper-link paper-link-review" href="papers/demo_paper_03_LSAR_review.pdf" target="_blank" rel="noopener">LSAR Review →</a> 415 </div> 416 </div> 417 418 <div class="paper-card"> 419 <div class="paper-meta"> 420 <span class="paper-badge paper-badge-demo">Demo Paper</span> 421 <span class="paper-badge paper-badge-review">LSAR Reviewed</span> 422 <span class="paper-date">September 2026</span> 423 </div> 424 <h4 class="paper-title">Predicting Two-Year Public Versus Four-Year Postsecondary Enrollment from Ninth-Grade Measures and the High School Transcript: Calibration and Subgroup Fairness in HSLS:09</h4> 425 <p class="paper-authors">EDM-ARS (Automated) · Target Venue: JEDM</p> 426 <div class="paper-links"> 427 <a class="paper-link" href="papers/demo_paper_04.pdf" target="_blank" rel="noopener">Paper PDF →</a> 428 <a class="paper-link paper-link-review" href="papers/demo_paper_04_LSAR_review.pdf" target="_blank" rel="noopener">LSAR Review →</a> 429 </div> 430 </div> 431 432 <div class="paper-card"> 433 <div class="paper-meta"> 434 <span class="paper-badge paper-badge-demo">Demo Paper</span> 435 <span class="paper-badge paper-badge-review">LSAR Reviewed</span> 436 <span class="paper-date">September 2026</span> 437 </div> 438 <h4 class="paper-title">
438High School Preparation, Not Background, Predicts Postsecondary Persistence: An Access–Persistence Contrast Using Machine Learning on HSLS:09</h4> 439 <p class="paper-authors">EDM-ARS (Automated) · Target Venue: JEDM</p> 440 <div class="paper-links"> 441 <a class="paper-link" href="papers/demo_paper_05.pdf" target="_blank" rel="noopener">Paper PDF →</a> 442 <a class="paper-link paper-link-review" href="papers/demo_paper_05_LSAR_review.pdf" target="_blank" rel="noopener">LSAR Review →</a> 443 </div> 444 </div> 445 446 <div class="paper-card"> 447 <div class="paper-meta"> 448 <span class="paper-badge paper-badge-demo">Demo Paper</span> 449 <span class="paper-badge paper-badge-review">LSAR Reviewed</span> 450 <span class="paper-date">March 2026</span> 451 </div> 452 <h4 class="paper-title">Fairness-Through-Unawareness Does Not Produce Equitable Predictions: Evidence from HSLS:09 on Subgroup Disparities in College Attendance Prediction</h4> 453 <p class="paper-authors">EDM-ARS (Automated) · Target Venue: EDM</p> 454 <div class="paper-links"> 455 <a class="paper-link" href="papers/demo_paper_01.pdf" target="_blank" rel="noopener">Paper PDF →</a> 456 <a class="paper-link paper-link-review" href="papers/demo_paper_01_LSAR_review.pdf" target="_blank" rel="noopener">LSAR Review →</a> 457 </div> 458 </div> 459 460 <div class="paper-card"> 461 <div class="paper-meta"> 462 <span class="paper-badge paper-badge-demo">Demo Paper</span> 463 <span class="paper-badge paper-badge-review">LSAR Reviewed</span> 464 <span class="paper-date">March 2026</span> 465 </div> 466 <h4 class="paper-title">Do Ninth-Grade Math and Science Identity Predict STEM Degree Non-Completion? Evidence from HSLS:09 and Machine Learning</h4> 467 <p class="paper-authors">EDM-ARS (Automated) · Target Venue: EDM</p> 468 <div class="paper-links"> 469 <a class="paper-link" href="papers/demo_paper_02.pdf" target="_blank" rel="noopener">Paper PDF →</a> 470 <a class="paper-link paper-link-review" href="papers/demo_paper_02_LSAR_review.pdf" target="_blank" rel="noopener">LSAR Review →</a> 471 </div> 472 </div> 473 474 </div> 475 </section> 476 477 <!-- ==================== TECH STACK ==================== -->
478 <section class="section reveal" id="tech"> 479 <p class="section-label">Built With</p> 480 <h3 class="section-title">Tech Stack</h3> 481 482 <div class="tech-groups"> 483 <div> 484 <div class="tech-group-title">Core Pipeline</div> 485 <div class="tech-tags"> 486 <span class="tech-tag accent">Python 3.11</span> 487 <span class="tech-tag accent">DeepSeek-v4-pro</span> 488 <span class="tech-tag">Anthropic / OpenAI / MiniMax</span> 489 <span class="tech-tag">Docker Sandbox</span> 490 <span class="tech-tag">PyYAML</span> 491 </div> 492 </div> 493 <div> 494 <div class="tech-group-title">Prediction & Causal</div> 495 <div class="tech-tags"> 496 <span class="tech-tag">scikit-learn</span> 497 <span class="tech-tag">XGBoost</span> 498 <span class="tech-tag">SHAP</span> 499 <span class="tech-tag">causal forests</span> 500 <span class="tech-tag">AIPW / TMLE</span> 501 <span class="tech-tag">pandas</span> 502 </div> 503 </div> 504 <div> 505 <div class="tech-group-title">Psychometrics</div> 506 <div class="tech-tags"> 507 <span class="tech-tag accent">R bridge</span> 508 <span class="tech-tag">IRT / GRM</span> 509 <span class="tech-tag">measurement invariance</span> 510 <span class="tech-tag">DIF</span> 511 <span class="tech-tag">DINA / G-DINA</span> 512 </div> 513 </div> 514 <div> 515 <div class="tech-group-title">Data & Literature</div> 516 <div class="tech-tags"> 517 <span class="tech-tag accent">HSLS:09</span> 518 <span class="tech-tag accent">ELS:2002</span> 519 <span class="tech-tag accent">ASSISTments</span> 520 <span class="tech-tag">ELS×HSLS panel</span> 521 <span class="tech-tag">Semantic Scholar + arXiv</span> 522 </div> 523 </div> 524 <div> 525 <div class="tech-group-title">Output & Verification</div> 526 <div class="tech-tags"> 527 <span class="tech-tag accent">ACM sigconf LaTeX</span> 528 <span class="tech-tag accent">APA-7 journal</span> 529 <span class="tech-tag">numeric reconciliation</span> 530 <span class="tech-tag">calibrated LSAR gate</span> 531 <span class="tech-tag">token metering</span> 532 </div> 533 </div> 534 </div> 535 </section> 536 537 <!-- ==================== ROADMAP ==================== --> 538 <section id="roadmap"> 539 <div class="inner reveal"> 540 <p class="section-label">Scope & Roadmap</p> 541 <h3 class="section-title">v5 & Beyond</h3> 542 <p class="section-body">Every phase on the original roadmap â polymorphism, findings memory, causal inference, multi-dataset support â has shipped. The current release spans five study types across four datasets, with the controlled human evaluation now in progress.</p> 543 544 <div class="roadmap-grid"> 545 <div> 546 <div class="roadmap-col-label done">â v5 — Delivered</div> 547 <div class="roadmap-item"> 548 <div class="roadmap-dot done"></div> 549 <div class="roadmap-text"><strong>Five Study Types</strong> — predictive modeling, three causal-inference designs (selection-on-observables, individualized treatment regimes, difference-in-differences) and psychometrics</div> 550 </div> 551 <div class="roadmap-item"> 552 <div class="roadmap-dot done"></div> 553 <div class="roadmap-text"><strong>Ten Certified Estimators</strong> — each validated against synthetic data with a known true effect before being allowed near a real dataset</div> 554 </div> 555 <div class="roadmap-item"> 556 <div class="roadmap-dot done"></div> 557 <div class="roadmap-text"><strong>Four Curated Datasets</strong> — HSLS:09, ELS:2002, ASSISTments 2009–10 and a constructed ELS×HSLS cross-cohort panel, each with its own registry and quirks</div> 558 </div> 559 <div class="roadmap-item"> 560 <div class="roadmap-dot done"></div> 561 <div class="roadmap-text"><strong>Skill-Based Architecture</strong> — 70 composable skills across four layers, matched at runtime and injected per agent — capabilities are added as skills, not as longer prompts</div> 562 </div> 563 <div class="roadmap-item"> 564 <div class="roadmap-dot done"></div> 565 <div class="roadmap-text"><strong>Calibrated Review Gate</strong> — thresholds set at the 25th percentile of papers each venue actually published, with median-of-three sampling for borderline verdicts</div> 566 </div> 567 <div class="roadmap-item"> 568 <div class="roadmap-dot done"></div> 569 <div class="roadmap-text"><strong>Deterministic Honesty Guards</strong> — numeric reconciliation against analysis artifacts, unverified-claim flagging, and review-health checks that reject a truncated review rather than scoring it</div> 570 </div> 571 <div class="roadmap-item"> 572 <div class="roadmap-dot done"></div> 573 <div class="roadmap-text"><strong>Measured Token Accounting</strong> — per-call prompt / completion / cached token counts with rates in configuration, so past runs re-price without re-running</div> 574 </div> 575 <div class="roadmap-item"> 576 <div class="roadmap-dot done"></div> 577 <div class="roadmap-text"><strong>Open Source Release</strong> — EDM-ARS and LSAR both published, runnable from a clean checkout, with their test suites green</div> 578 </div> 579 </div> 580 581 <div> 582 <div class="roadmap-col-label next">ð Next</div> 583 <div class="roadmap-item"> 584 <div class="roadmap-dot next"></div> 585 <div class="roadmap-text"><strong>Controlled Evaluation</strong> — blind human review comparing EDM-ARS output against other automated research systems on matched topics — in progress</div> 586 </div> 587 <div class="roadmap-item"> 588 <div class="roadmap-dot next"></div> 589 <div class="roadmap-text"><strong>
589Wider Venue Calibration</strong> — anchor corpora for the eight venues that currently have review criteria but no calibrated threshold</div> 590 </div> 591 <div class="roadmap-item"> 592 <div class="roadmap-dot next"></div> 593 <div class="roadmap-text"><strong>Additional Datasets</strong> — PISA 2022 (plausible values) and ECLS-K, extending the registry approach to new survey designs</div> 594 </div> 595 <div class="roadmap-item"> 596 <div class="roadmap-dot next"></div> 597 <div class="roadmap-text"><strong>Reproducibility Packaging</strong> — release the locked specifications alongside published results so any run can be reproduced end to end</div> 598 </div> 599 </div> 600 </div> 601 </div> 602 </section> 603 604 <!-- ==================== FOOTER ==================== --> 605 <footer> 606 <div class="footer-left"> 607 <p><span>EDM-ARS</span> · <span>LSAR</span> · Educational Data Mining Research Tools</p> 608 <p style="font-size:12px; margin-top:4px;">Built by Chenguang Pan and his friends</p> 609 </div> 610 <div class="footer-links"> 611 <a href="https://github.com/cgpan/edm-ars-public" aria-label="View EDM-ARS on GitHub">EDM-ARS GitHub</a> 612 <a href="https://github.com/cgpan/LSAR-public" aria-label="View LSAR on GitHub">LSAR GitHub</a> 613 <a href="papers/edmars_tech_report_v2.pdf" aria-label="Read the technical report">Paper</a> 614 </div> 615 </footer> 616 617 <!-- Architecture run replay --> 618
618<script src="replay.js"></script>
618 619 620 <div class="visitor-map"> 621
621<script type="text/javascript" id="mapmyvisitors" src="//mapmyvisitors.com/map.js?d=WaY7B-DOz4UKe0AudlVCfZ2S67u4ebs9T7ZNI8mF1c0&cl=ffffff&w=300"></script>
621 622 </div> 623 624 <!-- Three.js Mars Canvas --> 625
625<script src="https://cdnjs.cloudflare.com/ajax/libs/three.js/r128/three.min.js"></script>
625 626
626<script src="script.js"></script>
626 627</body> 628</html>
Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.