Skip to content

nexu-io/open-design / plugins/_official/examples/hps-academic-paper

hps-academic-paper

Verified

A review deck on compositional generalization in large language models — the field map, the gap, the evidence, and open questions. Built as a decision-grade academic research deck for PI, lab group, reviewers.

FileKindSize
LICENSEother1.1 KB
SKILL.mdskill md9.9 KB
example.htmlother28.9 KB
open-design.jsonconfig5.5 KB

example.html

View on GitHub
<!DOCTYPE html>
<html lang="en" data-theme="academic-paper">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width,initial-scale=1">
<title>Write an Academic Review Deck like a Top-Journal Survey Author - Academic research</title>
<style>
/* hps-academic-paper :: webfonts (Google Fonts @import only — no other external deps) */
@import url('https://fonts.googleapis.com/css2?family=Playfair+Display:ital,wght@0,400;0,600;0,700;0,800;1,400;1,600&display=swap');
@import url('https://fonts.googleapis.com/css2?family=Noto+Serif+SC:wght@400;600;700&display=swap');
@import url('https://fonts.googleapis.com/css2?family=JetBrains+Mono:wght@400;500;700&display=swap');
</style>
<style>
/* ============================================================
   TOKEN SHEET — academic-paper, LOCKED.
   Ported from lewislulu/html-ppt-skill assets/themes/academic-paper.css
   on top of the upstream assets/base.css variable contract.
   All three font slots are serif: Latin Modern Roman (local LaTeX
   font when installed) → Playfair Display → Noto Serif SC → Georgia.
   ============================================================ */
:root{
  --bg:#fdfcf8; --bg-soft:#f7f5ed; --surface:#ffffff; --surface-2:#f5f3ea;
  --border:rgba(20,20,20,.14); --border-strong:rgba(20,20,20,.35);
  --text-1:#0a0a0a; --text-2:#333333; --text-3:#707070;
  --accent:#1a3a7a; --accent-2:#0a0a0a; --accent-3:#8a1a1a;
  --good:#1a5a2a; --warn:#8a6a1a; --bad:#8a1a1a;
  --grad:linear-gradient(135deg,#1a3a7a,#0a0a0a);
  --grad-soft:linear-gradient(135deg,#e8edf8,#f5f3ea);
  --radius:0px; --radius-sm:0px; --radius-lg:0px;
  --shadow:none;
  --shadow-lg:0 1px 2px rgba(0,0,0,.1);
  --font-sans:'Latin Modern Roman','Playfair Display','Noto Serif SC',Georgia,serif;
  --font-serif:'Latin Modern Roman','Playfair Display','Noto Serif SC',Georgia,serif;
  --font-mono:'JetBrains Mono',SFMono-Regular,Menlo,monospace;
  --font-display:'Latin Modern Roman','Playfair Display','Noto Serif SC',Georgia,serif;
  --letter-tight:-.01em; --letter-normal:0;
  --ease:cubic-bezier(.4,0,.2,1);
}
</style>
<style>
/* ============================================================
   BASE + SLIDE SYSTEM — horizontal scroll-snap deck.
   One <section class="slide"> = exactly one viewport screen.
   Designed at 16:9 (1280×720 baseline), fluid to the viewport.
   ============================================================ */
*,*::before,*::after{box-sizing:border-box}
html,body{margin:0;padding:0;height:100%}
body{background:var(--bg);color:var(--text-1);font-family:var(--font-serif);
  font-weight:400;line-height:1.6;letter-spacing:var(--letter-normal);
  -webkit-font-smoothing:antialiased;-moz-osx-font-smoothing:grayscale;overflow:hidden}
img,svg,video{max-width:100%;display:block}
code,kbd,pre,samp{font-family:var(--font-mono)}
a{color:var(--accent);text-decoration:underline;text-underline-offset:3px}

.deck{display:flex;width:100vw;height:100vh;overflow-x:auto;overflow-y:hidden;
  scroll-snap-type:x mandatory;scrollbar-width:none;-ms-overflow-style:none}
.deck::-webkit-scrollbar{display:none}
.slide{position:relative;flex:0 0 100vw;height:100vh;scroll-snap-align:start;
  display:flex;flex-direction:column;justify-content:center;
  padding:72px 96px;overflow:hidden}

/* ============ TYPOGRAPHY ============ */
.eyebrow{font-size:13px;font-weight:600;letter-spacing:.14em;text-transform:uppercase;color:var(--text-3);margin:0}
.kicker{font-size:clamp(14px,1.4vw,17px);font-weight:400;font-style:italic;color:var(--accent);
  letter-spacing:0;text-transform:none;margin:0 0 12px}
.h1{font-family:var(--font-display);font-size:clamp(40px,5vw,64px);line-height:1.12;font-weight:700;letter-spacing:var(--letter-tight);margin:0 0 18px;color:var(--text-1)}
.h2{font-family:var(--font-display);font-size:clamp(30px,3.8vw,48px);line-height:1.15;font-weight:700;letter-spacing:var(--letter-tight);margin:0 0 14px;color:var(--text-1)}
.h3{font-size:clamp(21px,2.4vw,30px);line-height:1.25;font-weight:700;margin:0 0 10px}
.h4{font-size:clamp(16px,1.7vw,21px);line-height:1.35;font-weight:700;margin:0 0 8px}
.lede{font-size:clamp(16px,1.7vw,21px);line-height:1.6;color:var(--text-2);font-weight:400;max-width:64ch;margin:0}
.dim{color:var(--text-2)}
.dim2{color:var(--text-3)}
.mono{font-family:var(--font-mono)}
.serif{font-family:var(--font-serif)}

/* ============ LAYOUT PRIMITIVES ============ */
.row{display:flex;gap:24px;align-items:center}
.row.wrap{flex-wrap:wrap}
.grid{display:grid;gap:24px}
.g2{grid-template-columns:repeat(2,1fr)}
.g3{grid-template-columns:repeat(3,1fr)}
.g4{grid-template-columns:repeat(4,1fr)}
.center{display:flex;align-items:center;justify-content:center;text-align:center}
.mt-s{margin-top:8px}.mt-m{margin-top:18px}.mt-l{margin-top:32px}

/* ============ CARDS / PILLS (paper skin: hairline border, zero radius, no shadow) ============ */
.card{background:var(--surface);border:1px solid var(--border);border-radius:var(--radius);
  padding:26px 28px;box-shadow:var(--shadow);position:relative;overflow:hidden}
.card-soft{background:var(--surface-2)}
.card-accent{border-top:2px solid var(--text-1)}
.pill{display:inline-block;padding:4px 14px;border-radius:0;font-size:13px;font-weight:400;font-style:italic;
  background:transparent;color:var(--text-2);border:1px solid var(--border-strong)}
.pill-accent{color:var(--accent);border-color:var(--accent)}
.divider-accent{height:2px;width:100%;background:var(--text-1)}

/* ============ ACADEMIC DEVICES (theme signature) ============ */
/* Double rule — the LaTeX titlepage device: heavy line over hairline. */
.rule{height:1px;background:var(--border-strong);border:0;margin:0}
.rule-double{height:0;border:0;border-top:2px solid var(--text-1);position:relative;margin:0}
.rule-double::after{content:"";position:absolute;left:0;right:0;top:3px;border-top:1px solid var(--text-1)}
/* Abstract block — italic body, vertical hairlines like a quoted environment. */
.abstract{font-style:italic;color:var(--text-2);font-size:clamp(15px,1.55vw,19px);line-height:1.7;
  padding:18px 28px;border-top:1px solid var(--border-strong);border-bottom:1px solid var(--border-strong)}
.abstract b{font-style:normal;font-weight:700;color:var(--text-1)}
/* Citation anchor — link-blue, underlined, [n] in mono. */
.cite{font-family:var(--font-mono);font-size:.78em;color:var(--accent);text-decoration:underline;text-underline-offset:3px;font-style:normal}
/* Figure environment with numbered caption. */
.figure{border:1px solid var(--border);background:var(--surface);padding:26px 28px}
.figcaption{font-size:clamp(13px,1.3vw,15px);color:var(--text-2);margin-top:14px;font-style:italic}
.figcaption b{font-style:normal;font-weight:700;color:var(--text-1)}
/* Booktabs table — heavy top/bottom rules, hairline midrule, no vertical rules. */
.booktabs{border-collapse:collapse;width:100%;font-size:clamp(14px,1.45vw,18px)}
.booktabs th,.booktabs td{padding:10px 16px;text-align:left;border:0}
.booktabs td.num,.booktabs th.num{text-align:right;font-variant-numeric:tabular-nums}
.booktabs thead th{border-top:2px solid var(--text-1);border-bottom:1px solid var(--text-1);font-weight:700}
.booktabs tbody tr:last-child td{border-bottom:2px solid var(--text-1)}
.booktabs .best{font-weight:700;color:var(--accent)}
/* Footnote — hairline rule over small print. */
.footnote{font-size:13px;color:var(--text-3);border-top:1px solid var(--border-strong);padding-top:10px;max-width:48ch}
/* References list — hanging [n] labels. */
.references{list-style:none;margin:0;padding:0;font-size:clamp(14px,1.45vw,17px);color:var(--text-2)}
.references li{padding:7px 0 7px 52px;position:relative;line-height:1.55}
.references li>.cite{position:absolute;left:0;top:9px;text-decoration:none}
.references i{color:var(--text-1)}

/* ============ DECK CHROME ============ */
.deck-header{position:fixed;top:24px;left:48px;right:48px;display:flex;align-items:center;justify-content:space-between;
  font-size:12px;color:var(--text-3);font-style:italic;z-index:10;pointer-events:none}
.deck-footer{position:fixed;bottom:24px;left:48px;right:48px;display:flex;align-items:center;justify-content:space-between;
  font-size:12px;color:var(--text-3);z-index:10;pointer-events:none}
.theme-chip{font-family:var(--font-mono);font-size:11px;font-style:normal;letter-spacing:.04em}
.progress-bar{position:fixed;left:0;right:0;bottom:0;height:2px;z-index:20;pointer-events:none}
.progress-bar>span{display:block;height:100%;width:0;background:var(--accent);transition:width .3s var(--ease)}

/* ============ SLIDE-LOCAL COMPONENTS ============ */
.section-num{position:absolute;right:72px;bottom:40px;font-size:clamp(120px,17vw,220px);font-weight:700;line-height:.9;
  color:var(--surface-2);letter-spacing:-.02em;z-index:0;pointer-events:none;font-family:var(--font-display)}
.slide>*{position:relative;z-index:1}
.kpi .num{font-size:clamp(34px,4vw,52px);font-weight:700;letter-spacing:var(--letter-tight);line-height:1.1;font-variant-numeric:tabular-nums}
.kpi .lbl{font-size:14px;color:var(--text-3);margin-top:6px}
.stat-big{font-size:clamp(100px,14vw,190px);font-weight:700;line-height:1;letter-spacing:-.02em;margin:0;font-variant-numeric:tabular-nums}
.bar-chart{display:flex;align-items:flex-end;gap:26px;height:240px;margin-top:46px;max-width:780px;
  border-bottom:2px solid var(--text-1);padding:0 12px}
.bar-chart .bar{flex:1;background:var(--accent);border-radius:0;position:relative;min-height:14px}
.bar-chart .bar.muted{background:var(--surface-2);border:1px solid var(--border-strong)}
.bar-chart .bar span{position:absolute;bottom:-32px;left:0;right:0;text-align:center;font-size:13px;font-style:italic;color:var(--text-3)}
.bar-chart .bar em{position:absolute;top:-30px;left:0;right:0;text-align:center;font-size:14px;font-weight:700;font-style:normal;color:var(--text-1);font-variant-numeric:tabular-nums}
.toc-list{list-style:none;margin:0;padding:0;counter-reset:toc;max-width:760px}
.toc-list li{counter-increment:toc;display:flex;align-items:baseline;gap:18px;padding:16px 4px;
  border-bottom:1px solid var(--border);font-size:clamp(18px,2vw,25px)}
.toc-list li::before{content:counter(toc) ".";font-weight:700;color:var(--accent);min-width:34px;font-variant-numeric:tabular-nums}
.toc-list li .dim2{margin-left:auto;font-size:.7em;font-style:italic}
.quote{font-family:var(--font-serif);font-size:clamp(28px,3.4vw,44px);line-height:1.4;font-weight:400;font-style:italic;max-width:26ch;margin:0}
.check{list-style:none;padding:0;margin:0;font-size:clamp(15px,1.5vw,19px)}
.check li{padding:9px 0 9px 32px;position:relative;color:var(--text-2)}
.check li::before{content:"";position:absolute;left:2px;top:13px;width:14px;height:14px;
  border:1.5px solid var(--text-1);background:transparent}
.check li::after{content:"";position:absolute;left:6px;top:16px;width:8px;height:4px;
  border-left:1.5px solid var(--accent);border-bottom:1.5px solid var(--accent);transform:rotate(-45deg)}

/* ============ SPEAKER NOTES (hidden from audience) ============ */
.notes{display:none!important}

/* ============ ANIMATIONS (restrained entry effects) ============ */
@keyframes kf-rise{from{opacity:0;transform:translateY(40px)}to{opacity:1;transform:none}}
@keyframes kf-fade-up{from{opacity:0;transform:translateY(24px)}to{opacity:1;transform:none}}
.anim-rise-in{animation:kf-rise .8s var(--ease) both}
.anim-fade-up{animation:kf-fade-up .6s var(--ease) both}
.anim-stagger-list>*{opacity:0;animation:kf-rise .6s var(--ease) both}
.anim-stagger-list>*:nth-child(1){animation-delay:.05s}
.anim-stagger-list>*:nth-child(2){animation-delay:.15s}
.anim-stagger-list>*:nth-child(3){animation-delay:.25s}
.anim-stagger-list>*:nth-child(4){animation-delay:.35s}
.anim-stagger-list>*:nth-child(n+5){animation-delay:.45s}
@media (prefers-reduced-motion: reduce){
  .anim-rise-in,.anim-fade-up,.anim-stagger-list>*{animation:none;opacity:1;transform:none}
  .progress-bar>span{transition:none}
}

/* ============ PRINT ============ */
@media print{
  body{overflow:visible}
  .deck{display:block;width:auto;height:auto;overflow:visible}
  .slide{width:100vw;height:100vh;page-break-after:always}
  .deck-header,.deck-footer,.progress-bar{display:none!important}
}
</style>
<style id="od-no-pager">.deck-controls,#counter,.deck-counter{display:none!important}</style>
</head>
<body>

<div class="deck-header"><span>Compositional Generalization in LLMs · A Survey</span><span class="theme-chip" id="theme-chip">academic-paper</span></div>

<div class="deck" id="deck">

  <!-- 1 · layout: cover (LaTeX titlepage: double rule above, hairline below) -->
  <section class="slide" data-title="Title">
    <p class="kicker">Academic research · Survey deck for lab group review</p>
    <div class="rule-double" style="max-width:880px"></div>
    <h1 class="h1 anim-rise-in" style="margin-top:30px;max-width:20ch">Compositional Generalization in Large Language Models: What We Know, What We Don't</h1>
    <p class="lede" style="margin-bottom:26px">A field map of 63 studies (2019–2026), the gap they keep finding, and the question still open.</p>
    <hr class="rule" style="max-width:880px">
    <p class="dim2 mt-m" style="font-style:italic">Prepared for PI review &nbsp;·&nbsp; Draft v3 &nbsp;·&nbsp; NLP &amp; Reasoning Lab &nbsp;·&nbsp; July 2026</p>
    <div class="notes">Title slide in LaTeX titlepage grammar: italic kicker for the positioning line, double rule over the title, support line below, hairline rule, then the badge row in small print. Everything is serif, ink on paper-white.</div>
  </section>

  <!-- 2 · layout: toc (numbered outline with hairline rules) -->
  <section class="slide" data-title="Outline">
    <p class="kicker">Outline</p>
    <h2 class="h2">Contents</h2>
    <ol class="toc-list mt-m anim-stagger-list">
      <li>The field map: how compositionality gets tested <span class="dim2">3 min</span></li>
      <li>The gap: where systematic generalization breaks <span class="dim2">6 min</span></li>
      <li>Method &amp; evidence: what 63 studies show <span class="dim2">7 min</span></li>
      <li>Contribution, open questions &amp; discussion <span class="dim2">4 min</span></li>
    </ol>
    <div class="notes">A paper talk outline reads like a table of contents: numbered entries, hairline separators, timing column in italic small print.</div>
  </section>

  <!-- 3 · layout: section-divider -->
  <section class="slide" data-title="Section 1">
    <span class="section-num">1</span>
    <p class="eyebrow">Section one</p>
    <h2 class="h2" style="max-width:18ch">The field map: how compositionality gets tested</h2>
    <div class="divider-accent mt-m" style="max-width:240px"></div>
    <div class="notes">Section divider: oversized ghost numeral in the paper margin, ink rule under the section title.</div>
  </section>

  <!-- 4 · layout: abstract + contributions (two-column) -->
  <section class="slide" data-title="Abstract">
    <p class="kicker">§1 · Introduction</p>
    <h2 class="h2">Abstract &amp; the gap</h2>
    <div class="grid g2 mt-l">
      <div class="abstract anim-fade-up">
        <b>Abstract.</b> Since SCAN <span class="cite">[1]</span> first showed seq2seq models fail to recombine known primitives into novel commands, a decade of benchmarks — COGS, CFQ, COGNATE — has asked whether scale closes that gap. We synthesize 63 studies spanning 2019–2026 and find the gap narrows but does not close: even 70B-parameter instruction-tuned models lose <b>28–41 accuracy points</b> on held-out compositions versus in-distribution ones. No architecture change or scale increase has eliminated it.
      </div>
      <div>
        <ul class="check">
          <li>The gap is systematic, not noise: it reappears across 6 benchmark families and 11 model families</li>
          <li>Scale reduces but does not close it — the curve bends, never reaches zero</li>
          <li>In-context exemplars help more than parameter count does, per unit compute</li>
          <li>No published method fully separates memorization from genuine recombination</li>
        </ul>
      </div>
    </div>
    <div class="notes">The abstract environment: italic body between two hairline rules, bold lead word, citation markers in link blue. Contributions as a squared checklist on the right.</div>
  </section>

  <!-- 5 · layout: stat-highlight -->
  <section class="slide center" data-title="Headline result">
    <p class="kicker">Headline number</p>
    <p class="stat-big anim-rise-in" style="color:var(--accent)">34pp</p>
    <p class="lede mt-m" style="max-width:44ch">Median accuracy drop from in-distribution to held-out novel compositions, pooled across the 63 surveyed studies.</p>
    <p class="footnote mt-l" style="text-align:left">† 63 studies · 11 model families (350M–70B params) · 6 benchmark families · gap shrinks with scale but never reaches zero (min. observed: 9pp, largest model, best prompting).</p>
    <div class="notes">One number carries the slide, set in link-blue ink. The dagger footnote with its hairline rule is the academic signature: the headline figure arrives with its supporting counts.</div>
  </section>

  <!-- 6 · layout: method figure (inline SVG diagram) -->
  <section class="slide" data-title="Method">
    <p class="kicker">§2 · Method</p>
    <h2 class="h2">The systematic-split coding protocol</h2>
    <div class="figure mt-m" style="max-width:880px">
      <svg viewBox="0 0 800 190" role="img" aria-label="Survey coding protocol diagram" style="width:100%;height:auto">
        <defs>
          <marker id="arr" viewBox="0 0 10 10" refX="9" refY="5" markerWidth="7" markerHeight="7" orient="auto-start-reverse">
            <path d="M0,0 L10,5 L0,10 z" fill="#0a0a0a"/>
          </marker>
        </defs>
        <g font-family="Georgia,serif" font-size="15" fill="#0a0a0a">
          <rect x="10" y="55" width="160" height="56" fill="#ffffff" stroke="#0a0a0a" stroke-width="1.5"/>
          <text x="90" y="79" text-anchor="middle">Extract</text>
          <text x="90" y="98" text-anchor="middle">primitives &amp; splits</text>
          <rect x="240" y="55" width="160" height="56" fill="#ffffff" stroke="#0a0a0a" stroke-width="1.5"/>
          <text x="320" y="79" text-anchor="middle">Normalize</text>
          <text x="320" y="98" text-anchor="middle">metric definitions</text>
          <rect x="470" y="55" width="160" height="56" fill="#e8edf8" stroke="#1a3a7a" stroke-width="1.5"/>
          <text x="550" y="79" text-anchor="middle" fill="#1a3a7a">Pool the</text>
          <text x="550" y="98" text-anchor="middle" fill="#1a3a7a">gap estimate</text>
          <rect x="660" y="20" width="130" height="50" fill="#ffffff" stroke="#0a0a0a" stroke-width="1.5"/>
          <text x="725" y="50" text-anchor="middle">Scale trend</text>
          <rect x="660" y="100" width="130" height="50" fill="#ffffff" stroke="#0a0a0a" stroke-width="1.5"/>
          <text x="725" y="130" text-anchor="middle">Method trend</text>
          <line x1="170" y1="83" x2="236" y2="83" stroke="#0a0a0a" stroke-width="1.5" marker-end="url(#arr)"/>
          <line x1="400" y1="83" x2="466" y2="83" stroke="#0a0a0a" stroke-width="1.5" marker-end="url(#arr)"/>
          <line x1="630" y1="70" x2="656" y2="50" stroke="#0a0a0a" stroke-width="1.5" marker-end="url(#arr)"/>
          <line x1="630" y1="96" x2="656" y2="120" stroke="#0a0a0a" stroke-width="1.5" marker-end="url(#arr)"/>
          <text x="640" y="38" font-style="italic" font-size="13" fill="#333333">by size</text>
          <text x="640" y="142" font-style="italic" font-size="13" fill="#333333">by design</text>
        </g>
      </svg>
      <p class="figcaption"><b>Figure 1:</b> Each of the 63 studies is re-coded onto one shared metric — accuracy on a held-out systematic split minus accuracy on an i.i.d. split of matched difficulty — so gap sizes are comparable across SCAN, COGS, and CFQ-style benchmarks before pooling.</p>
    </div>
    <div class="notes">Figures are ink-line SVG boxes with serif labels, never screenshots. The numbered caption in italic is mandatory — every figure must be citable.</div>
  </section>

  <!-- 7 · layout: chart-bar (CSS bars, Figure 2) -->
  <section class="slide" data-title="Results">
    <p class="kicker">§3 · Evidence</p>
    <h2 class="h2">The gap shrinks with scale, but plateaus above zero</h2>
    <div class="bar-chart">
      <div class="bar muted" style="height:98%"><em>41pp</em><span>350M</span></div>
      <div class="bar muted" style="height:78%"><em>33pp</em><span>1.3B</span></div>
      <div class="bar muted" style="height:60%"><em>26pp</em><span>7B</span></div>
      <div class="bar muted" style="height:48%"><em>19pp</em><span>13B</span></div>
      <div class="bar" style="height:30%"><em>12pp</em><span>70B</span></div>
    </div>
    <p class="figcaption" style="margin-top:52px;max-width:760px"><b>Figure 2:</b> Mean held-out accuracy gap (i.i.d. minus systematic split) by parameter count, pooled across 6 benchmark families and 34 of the 63 studies that report a clean scale sweep. The 70B point in link blue is the best reported result — still 12 points, not zero.</p>
    <div class="notes">A results chart in paper grammar: muted hatch-grey bars against one blue lead bar, sitting on a heavy ink axis, with a numbered caption underneath. Pure CSS — no chart library.</div>
  </section>

  <!-- 8 · layout: booktabs table (Table 1) -->
  <section class="slide" data-title="Comparison">
    <p class="kicker">§3 · Comparison</p>
    <h2 class="h2">Which interventions actually move the gap</h2>
    <div class="mt-l" style="max-width:880px">
      <table class="booktabs">
        <thead>
          <tr><th>Intervention</th><th class="num">Studies</th><th class="num">Median Δgap</th><th class="num">Best case</th><th>Robustness</th></tr>
        </thead>
        <tbody>
          <tr><td>Scale alone (10x params)</td><td class="num">19</td><td class="num">−6pp</td><td class="num">−11pp</td><td class="dim2">consistent, sub-linear</td></tr>
          <tr><td>Data augmentation</td><td class="num">14</td><td class="num">−9pp</td><td class="num">−17pp</td><td class="dim2">benchmark-specific</td></tr>
          <tr><td>Chain-of-thought prompting</td><td class="num">11</td><td class="num">−8pp</td><td class="num">−15pp</td><td class="dim2">fragile to phrasing</td></tr>
          <tr><td><b>Structured decoding (ours-adjacent)</b></td><td class="num best">9</td><td class="num best">−15pp</td><td class="num best">−22pp</td><td class="dim2">most robust, least tested at scale</td></tr>
        </tbody>
      </table>
      <p class="figcaption"><b>Table 1:</b> Structured-decoding methods post the largest median gap reduction (−15pp) but appear in only 9 of 63 studies and none above 13B parameters — the evidence base is real but thin, which is why we flag it as promising rather than proven.</p>
    </div>
    <div class="notes">Booktabs discipline: heavy top and bottom rules, one hairline under the header, no vertical rules ever. Best numbers bolded in the accent blue, license terms in italic small print.</div>
  </section>

  <!-- 9 · layout: big-quote (reviewer epigraph) -->
  <section class="slide center" data-title="Quote">
    <p class="quote anim-fade-up">“The field keeps rediscovering the same gap under a new benchmark name. What's missing isn't another leaderboard — it's a shared unit for the gap itself.”</p>
    <p class="eyebrow mt-l">— Anonymous reviewer, ICLR 2026 area chair notes</p>
    <div class="notes">An epigraph page: one italic serif quotation, attribution in small caps style. In an academic deck the quote is a citation, so it names its source precisely.</div>
  </section>

  <!-- 10 · layout: conclusions + future work -->
  <section class="slide" data-title="Conclusions">
    <p class="kicker">§4 · Discussion</p>
    <h2 class="h2">Contribution &amp; the next question</h2>
    <div class="grid g2 mt-l">
      <div class="card card-accent">
        <h4 class="h4">What this survey contributes</h4>
        <ul class="check">
          <li>One pooled gap metric across 6 benchmark families — prior surveys compared scores, not gaps</li>
          <li>First scale sweep spanning 350M–70B on a matched split protocol</li>
          <li>A ranked intervention table showing structured decoding is promising but under-tested, not proven</li>
        </ul>
      </div>
      <div class="card">
        <h4 class="h4">Open questions we do not answer</h4>
        <ul class="check">
          <li>Does the gap reach zero at any feasible scale, or is there a floor?</li>
          <li>Can structured decoding's −15pp median hold above 13B params?</li>
          <li>Is the residual gap measuring recombination failure, or benchmark artifact?</li>
        </ul>
      </div>
    </div>
    <p class="footnote mt-l">We claim a better measurement, not a solved problem: the 12pp floor at 70B remains unexplained. Coded dataset and split protocol available on request.</p>
    <div class="notes">Closing argument in two ruled cards — claims on the left under a heavy top rule, open problems on the right. The artifact link footnote signals reproducibility.</div>
  </section>

  <!-- 11 · layout: thanks + references -->
  <section class="slide" data-title="References">
    <p class="kicker">Thank you · Questions welcome</p>
    <h2 class="h2">Discussion &amp; selected references</h2>
    <ol class="references mt-m anim-stagger-list" style="max-width:920px">
      <li><span class="cite">[1]</span> <i>Lake &amp; Baroni.</i> Generalization without systematicity: SCAN. <i>ICML</i>, 2018.</li>
      <li><span class="cite">[2]</span> <i>Kim &amp; Linzen.</i> COGS: A compositional generalization challenge based on semantic interpretation. <i>EMNLP</i>, 2020.</li>
      <li><span class="cite">[3]</span> <i>Keysers et al.</i> Measuring compositional generalization: CFQ. <i>ICLR</i>, 2020.</li>
      <li><span class="cite">[4]</span> <i>Full 63-study coding sheet and split protocol</i> — available from the corresponding author on request.</li>
    </ol>
    <hr class="rule mt-l" style="max-width:920px">
    <p class="dim2 mt-m" style="font-style:italic">NLP &amp; Reasoning Lab · ← → navigate, #/6 deep-links Figure 1</p>
    <div class="notes">A paper talk ends on its bibliography, repurposed as a call to action: hanging [n] labels in link blue, actions in italic. Links line in small print.</div>
  </section>

</div>

<div class="deck-footer"><span>hps-academic-paper · MIT · based on lewislulu/html-ppt-skill</span><span id="counter">1 / 11</span></div>
<div class="progress-bar"><span id="bar"></span></div>

<script>
(function () {
  var deck = document.getElementById('deck');
  var slides = Array.prototype.slice.call(document.querySelectorAll('.slide'));
  var counter = document.getElementById('counter');
  var bar = document.getElementById('bar');
  var active = 0;
  var lastKeyEvent = null;

  function clamp(i) { return Math.max(0, Math.min(slides.length - 1, i)); }

  function paint(i) {
    counter.textContent = (i + 1) + ' / ' + slides.length;
    bar.style.width = (((i + 1) / slides.length) * 100) + '%';
  }

  // Hash routing: #/N (1-based). replaceState keeps back-button clean and
  // never throws inside a srcdoc iframe thanks to the try/catch.
  function setHash(i) {
    try { history.replaceState(null, '', '#/' + (i + 1)); } catch (_) { /* sandboxed */ }
  }
  function fromHash() {
    var m = (location.hash || '').match(/^#\/(\d+)/);
    return m ? clamp(parseInt(m[1], 10) - 1) : 0;
  }

  function go(i, instant) {
    active = clamp(i);
    deck.scrollTo({ left: active * deck.clientWidth, behavior: instant ? 'auto' : 'smooth' });
    paint(active);
    setHash(active);
  }

  function syncFromScroll() {
    var i = Math.round(deck.scrollLeft / deck.clientWidth);
    if (i !== active && i >= 0 && i < slides.length) { active = i; paint(i); setHash(i); }
  }

  function onKey(e) {
    if (e === lastKeyEvent) return; // dedupe: same Event passes both listeners
    lastKeyEvent = e;
    var t = e.target;
    if (t && (t.tagName === 'INPUT' || t.tagName === 'TEXTAREA')) return;
    if (e.key === 'ArrowRight' || e.key === ' ' || e.key === 'PageDown') { e.preventDefault(); go(active + 1); }
    else if (e.key === 'ArrowLeft' || e.key === 'PageUp') { e.preventDefault(); go(active - 1); }
    else if (e.key === 'Home') { e.preventDefault(); go(0); }
    else if (e.key === 'End') { e.preventDefault(); go(slides.length - 1); }
  }

  // Listen on both window and document in capture phase so keys work no
  // matter which element holds focus inside a host iframe.
  window.addEventListener('keydown', onKey, true);
  document.addEventListener('keydown', onKey, true);
  deck.addEventListener('scroll', function () { window.requestAnimationFrame(syncFromScroll); }, { passive: true });
  window.addEventListener('hashchange', function () { go(fromHash()); });
  window.addEventListener('resize', function () { go(active, true); });

  // Auto-focus so arrow keys work without a click.
  document.body.setAttribute('tabindex', '-1');
  document.body.style.outline = 'none';
  function focusDeck() {
    try { window.focus(); document.body.focus({ preventScroll: true }); } catch (_) { /* sandboxed */ }
  }
  document.addEventListener('mousedown', focusDeck);
  window.addEventListener('load', focusDeck);

  go(fromHash(), true);
  paint(active);
  focusDeck();
})();
</script>

</body>
</html>
add-product-e2e-evalpaperclipai/paperclipAdd or extend a Paperclip full-stack runner E2E workflow, fixture, matcher, or report evidence path for local or Daytona execution.Stars: 115kUpdated Claude CodeCodexVerifiedInstall
add-runner-evalpaperclipai/paperclipAdd or extend a Paperclip Runner protocol evaluation definition, roster, assertion, or report fixture with provenance and narrow validation.Stars: 115kUpdated Claude CodeCodexVerifiedInstall
diagnose-why-work-stoppedpaperclipai/paperclipCompatibility entry point for failed, stalled, looping, or over-recovered Paperclip work. Use fix-failed-paperclip-run for the full diagnosis, regression, repair, real verification, and green-PR workflow; honor an explicitly requested diagnosis-only scope.Stars: 115kUpdated Claude CodeCodexVerifiedInstall
fix-failed-paperclip-runpaperclipai/paperclipDiagnose and fix failed, stuck, slow, or incorrectly blocked Paperclip tasks and runs. Use with a task/run URL or failure report to trace the root cause, prove a regression red then green, check second-order effects, verify the real workflow, and deliver a PR with passing checks.Stars: 115kUpdated Claude CodeCodexVerifiedInstall