SabaPivot's picture
download
raw
58.7 kB
<!DOCTYPE html>
<!--
Vendored + adapted into posterly from ARIS (Auto-claude-code-research-in-sleep),
skill paper-poster-html. Origin: https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep
Upstream tokenization: MIT (c) 2026 wanshuiyin -- ../LICENSES/aris-MIT.txt
This is the tokenized form of posterly's own neutral template; the class names and
content are posterly's own (c) 2026 Ruishuo Chen. As an adapted derivative this file
ships as part of posterly under AGPL-3.0. See ../NOTICE.md.
-->
<!--
============================================================
TEMPLATE: landscape_4col (ARIS fork)
CANVAS: 60in × 36in landscape (ICML / NeurIPS / generic landscape)
LAYOUT: header → optional banner → 4 body columns → optional takeaways → footer
USE WHEN: standard ML conference poster with 3-5 content cards per column,
mix of figures + equations + small tables.
HOW TO USE:
1. Copy this file to your working directory as `poster.html`.
2. Edit the DESIGN TOKENS in `:root` to your lab/venue colors.
3. Replace TODO content placeholders (search for "TODO").
4. Run `python tools/run_gates.py poster.html --tokens design_tokens.json` (the default driver: preflight -> style -> measure -> polish; design_tokens.json is the design-direction pack written at lock time -- SKILL.md Step 2.5) to align columns and check style/structure.
5. Run `python tools/render_preview.py poster.html` to produce the PDF.
NOTE: as shipped, this scaffold passes `preflight` (structure) but is
EXPECTED to FAIL `measure`/`polish` -- figures are commented out and copy
is TODO stubs, so columns only fill the top of the canvas. Those two gates
judge a FILLED poster; balance them once you've added real content. See
templates/README.md ("Scaffolds, not finished posters").
MEASURE ROLES: every layout-critical element carries `data-measure-role`.
Generic measurement scripts depend on these — do not remove.
CANVAS RETARGETING: to print at another size, change the canvas in exactly
TWO places that must stay in sync -- the `@page { size: ... }` rule and the
`.poster { width/height }` (also the `@media print .poster { width/height }`).
Example: ICLR 2026 main conference uses 185cm 90cm landscape (official print
service spec) -- set @page to `185cm 90cm`, width to `1850 * var(--u)` and
height to `900 * var(--u)` (and the print-override .poster to `185cm`/`90cm`).
Adapted from posterly (MIT, © 2026 Ruishuo Chen) — see LICENSES/ & NOTICE.md;
ARIS modifications: flat de-gradient, --fs token scale, zero-inline-style
utilities, data-source/data-color-exempt contracts.
============================================================
-->
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Reproducing: Floating-Point Networks with AD Can Represent Almost All FP Functions and Their Gradients — ICML 2026</title>
<!-- MathJax v3 for inline equations. CDN by default when the file is
opened by hand; the posterly check tools (measure/polish/pack/…)
intercept this request and serve the skill's bundled copy
(assets/mathjax/tex-svg.js, MathJax 3.2.2), so gates render math
deterministically even offline. To make a hand-opened poster
offline too, copy that bundle next to the poster and point the
<script> `src` at it. -->
<script>
window.MathJax = {
tex: {
inlineMath: [['$', '$'], ['\\(', '\\)']],
displayMath: [['$$', '$$'], ['\\[', '\\]']],
packages: {'[+]': ['ams']}
},
svg: { fontCache: 'global' }
};
</script>
<script id="MathJax-script" async src="https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-svg.js"></script>
<style>
/* =========================================================
CANVAS — 60" × 36" landscape
========================================================= */
@page { size: 60in 36in; margin: 0; }
:root {
/* ===== DESIGN TOKENS ===== */
/* Primary accent (header underline, card-highlight bar, section nums, .keyword) */
--accent: #B03A2E; /* round-off red; white ink on it measures 6.05:1 (>= AA) */
--accent-deep: #7B1F16;
--accent-light: #FBEAE7;
--accent-soft: #F0C7C0;
--accent-ink: #FFFFFF; /* ink ON the accent (chips, callout, thead) */
/* Emphasis register for "ours / best" (table .ours row, "★" callouts).
Default is the warm gold; a redesign may choose any register that
keeps 4.5:1 with --emph-ink (see templates/THEMES.md Mechanism 1). */
--emph: #C9A24A;
--emph-soft: #FFF7E0;
--emph-ink: #14314A; /* ink ON the emphasis fill. NOT var(--accent-deep):
#1F4566 on the default gold measures 4.16:1 (< AA);
#14314A measures 5.58:1 (>= 4.5:1). */
/* Text */
--text-primary: #1A1A1A;
--text-secondary: #555555;
--text-muted: #888888;
/* Backgrounds */
--bg-page: #F5F6F9;
--bg-card: #FFFFFF;
--bg-card-tint: #FAFAFB;
--bg-emphasis: var(--accent-light);
/* Borders */
--border-soft: #D8D8D8;
--border-strong: var(--accent);
/* Screen-only dark mat behind the poster (the off-canvas viewport
background). Print resets html/body to white, so this never reaches
paper; kept as a token so no color literal lives outside this block. */
--bg-viewport: #2B2B2B;
/* Base unit. Print: 1mm; screen preview: 1.6px (~3.78px = 1mm at 96dpi). */
--u: 1.6px;
/* Font-size scale (9 archetypes). Every `font-size` in this template
references one of these. `calc(var(--fs-N) * k)` is permitted ONLY for
a COMPONENTS.md-defined variant (e.g. `.eqn--large`). */
--fs-1: calc(9 * var(--u)); /* micro label */
--fs-2: calc(10 * var(--u)); /* small caption */
--fs-3: calc(11 * var(--u)); /* caption / table */
--fs-4: calc(12 * var(--u)); /* body text */
--fs-5: calc(13 * var(--u)); /* equation / emphasis */
--fs-6: calc(15 * var(--u)); /* subtitle */
--fs-7: calc(16 * var(--u)); /* section title */
--fs-8: calc(22 * var(--u)); /* banner number */
--fs-9: calc(32 * var(--u)); /* main title */
/* Fonts. Override at :root if your venue mandates a specific family. */
--font-serif: "Charter", "Source Serif Pro", "Georgia", serif;
--font-sans: "Inter", "Helvetica Neue", sans-serif;
/* Shadows & watermark ink — tokenized so rules 1/3 stay literal-free
(the faint radial page tint is the ONE allowed literal outside this
block; rule 5 validates it WHEN ENABLED -- the style gate ships with rules 4-5 off). */
--shadow-screen: 0 0 60px rgba(0, 0, 0, 0.5);
--shadow-card: 0 calc(2 * var(--u)) calc(6 * var(--u)) rgba(176, 58, 46, 0.06);
--ornament-ink: rgba(29, 78, 216, 0.06);
/* Corner-radius scale. Every border-radius calc() multiplies by this;
1 = the shipped soft look, 0 = square/flat (see templates/THEMES.md). */
--rs: 1;
/* Figure mount: the ground behind paper figures (transparent PNGs sit on
it) and the keyline around them (.figure img / .ff-fig img). Re-theme
with the Axis 6 frame decision so figures sit mounted in the design
instead of pasted on it; --fig-frame: transparent = frameless mount.
QR backgrounds stay literal white -- that's scannability, not styling. */
--fig-bg: white;
--fig-frame: var(--border-soft);
/* ===== END DESIGN TOKENS ===== */
}
/* =========================================================
RESET + BASE
========================================================= */
* { box-sizing: border-box; margin: 0; padding: 0; }
/* =========================================================
BASE DEFENSES -- wrap & ink safety. KEEP this block in ANY
skeleton, including a fully custom one (copy it over and
EXTEND the selector lists with your own prose/display
classes -- a hand-rolled skeleton that drops it strands
single-word widows and ragged titles; polish warns as
TEXT-WRAP).
- pretty: fills each line, protects the last-line orphan
(single-word widow). Safe on all prose.
- balance: evens CENTERED display text only -- never pair
it with left-aligned multi-sentence prose (SKILL.md wrap
rules).
- Any inline class you add that paints a background (a
.mark highlight, a keyword chip) must declare its own
`color` (Gate G -- never inherit ink across a ground
change) plus `box-decoration-break: clone;
-webkit-box-decoration-break: clone;` so a wrapped
highlight keeps its padding on both fragments (then it
never needs &nbsp;-gluing to hold one line).
========================================================= */
p, li, dd, figcaption,
.body-text, .caption, .callout, .section-title { text-wrap: pretty; }
.title { text-wrap: balance; }
html, body {
background: var(--bg-viewport);
font-family: var(--font-serif);
color: var(--text-primary);
-webkit-font-smoothing: antialiased;
}
/* =========================================================
UTILITY CLASSES — replace inline `style=` in the body markup.
The template (and any finished poster) must carry ZERO `style=`
attributes. Use these classes instead. Two documented exceptions
exist ONLY as examples and may appear in comments, never live:
- a logo/seal SVG flagged data-color-exempt="logo"
- a paper figure `<img ... data-source="paper" style="width: NN%">`
(AR width tweak; prefer the .w-NN classes below where possible)
========================================================= */
/* Poster container — exact print dimensions; data-measure-role="poster" so poster_check.py
can verify the canvas size. */
.poster {
width: calc(1524 * var(--u));
height: calc(914 * var(--u));
background: var(--bg-page);
/* Page tint kept (radial only, all color stops alpha <= 0.06 per the
de-gradient policy: linear-gradient is banned, this low-alpha radial
wash is the single allowed exception). */
background-image:
radial-gradient(ellipse at top left, rgba(29, 78, 216, 0.06), transparent 40%),
radial-gradient(ellipse at bottom right, rgba(201, 162, 74, 0.05), transparent 50%);
margin: 20px auto;
padding: calc(10 * var(--u)) calc(14 * var(--u));
display: grid;
grid-template-columns: minmax(0, 1fr); /* column-axis twin of the minmax(0,1fr) body-row defense: without it the implicit `auto` column grows to a wide child's max-content and a full-width row overflows the canvas (measure's canvas-overflow gate). Keep on any custom skeleton. */
grid-template-rows: auto auto minmax(0, 1fr) auto; /* header | banner | body | footer. minmax(0,1fr) (not bare 1fr): an over-tall body compresses instead of pushing footer off-canvas */
gap: calc(6 * var(--u));
box-shadow: var(--shadow-screen);
position: relative;
overflow: hidden;
}
/* Decorative top bar — flat solid accent (de-gradient). */
.poster::before {
content: "";
position: absolute; top: 0; left: 0; right: 0;
height: calc(8 * var(--u));
background: var(--accent);
}
/* =========================================================
HEADER
========================================================= */
.header {
display: grid;
grid-template-columns: 1fr minmax(50%, auto) 1fr; /* equal side tracks (1fr) -> the title track is centred on the poster, not just between the side blocks; the centre track is floored at 50% so a one-line title still fills it (else a short title measures narrow and trips a false HEADER/TITLE-SQUEEZED). Best-effort: a side block wide enough to clamp its 1fr track can still pull the title off-centre. */
align-items: center;
gap: calc(16 * var(--u));
padding: calc(2 * var(--u)) calc(4 * var(--u)) calc(5 * var(--u));
border-bottom: calc(2 * var(--u)) solid var(--accent);
}
.venue-badge {
justify-self: start; /* anchor to the far-left margin so the centre track stays centred */
display: flex; flex-direction: column;
align-items: center; justify-content: center;
min-width: calc(95 * var(--u));
text-align: center;
border-right: calc(1 * var(--u)) solid var(--border-soft);
padding-right: calc(12 * var(--u));
}
.venue-badge .vb-venue {
font-family: var(--font-sans);
font-weight: 800;
font-size: var(--fs-9);
color: var(--accent-deep);
line-height: 1;
letter-spacing: -0.5px;
}
.venue-badge .vb-year {
font-family: var(--font-sans);
font-size: var(--fs-5);
color: var(--text-secondary);
margin-top: calc(3 * var(--u));
letter-spacing: 1.2px;
}
.venue-badge .vb-tag {
font-family: var(--font-sans);
font-size: var(--fs-2);
color: var(--accent);
font-weight: 700;
margin-top: calc(2 * var(--u));
letter-spacing: 1.2px;
}
.title-block { text-align: center; min-width: 0; }
.title {
font-family: var(--font-sans);
font-weight: 800;
font-size: var(--fs-9);
line-height: 1.05;
color: var(--accent-deep);
letter-spacing: -0.5px;
}
.title .accent { color: var(--accent); }
.subtitle {
font-family: var(--font-sans);
font-weight: 500;
font-size: var(--fs-6);
color: var(--text-secondary);
margin-top: calc(2 * var(--u));
font-style: italic;
}
.authors-line {
font-family: var(--font-sans);
font-size: var(--fs-4);
color: var(--accent);
font-weight: 600;
margin-top: calc(3 * var(--u));
}
.authors-line .author { margin: 0 calc(4 * var(--u)); }
.authors-line sup { font-size: 0.7em; color: var(--accent); }
.authors-line .aff {
color: var(--text-secondary);
font-weight: 400;
display: block;
margin-top: calc(2 * var(--u));
font-size: var(--fs-4);
}
.right-block {
justify-self: end; /* keep logo+QR in the far-right corner */
display: flex; align-items: center;
gap: calc(10 * var(--u));
}
.qr-block { display: flex; flex-direction: column; align-items: center; gap: calc(2 * var(--u)); }
.qr-block img {
width: calc(85 * var(--u));
height: calc(85 * var(--u));
border: calc(2 * var(--u)) solid var(--accent);
border-radius: calc(4 * var(--u) * var(--rs));
background: white;
padding: calc(2 * var(--u));
}
.qr-label {
font-family: var(--font-sans);
font-size: var(--fs-3);
color: var(--accent);
font-weight: 600;
}
/* Optional logo slot — drop your lab logo here. Keep its rendered height
close to the QR (~85u) so the header doesn't grow disproportionately.
A raster logo needs no exemption. An inline-SVG logo with brand colors
must carry data-color-exempt="logo" so style_check skips its fills, e.g.:
<svg data-color-exempt="logo" ...> ... </svg> */
.logo-slot img { height: calc(85 * var(--u)); width: auto; max-width: calc(360 * var(--u)); object-fit: contain; display: block; }
/* Logo size classes on .logo-slot -- pick from the file's aspect ratio (SKILL.md Gate E).
tall/square sit at the QR-matched slot height; wide is intentionally shorter (~68% of the
QR) and width-capped so a long wordmark doesn't out-mass the title. Do NOT combine these
with logo-stack (that row is width-normalized, below). */
.logo-slot.logo-tall img,
.logo-slot.logo-square img { height: calc(85 * var(--u)); }
.logo-slot.logo-wide img { height: calc(58 * var(--u)); max-width: calc(300 * var(--u)); }
/* Logo chip: a solid backing that keeps a transparent / white-bordered logo legible on a
colored/dark header (white chip) or a light header (.logo-chip-dark); the padding + radius
fold a stray white box into a deliberate rounded tile (SKILL.md Gate E, Background). */
.logo-chip {
display: inline-flex; align-items: center; justify-content: center;
background: var(--bg-card); border-radius: calc(3 * var(--u) * var(--rs));
padding: calc(3 * var(--u)) calc(5 * var(--u));
}
.logo-chip.logo-chip-dark { background: var(--text-primary); }
/* logo-row: institution logos in the header right block (REAL logos the
user provided — never fabricate a seal). Each img MUST carry
data-color-exempt="logo"; height pairs with the QR. */
.logo-row { display: flex; align-items: center; gap: calc(5 * var(--u)); }
.logo-row img { height: calc(68 * var(--u)); width: auto; display: block; }
/* --boxed variant: each logo in a labeled tile (logo + institution name) —
more presence at poster distance; white tile + soft border, no new hues. */
.logo-row .lr-item {
display: flex; flex-direction: column; align-items: center;
gap: calc(2 * var(--u));
background: var(--bg-card);
border: 1px solid var(--border-soft);
border-radius: calc(3 * var(--u) * var(--rs));
padding: calc(4 * var(--u)) calc(6 * var(--u));
}
.logo-row .lr-item img { height: calc(58 * var(--u)); }
.logo-row .lr-label {
font-family: var(--font-sans); font-weight: 600; font-size: var(--fs-1);
color: var(--text-secondary); text-align: center; line-height: 1.15;
}
/* logo-stack variant: WIDE WORDMARKS (AR >= ~2) normalized to EQUAL WIDTH
and stacked vertically, left-aligned. Use when two wide wordmarks of
different aspect ratio read unbalanced height-matched in a row (equal
width lets the less-wide mark grow taller and aligns a clean block).
NOT for a square seal or tall mark (equal width blows it up — keep those
height-matched). Do NOT also apply the logo-wide/tall/square classes. */
.logo-row.logo-stack { flex-direction: column; align-items: flex-start; gap: calc(8 * var(--u)); }
.logo-row.logo-stack img { width: calc(170 * var(--u)); height: auto; }
/* venue badge may carry the official venue logo above its text line */
.venue-badge img { height: calc(62 * var(--u)); width: auto; display: block; margin: 0 auto calc(2 * var(--u)); }
.venue-badge .vb-title { font-family: var(--font-sans); font-weight: 800; font-size: var(--fs-5); color: var(--accent-deep); letter-spacing: 0.5px; }
/* =========================================================
OPTIONAL FRAMEWORK BANNER (delete this whole section if you don't need it)
========================================================= */
.framework-banner {
display: flex;
align-items: center;
gap: calc(16 * var(--u));
/* Flat solid emphasis fill (de-gradient). */
background: var(--bg-emphasis);
border: calc(1 * var(--u)) solid var(--border-soft);
border-left: calc(6 * var(--u)) solid var(--accent);
border-radius: calc(6 * var(--u) * var(--rs));
padding: calc(6 * var(--u)) calc(14 * var(--u));
}
.framework-banner img { height: calc(120 * var(--u)); width: auto; display: block; }
.framework-banner .banner-stats {
flex: 1;
display: grid;
grid-template-columns: repeat(4, 1fr);
gap: calc(6 * var(--u));
}
.framework-banner .bs-item {
background: white;
border: calc(1 * var(--u)) solid var(--border-soft);
border-left: calc(3 * var(--u)) solid var(--accent);
border-radius: calc(3 * var(--u) * var(--rs));
padding: calc(2 * var(--u)) calc(8 * var(--u));
text-align: center;
}
.framework-banner .bs-num {
font-family: var(--font-sans);
font-weight: 800;
font-size: var(--fs-8);
color: var(--accent);
line-height: 1;
}
.framework-banner .bs-label {
font-family: var(--font-sans);
font-size: var(--fs-3);
color: var(--text-secondary);
margin-top: calc(2 * var(--u));
line-height: 1.2;
}
.framework-banner .fb-text {
flex: 1.6;
font-family: var(--font-serif);
font-size: var(--fs-6);
line-height: 1.5;
text-wrap: balance; /* centered banner headline: even the two lines so neither widows (Gate B) */
text-align: center;
}
.framework-banner .fb-text strong { color: var(--accent-deep); }
.framework-banner .fb-label {
display: inline-block;
background: var(--accent);
color: var(--accent-ink);
font-family: var(--font-sans);
font-size: var(--fs-5);
font-weight: 700;
padding: calc(2 * var(--u)) calc(8 * var(--u));
border-radius: calc(4 * var(--u) * var(--rs));
text-transform: uppercase;
letter-spacing: 1px;
vertical-align: middle;
line-height: 1;
position: relative;
top: calc(-1 * var(--u));
}
/* Method-overview figure for the banner (banner-figure component,
catalogued in COMPONENTS.md). Usually CAPTIONLESS -- the banner text
block beside it is the figure's explanation. width:min-content collapses
the slot to the IMAGE (a captionless figure's slot IS the image); if a
short caption is used it wraps at the image box and can NEVER set the
flex-item width; margin-inline centres a block image (text-align does
not); overflow-wrap guards a long unbreakable token. Do NOT hand-roll a
.fb-fig wrapper or a bare <img class="w-100"> here. */
.framework-banner .banner-figure {
flex: 0 0 auto;
width: min-content;
margin: 0;
text-align: center;
}
.framework-banner .banner-figure img {
height: calc(120 * var(--u));
width: auto;
display: block;
margin-inline: auto;
}
.framework-banner .banner-figure figcaption {
width: 100%;
margin-top: calc(2 * var(--u));
font-family: var(--font-sans);
font-size: var(--fs-2);
line-height: 1.2;
color: var(--text-secondary);
text-align: center;
text-wrap: pretty;
overflow-wrap: anywhere;
}
.framework-banner .banner-figure figcaption strong { color: var(--accent-deep); }
/* =========================================================
BODY: 4 columns, variable-width hint via fr values
========================================================= */
.body-grid {
display: grid;
grid-template-columns: 1fr 1.05fr 1.05fr 1fr;
gap: calc(10 * var(--u));
/* No overflow:hidden here — it would clip card shadows (pitfall #6).
.poster already clips at the page boundary; min-height:0 is the
grid-blowout guard. */
min-height: 0;
}
.column {
display: flex; flex-direction: column;
gap: calc(11 * var(--u));
min-height: 0;
height: 100%;
padding-bottom: calc(4 * var(--u)); /* shadow breathing room above takeaways */
}
/* =========================================================
CARDS
========================================================= */
.card {
background: var(--bg-card);
border-radius: calc(5 * var(--u) * var(--rs));
padding: calc(4 * var(--u)) calc(9 * var(--u));
border: calc(1 * var(--u)) solid var(--border-soft);
box-shadow: var(--shadow-card);
position: relative;
}
.card.tinted { background: var(--bg-card-tint); }
.card.card--compact { padding: calc(3 * var(--u)) calc(6 * var(--u)); } /* predefined variant: tighter padding (fix (f)) */
.card.highlight {
border-left: calc(6 * var(--u)) solid var(--accent);
/* Flat emphasis tint (de-gradient). */
background: var(--bg-emphasis);
}
.section-title {
font-family: var(--font-sans);
font-weight: 700;
font-size: var(--fs-7);
color: var(--accent-deep);
margin-bottom: calc(3 * var(--u));
display: flex; align-items: center;
gap: calc(5 * var(--u));
}
/* Title text + any ★ marker share ONE .st-text span, so the heading wraps as
natural text (hanging indent) instead of flex-wrapping atomic items: the
number badge is never stranded on its own line and the ★ never widows. */
.section-title .st-text { flex: 1; min-width: 0; line-height: 1.18; }
/* Graceful fallback if a title is NOT wrapped in .st-text: float the badge so
bare inline text still wraps beside it rather than dropping below it. */
.section-title:not(:has(.st-text)) { display: block; line-height: 1.18; }
.section-title:not(:has(.st-text)) .num { float: left; margin-right: calc(5 * var(--u)); }
.section-title .num {
display: inline-flex; align-items: center; justify-content: center;
width: calc(22 * var(--u)); height: calc(22 * var(--u));
background: var(--accent); color: var(--accent-ink);
border-radius: 50%;
font-size: var(--fs-5); font-weight: 700;
flex-shrink: 0;
}
/* Inline "★ KEY" / "★ Headline" marker beside a section title. */
.section-title .key-mark { color: var(--emph); font-size: var(--fs-3); }
.body-text, .card p, .card li {
font-family: var(--font-serif);
font-size: var(--fs-4);
line-height: 1.3;
color: var(--text-primary);
}
.card ul, .card ol { padding-left: calc(18 * var(--u)); }
.card li { margin-bottom: calc(2 * var(--u)); }
.keyword { color: var(--accent); font-weight: 700; }
.keyword-emph { color: var(--emph); font-weight: 700; }
.highlight-text {
background: var(--bg-emphasis);
padding: 0 calc(3 * var(--u));
border-radius: calc(2 * var(--u) * var(--rs));
}
/* Equation block */
.eqn {
background: var(--bg-emphasis);
border-left: calc(3 * var(--u)) solid var(--accent);
padding: calc(4 * var(--u)) calc(10 * var(--u));
margin: calc(4 * var(--u)) 0;
font-size: var(--fs-5);
overflow-x: hidden;
}
/* Predefined variant for a larger / more legible display equation.
`calc(var(--fs-N) * k)` is allowed here because this is a
COMPONENTS.md-registered variant (DESIGN_FINAL §A / §10 f). */
.eqn--large { font-size: calc(var(--fs-5) * 1.25); }
.eqn .label {
display: block;
font-family: var(--font-sans);
font-size: var(--fs-2);
color: var(--accent);
font-weight: 600;
margin-bottom: calc(2 * var(--u));
text-transform: uppercase;
letter-spacing: 1px;
}
/* Callout: solid accent for primary, solid emphasis fill for theorems / "★ key" strips */
.callout {
background: var(--accent);
color: var(--accent-ink);
padding: calc(5 * var(--u)) calc(10 * var(--u));
border-radius: calc(4 * var(--u) * var(--rs));
font-size: var(--fs-4);
margin: calc(4 * var(--u)) 0;
}
.callout strong { color: var(--emph-soft); }
/* Flat solid emphasis fill (de-gradient). */
.callout.emph {
background: var(--emph);
color: var(--emph-ink); /* route through the token — accent-deep here would dodge the register swap AND fail AA on the default gold */
}
.callout.emph strong { color: var(--emph-ink); }
/* Figure container */
.figure { margin: calc(4 * var(--u)) 0; text-align: center; }
.figure img:not([class*="w-"]) { width: 100%; }
.figure--wide img { width: 100%; } /* predefined variant: force full card width over any .w-NN (fix (f)) */
.figure img {
border-radius: calc(4 * var(--u) * var(--rs));
border: calc(1 * var(--u)) solid var(--fig-frame);
background: var(--fig-bg);
}
.figure .caption {
font-family: var(--font-sans);
font-size: var(--fs-3);
color: var(--text-secondary);
margin-top: calc(3 * var(--u));
line-height: 1.3;
text-align: left;
}
.figure .caption strong { color: var(--accent-deep); }
/* Float a figure BESIDE text in a TEXT-RICH card: text wraps to its side
and then below it. The clearfix grows the card to contain the float;
mark the <img> data-fig-layout="beside-text" so the AR gates honour the
intentionally small width. Use ONLY when there is enough text to fill the
figure's height -- a text-sparse card should center .figure instead. A
short-text float leaves an L-shaped void below the text, which polish's
FIG/BESIDE-TEXT-VOID flags. */
.fig-wrap::after { content: ""; display: table; clear: both; }
.ff-fig {
float: right;
width: 48%; max-width: 58%; min-width: 38%;
margin: calc(1 * var(--u)) 0 calc(3 * var(--u)) calc(11 * var(--u));
text-align: center;
}
.ff-fig.left {
float: left;
margin: calc(1 * var(--u)) calc(11 * var(--u)) calc(3 * var(--u)) 0;
}
.ff-fig img {
display: block;
width: 100%;
border-radius: calc(4 * var(--u) * var(--rs));
border: calc(1 * var(--u)) solid var(--fig-frame);
background: var(--fig-bg);
}
.ff-fig .caption {
font-family: var(--font-sans);
font-size: var(--fs-3);
color: var(--text-secondary);
margin-top: calc(3 * var(--u));
line-height: 1.3;
text-align: center;
}
/* Result table with .ours row highlighted with the emphasis tint */
.result-table {
width: 100%;
border-collapse: collapse;
font-family: var(--font-sans);
font-size: var(--fs-3);
margin-top: calc(3 * var(--u));
}
.result-table th, .result-table td {
padding: calc(2 * var(--u)) calc(4 * var(--u));
text-align: center;
border-bottom: calc(1 * var(--u)) solid var(--border-soft);
}
.result-table thead th {
background: var(--accent); color: var(--accent-ink);
font-weight: 600; font-size: var(--fs-2);
}
.result-table tbody tr.group-row td {
background: var(--bg-emphasis); font-weight: 700;
text-align: left;
color: var(--accent-deep);
padding-left: calc(8 * var(--u));
border-bottom: calc(2 * var(--u)) solid var(--accent);
}
.result-table tbody tr.ours td { background: var(--emph-soft); font-weight: 700; }
.result-table tbody tr.ours td:first-child { color: var(--accent-deep); }
/* Greyed reference row (was inline color:#888). */
.result-table tbody tr.reference td { color: var(--text-muted); }
.result-table .method { text-align: left; padding-left: calc(8 * var(--u)); }
.result-table .best { color: var(--accent); font-weight: 700; }
/* 3-up stat box */
.keybox {
display: grid;
grid-template-columns: repeat(3, 1fr);
gap: calc(4 * var(--u));
margin: calc(4 * var(--u)) 0 0;
}
.keybox .kb-item {
background: var(--bg-emphasis);
border-top: calc(2 * var(--u)) solid var(--accent);
padding: calc(3 * var(--u));
text-align: center;
/* The grid stretches every tile to the tallest one; center the content
vertically so a 1-line tile's number aligns with a 2-line neighbour's
instead of top-ragged (SKILL Layout pitfall — stat-tile alignment). */
display: flex; flex-direction: column; justify-content: center;
border-radius: 0 0 calc(3 * var(--u) * var(--rs)) calc(3 * var(--u) * var(--rs));
}
.kb-item .kb-num {
font-family: var(--font-sans);
font-weight: 800;
font-size: var(--fs-6);
color: var(--accent);
line-height: 1;
}
.kb-item .kb-label {
font-family: var(--font-sans);
font-size: var(--fs-1);
color: var(--text-secondary);
margin-top: calc(2 * var(--u));
line-height: 1.1;
/* Reserve two label lines (2lh tracks this line-height) so 1-line
and 2-line labels occupy the same height and the big numbers
align across tiles. Keep labels <= 2 lines (COMPONENTS §keybox). */
min-height: 2lh;
}
/* Optional takeaways strip (delete the section if not needed) */
.takeaways-strip {
display: grid;
grid-template-columns: auto repeat(4, 1fr);
align-items: center;
gap: calc(10 * var(--u));
/* Flat solid emphasis fill (de-gradient). */
background: var(--bg-emphasis);
border: calc(1 * var(--u)) solid var(--border-soft);
border-radius: calc(5 * var(--u) * var(--rs));
padding: calc(8 * var(--u)) calc(12 * var(--u));
}
.takeaways-strip .ts-title {
font-family: var(--font-sans);
font-size: var(--fs-6);
font-weight: 800;
color: var(--accent-deep);
display: flex; align-items: center; gap: calc(6 * var(--u));
}
.takeaways-strip .ts-title .num {
display: inline-flex; align-items: center; justify-content: center;
width: calc(22 * var(--u)); height: calc(22 * var(--u));
background: var(--accent); color: var(--accent-ink);
border-radius: 50%;
font-size: var(--fs-5); font-weight: 700;
}
.takeaways-strip .ts-item {
border-left: calc(3 * var(--u)) solid var(--accent);
padding-left: calc(8 * var(--u));
line-height: 1.3;
text-align: center;
text-wrap: balance;
}
.takeaways-strip .ts-key {
font-family: var(--font-sans);
font-size: var(--fs-3);
font-weight: 700;
color: var(--accent);
text-transform: uppercase;
letter-spacing: 1px;
}
.takeaways-strip .ts-text {
font-family: var(--font-serif);
font-size: var(--fs-4);
margin-left: calc(4 * var(--u));
}
/* Footer */
.footer {
grid-column: 1 / -1;
display: flex; justify-content: space-between; align-items: baseline;
/* When the two blocks can't sit side by side they stack (wrap) instead of
overflowing; row-gap keeps the stacked blocks legible. */
flex-wrap: wrap; gap: calc(2 * var(--u)) calc(10 * var(--u));
padding-top: calc(8 * var(--u));
border-top: calc(1 * var(--u)) solid var(--border-soft);
font-family: var(--font-sans);
font-size: var(--fs-5);
color: var(--text-muted);
}
.footer .method-name { color: var(--accent-deep); }
/* Long repo URL / email breaks mid-token rather than overflowing the margin. */
.footer .repo { color: var(--accent); font-weight: 600; overflow-wrap: anywhere; }
/* Optional watermark */
.ornament {
position: absolute;
right: calc(20 * var(--u));
bottom: calc(20 * var(--u));
font-family: var(--font-sans);
font-size: var(--fs-9);
color: var(--ornament-ink);
font-weight: 900;
letter-spacing: 4px;
pointer-events: none;
user-select: none;
}
/* =========================================================
DENSITY COMPONENTS (catalogued — see COMPONENTS.md).
Guardrail: no component-local color semantics. Distinction is
carried by labels, order and typography — never by new hues.
========================================================= */
/* equation-stack: compact multi-row formula stack (denser than one big .eqn) */
.equation-stack { margin: calc(2 * var(--u)) 0; }
.equation-stack .eqn { margin: calc(2 * var(--u)) 0; padding: calc(2 * var(--u)) calc(8 * var(--u)); }
/* eqn-anatomy: term-by-term anatomy grid (2x2 default; --row variant = 1x4) */
.eqn-anatomy {
display: grid; grid-template-columns: 1fr 1fr;
gap: calc(2.5 * var(--u)); margin: calc(3 * var(--u)) 0;
}
.eqn-anatomy.eqn-anatomy--row { grid-template-columns: repeat(4, 1fr); }
.eqn-anatomy .ea-item {
background: var(--bg-card-tint);
border: 1px solid var(--border-soft);
border-left: calc(2 * var(--u)) solid var(--accent);
border-radius: calc(2 * var(--u) * var(--rs));
padding: calc(2 * var(--u)) calc(4 * var(--u));
font-size: var(--fs-2); line-height: 1.3;
}
.eqn-anatomy .ea-tag {
display: inline-block; font-family: var(--font-sans); font-weight: 700;
font-size: var(--fs-1); color: var(--accent-deep);
background: var(--accent-light); border-radius: calc(1.5 * var(--u) * var(--rs));
padding: calc(0.5 * var(--u)) calc(2.5 * var(--u));
margin-right: calc(1.5 * var(--u));
}
/* flow-strip: labeled pipeline, ALL steps the same accent; ONLY the
final step may carry the emphasis top bar. Not an "algorithm" unless the
paper has one — see COMPONENTS.md. */
.flow-strip { display: flex; align-items: stretch; gap: calc(1.5 * var(--u)); margin: calc(3 * var(--u)) 0; }
.flow-strip .step {
flex: 1; background: var(--accent-light);
border: 1px solid var(--accent-soft); border-radius: calc(2 * var(--u) * var(--rs));
padding: calc(2 * var(--u)) calc(2.5 * var(--u));
text-align: center; font-size: var(--fs-2); line-height: 1.25;
}
.flow-strip .step .step-name {
display: block; font-family: var(--font-sans); font-weight: 700;
font-size: var(--fs-1); color: var(--accent-deep);
letter-spacing: 0.5px; text-transform: uppercase;
margin-bottom: calc(1 * var(--u));
}
.flow-strip .step--final { border-top: calc(1.5 * var(--u)) solid var(--emph); background: var(--emph-soft); }
.flow-strip .arrow {
align-self: center; color: var(--accent);
font-family: var(--font-sans); font-weight: 700; font-size: var(--fs-4);
flex: 0 0 auto;
}
/* figure--duo: two paper figures sharing one caption; each img MUST be
a data-source="paper" asset and carry .w-45/.w-50 (42-48% each). */
.figure--duo { display: flex; gap: calc(3 * var(--u)); align-items: flex-start; justify-content: center; }
.figure--duo img { max-width: 48%; min-width: 0; } /* hard cap: duo contract is 42-48% each; an oversized w-NN cannot overflow the card */
/* result-table derived column (emph-soft = DERIVED arithmetic; label it) */
.result-table th.derived, .result-table td.derived {
background: var(--emph-soft); font-family: var(--font-sans); font-weight: 700;
}
/* Header cells keep the accent ink rule: on the emph-soft ground the white
head ink would read 1.07:1, so route the header's derived cell through
the emphasis ink token (Gate G -- never inherit ink across a ground change). */
.result-table thead th.derived { color: var(--emph-ink); }
/* keybox 4-up variant */
.keybox.keybox--4 { grid-template-columns: repeat(4, 1fr); }
/* algo: compact numbered procedure — ONLY when the paper itself states
an explicit algorithm/procedure; never invent steps. */
ol.algo { padding-left: calc(16 * var(--u)); }
ol.algo li { margin-bottom: calc(1.5 * var(--u)); font-size: var(--fs-3); line-height: 1.3; }
/* claim-pills: provenance mini-table (numeric-heavy posters only) */
.claim-pills { width: 100%; border-collapse: collapse; font-family: var(--font-sans); font-size: var(--fs-2); }
.claim-pills td {
border-bottom: 1px solid var(--border-soft);
padding: calc(1.5 * var(--u)) calc(3 * var(--u));
}
.claim-pills .cp-id {
font-weight: 700; color: var(--accent-deep); background: var(--accent-light);
border-radius: calc(1.5 * var(--u) * var(--rs)); padding: 0 calc(2.5 * var(--u) * var(--rs)); white-space: nowrap;
}
.claim-pills .cp-fact { color: var(--accent-deep); font-weight: 700; }
.claim-pills .cp-derived { color: var(--text-secondary); font-weight: 700; }
/* Grid-blowout guards: (1) columns may shrink below min-content so a
wide equation can never widen its track and squeeze siblings;
(2) over-wide display math scales down to fit its box instead of
blowing out the card (SVG keeps aspect via height:auto). */
.body-grid > .column { min-width: 0; }
.card { min-width: 0; }
.eqn mjx-container > svg { max-width: 100%; height: auto; }
/* ── Utility classes — defined LAST so equal-specificity ties
resolve in the utility's favor (source order). ── */
.fs-1 { font-size: var(--fs-1); }
.fs-2 { font-size: var(--fs-2); }
.fs-3 { font-size: var(--fs-3); }
.fs-4 { font-size: var(--fs-4); }
.fs-5 { font-size: var(--fs-5); }
.fs-6 { font-size: var(--fs-6); }
.fs-7 { font-size: var(--fs-7); }
.fs-8 { font-size: var(--fs-8); }
.fs-9 { font-size: var(--fs-9); }
.mt-1 { margin-top: calc(1 * var(--u)); }
.mt-2 { margin-top: calc(2 * var(--u)); }
.mt-3 { margin-top: calc(3 * var(--u)); }
.mt-4 { margin-top: calc(4 * var(--u)); }
.mt-5 { margin-top: calc(5 * var(--u)); }
.mt-6 { margin-top: calc(6 * var(--u)); }
.mb-1 { margin-bottom: calc(1 * var(--u)); }
.mb-2 { margin-bottom: calc(2 * var(--u)); }
.mb-3 { margin-bottom: calc(3 * var(--u)); }
.mb-4 { margin-bottom: calc(4 * var(--u)); }
.w-45 { width: 45%; }
.w-50 { width: 50%; }
.w-55 { width: 55%; }
.w-60 { width: 60%; }
.w-65 { width: 65%; }
.w-70 { width: 70%; }
.w-75 { width: 75%; }
.w-80 { width: 80%; }
.w-85 { width: 85%; }
.w-90 { width: 90%; }
.w-95 { width: 95%; }
.w-100 { width: 100%; }
.text-secondary { color: var(--text-secondary); }
.text-muted { color: var(--text-muted); }
.nowrap { white-space: nowrap; }
.text-center { text-align: center; }
/* =========================================================
PRINT OVERRIDE — KEEP LAST so it wins source-order ties.
========================================================= */
@media print {
html, body { background: white; }
.poster { margin: 0; box-shadow: none; width: 60in; height: 36in; page-break-after: avoid; }
:root { --u: 1mm; }
}
</style>
</head>
<body>
<div class="poster" data-measure-role="poster">
<!-- =========================== HEADER =========================== -->
<header class="header" data-measure-role="header">
<div class="venue-badge">
<div class="vb-venue">ICML</div>
<div class="vb-year">2026</div>
<div class="vb-tag">REPRO CHALLENGE</div>
</div>
<div class="title-block">
<h1 class="title">Reproducing: <span class="accent">Floating-Point Networks</span> with Automatic Differentiation Can Represent Almost All Floating-Point Functions <span class="accent">and Their Gradients</span></h1>
<div class="subtitle">Independent reproduction &mdash; ICML 2026 Reproducibility Challenge</div>
<div class="authors-line">
<span class="author">SabaPivot</span>
<span class="aff">Independent re-derivation of arXiv&nbsp;2605.01702 &middot; constructive existence paper &middot; author code exists (yechanp/fp-grad-rep, ReLU only)</span>
</div>
</div>
<div class="right-block">
<div class="qr-block">
<div class="qr-label">arXiv 2605.01702</div>
<div class="qr-label">OpenReview g89qqA6qmD</div>
<div class="qr-label">6 / 6 claims verified</div>
</div>
</div>
</header>
<!-- ===================== VERDICT BANNER ===================== -->
<section class="framework-banner" data-measure-role="banner">
<div class="fb-text">
<span class="fb-label">Verdict</span>
&nbsp;We <strong>build</strong> the networks, run real IEEE-754 automatic differentiation, and check both outputs against two independently chosen targets at every point of a finite domain. A floating-point network carries an <strong>arbitrary value and an arbitrary, unrelated AD gradient at once</strong> &mdash; bit for bit, all six activations, binary32 and binary64. In exact rational arithmetic the same networks fail. <strong>All six claims verified.</strong>
</div>
<div class="banner-stats">
<div class="bs-item"><div class="bs-num">6/6</div><div class="bs-label">claims<br>verified</div></div>
<div class="bs-item"><div class="bs-num">0.0</div><div class="bs-label">max abs error,<br>values &amp; grads</div></div>
<div class="bs-item"><div class="bs-num">1188</div><div class="bs-label">exact checks,<br>36 networks</div></div>
<div class="bs-item"><div class="bs-num">$0</div><div class="bs-label">CPU only,<br>~6 min</div></div>
</div>
</section>
<!-- ============================ BODY ============================ -->
<div class="body-grid" data-measure-role="body">
<!-- ============ COLUMN 1 ============ -->
<div class="column" data-measure-role="column">
<div class="card highlight" data-measure-role="card"
data-logbook-target="executive-summary" data-logbook-label="Executive summary">
<div class="section-title"><span class="num">1</span><span class="st-text">Setting &amp; Question</span></div>
<p class="body-text">
A network on a computer is not a differentiable function: it is a
<span class="keyword">sequence of rounded operations</span>, and its "gradient" is whatever AD computes by
walking that sequence backwards,
$$D_{f,x}(g) = g \otimes A_L \otimes \sigma'(y_{L-1}) \otimes \cdots \otimes \sigma'(y_1) \otimes A_1$$
evaluated <span class="keyword">left to right</span>, every $\otimes$, $\oplus$ correctly rounded. Exact
arithmetic ties value and gradient together by the chain rule; floating point unties them.
</p>
<div class="flow-strip">
<div class="step"><span class="step-name">Lem 3.4</span>$f=f^*$, $D=0$</div>
<div class="arrow">+</div>
<div class="step"><span class="step-name">Lem 3.5</span>$f=0$, $D=g^*$</div>
<div class="arrow">&rarr;</div>
<div class="step step--final"><span class="step-name">Thm 3.1</span>both</div>
</div>
</div>
<div class="card" data-measure-role="card">
<div class="section-title"><span class="num">2</span><span class="st-text">What We Actually Did</span></div>
<p class="body-text fs-3">
A constructive existence paper, so the reproduction is a construction: our <span class="keyword">own</span>
simulator and our <span class="keyword">own</span> construction, no author code reused.</p>
<ul class="fs-3 mt-2">
<li>Simulator vs <span class="keyword">PyTorch autograd</span>: $\le 6\times 10^{-7}$ rel. (binary32),
$\le 4\times 10^{-15}$ (binary64) &mdash; residue is PyTorch's summation order.</li>
<li>Domain: every other binary32 number of $[1,2)$ &mdash; <span class="keyword">adjacent floats</span>.</li>
<li>Every check is <span class="keyword">bit-exact equality</span>, never a tolerance.</li>
</ul>
</div>
<div class="card" data-measure-role="card"
data-logbook-target="conclusion" data-logbook-label="Conclusion">
<div class="section-title"><span class="num">3</span><span class="st-text">Scope, Cost &amp; Verdict</span></div>
<ul class="fs-3 mt-2">
<li><span class="keyword">All six claims verified.</span> No contradiction found anywhere.</li>
<li><span class="keyword">Scope:</span> $|X|$ parallel branches, so the full domain
$[-2^{125},2^{125}]_{\mathbb F}$ is unbuildable by anyone; we verify on finite grids.</li>
<li><span class="keyword">Our depth bound is smaller:</span> we reach the same guarantee at $L=5$, so
"$L \ge 9$" is sufficient but not tight (arXiv v2 itself says $L \ge 7$).</li>
</ul>
<div class="keybox mt-6">
<div class="kb-item"><div class="kb-num">CPU</div><div class="kb-label">no GPU,<br>no API key</div></div>
<div class="kb-item"><div class="kb-num">6 min</div><div class="kb-label">total<br>wall clock</div></div>
<div class="kb-item"><div class="kb-num">$0</div><div class="kb-label">total<br>cost</div></div>
</div>
<div class="callout emph mt-6 mb-4">
<strong>Bottom line.</strong> A network's values tell you nothing about the gradients AD will hand back:
the two are independently choosable, and no forward test reveals it.
</div>
</div>
</div>
<!-- ============ COLUMN 2 ============ -->
<div class="column" data-measure-role="column">
<div class="card highlight" data-measure-role="card"
data-logbook-target="claim-1-thm-3-1-values-independent-gradient-9-layers"
data-logbook-label="Claim 1: Thm 3.1 values + independent gradient">
<div class="section-title"><span class="num">4</span><span class="st-text">Claim 1 &mdash; Values and an Independent Gradient</span></div>
<p class="body-text fs-3">
Draw $f^*, h^*, g^*$ independently at random on a finite domain $X$ (with the theorem's only hypothesis,
$h^*(x)=0 \Rightarrow g^*(x)=0$) and build a 9-layer network. Then for every $x \in X$
</p>
<div class="eqn">
<span class="label">both targets, simultaneously, bit for bit</span>
$$f(x) = f^*(x) \qquad\text{and}\qquad D_{f,x}\big(h^*(x)\big) = g^*(x)$$
</div>
<div class="figure">
<img src="fig2_exactness.png" data-source="paper" data-asset-id="fig-exactness" class="w-45"
alt="Grid of zero mismatches for six activations in binary32 and binary64">
<div class="caption fs-2">
<strong>0 mismatches</strong> out of 33 points &times; 3 seeds for each of the 12 activation &times;
format cells. Max absolute error <strong>0.0</strong>.
</div>
</div>
<p class="body-text fs-3 mt-1">
Also verified in <span class="keyword">two input dimensions</span> (6&times;6 grid) and at $L=5,\dots,12$.
1188 exact value checks and 1188 exact gradient checks in total.
</p>
</div>
<div class="card" data-measure-role="card">
<div class="section-title"><span class="num">5</span><span class="st-text">The Layer Count Is Not Tight</span></div>
<table class="result-table">
<thead>
<tr><th class="method">depth $L$</th><th>value mismatches</th><th class="derived">gradient mismatches</th></tr>
</thead>
<tbody>
<tr><td class="method">1 (affine)</td><td colspan="2">impossible: $D_{f,x}(y)=y\otimes A_1$ is constant in $x$</td></tr>
<tr class="ours"><td class="method">5</td><td>0</td><td class="derived">0</td></tr>
<tr><td class="method">7 (arXiv v2 bound)</td><td>0</td><td class="derived">0</td></tr>
<tr><td class="method">9 (claim as stated)</td><td>0</td><td class="derived">0</td></tr>
</tbody>
</table>
<p class="body-text fs-3 mt-1">
"$L \ge 9$" is <span class="keyword">sufficient</span> and holds; not necessary &mdash; we reach the same
guarantee at $L=5$, and arXiv v2 lowers the authors' bound to $L \ge 7$.
</p>
</div>
<div class="card card--compact" data-measure-role="card"
data-logbook-target="claim-2-thm-3-1-holds-for-relu-elu-gelu-swish-sigmoid-tanh"
data-logbook-label="Claim 2: six activations">
<div class="section-title"><span class="num">6</span><span class="st-text">Claim 2 &mdash; All Six Activations</span></div>
<p class="body-text fs-3">
Which mechanism applies is decided by the activation, as the paper's Conditions 1&ndash;3 predict:
</p>
<ul class="fs-3">
<li><span class="keyword">ReLU, ELU, GELU, Swish</span> never have $\sigma'=0$ where $\sigma \ne 0$: the
gradient must be annihilated by a <span class="keyword">$+C/-C$ pair</span> in the AD sum.</li>
<li><span class="keyword">Sigmoid, tanh</span> saturate: $\sigma = \pm 1$ exactly while $\sigma'$ has
<span class="keyword">underflowed to 0</span>. No pair needed.</li>
</ul>
<p class="body-text fs-3">
Operating points are searched numerically per activation and format.
</p>
</div>
</div>
<!-- ============ COLUMN 3 ============ -->
<div class="column" data-measure-role="column">
<div class="card highlight" data-measure-role="card"
data-logbook-target="claim-6-mechanism-fp-non-associativity-decouples-ad-gradient"
data-logbook-label="Claim 6: non-associativity is the mechanism">
<div class="section-title"><span class="num">7</span><span class="st-text">Claim 6 &mdash; Non-Associativity Is the Mechanism</span></div>
<p class="body-text fs-3">
Both halves are one identity used in two accumulation orders:
</p>
<div class="eqn">
<span class="label">the wipe identity</span>
$$(t \oplus C) \ominus C = 0 \;\;\text{when}\;\; |t| &lt; \tfrac{1}{2}\mathrm{ulp}(C), \qquad\text{but}\qquad 0 \oplus C \ominus C \oplus u = u$$
</div>
<div class="figure">
<img src="fig1_mechanism.png" data-source="paper" data-asset-id="fig-mechanism" class="w-45"
alt="Two accumulation chains showing the same constant pair annihilating one and not the other">
<div class="caption fs-2">
The two units are <strong>identical</strong>, downstream weights $+w,-w$: the forward pass cancels them
exactly. The AD accumulation meets them <em>after</em> the value path and is destroyed.
</div>
</div>
</div>
<div class="card" data-measure-role="card">
<div class="section-title"><span class="num">8</span><span class="st-text">The Decisive Test: Exact Arithmetic</span></div>
<p class="body-text fs-3">
The <em>same</em> ReLU networks in Python <span class="keyword">Fractions</span>: identical weights and
graph, no rounding.
</p>
<table class="result-table">
<thead>
<tr><th class="method">network</th><th>IEEE-754</th><th class="derived">exact rationals</th></tr>
</thead>
<tbody>
<tr><td class="method">Lem 3.4: $\max|D|$</td><td class="best">0</td><td class="derived">$3.6\times 10^{9}$</td></tr>
<tr class="ours"><td class="method">Lem 3.5: $\max|f|$</td><td>0</td><td class="derived">$2.9\times 10^{17}$</td></tr>
</tbody>
</table>
<p class="body-text fs-3 mt-2">
Remove the rounding and <span class="keyword">both constructions collapse</span> &mdash; the effect is the
rounding, not the simulator.
</p>
<div class="section-title mt-6 mb-2"><span class="num">9</span><span class="st-text">Two Ablations</span></div>
<ul class="fs-3 mt-6">
<li><span class="keyword">Delete the pair:</span> values still bit-exact, AD gradient jumps
$0 \to 7.7\times 10^{9}$.</li>
<li><span class="keyword">Permute the layer-1 unit order:</span> same weights, same graph, only the
summation order &mdash; identical function in exact arithmetic. Gradient again jumps to
$7.7\times 10^{9}$, $f$ stays bit-exact.</li>
</ul>
<div class="callout mt-6 mb-4 fs-3">
The permutation is the cleanest statement of Section 1.1: the AD gradient is a property of the
<strong>order of operations</strong>, not of the function. What would have falsified the claim &mdash; a
gradient surviving the pair, a value the pair disturbed, exact arithmetic agreeing with floating point
&mdash; we looked for and did not find.
</div>
</div>
</div>
<!-- ============ COLUMN 4 ============ -->
<div class="column" data-measure-role="column">
<div class="card" data-measure-role="card"
data-logbook-target="claim-4-lemma-3-4-correct-values-gradient-suppressed-to-zero"
data-logbook-label="Claim 4: Lemma 3.4">
<div class="section-title"><span class="num">10</span><span class="st-text">Claim 4 &mdash; Values Kept, Gradient Killed</span></div>
<p class="body-text fs-3">
$f = f^*$ bit-exact while $D_{f,x}(y) = 0$ for <em>every</em> $y$: 46 magnitudes over binary32, 276 over
binary64.</p>
<div class="figure">
<img src="fig3_lemma34_range.png" data-source="paper" data-asset-id="fig-lem34" class="w-45"
alt="Bar chart of the largest input gradient magnitude for which suppression is exact, per activation">
<div class="caption fs-2">
Exact to $|y| \approx 2^{61}$ (ReLU/ELU), $2^{117}$ (Sigmoid), then <strong>degrades as predicted</strong>
&mdash; the annihilating constant overflows. Values stay exact.
</div>
</div>
</div>
<div class="card" data-measure-role="card"
data-logbook-target="claim-5-lemma-3-5-zero-output-arbitrary-ad-gradient"
data-logbook-label="Claim 5: Lemma 3.5">
<div class="section-title"><span class="num">11</span><span class="st-text">Claim 5 &mdash; Zero Output, Arbitrary Gradient</span></div>
<p class="body-text fs-3">
The dual gadget: $f \equiv 0$ exactly, $D = g^*$ arbitrary.</p>
<table class="result-table">
<thead>
<tr><th class="method">check</th><th>result</th></tr>
</thead>
<tbody>
<tr><td class="method">$f\equiv 0$ on $X$, 12 nets</td><td class="best">204 / 204</td></tr>
<tr><td class="method">$f(x)=0$ off-grid</td><td class="best">768 / 768</td></tr>
<tr class="ours"><td class="method">$D = g^*$ bit-exact</td><td>204 / 204</td></tr>
</tbody>
</table>
<p class="body-text fs-3 mt-1">A network that looks inert in any forward test hands an attacker-chosen
gradient to the optimiser.</p>
</div>
<div class="card highlight" data-measure-role="card"
data-logbook-target="claim-3-thm-3-2-gradient-varies-with-phi-prime-under-antisymmetry"
data-logbook-label="Claim 3: Thm 3.2 antisymmetry">
<div class="section-title"><span class="num">12</span><span class="st-text">Claim 3 &mdash; Gradient That Tracks the Loss Derivative</span></div>
<p class="body-text fs-3">
Antisymmetry is <span class="keyword">forced</span>, not assumed: IEEE-754 $\otimes,\oplus$ are
sign-symmetric, so $D_{f,x}(-y)=-D_{f,x}(y)$ always &mdash; exhaustive over all $4.03\times 10^{9}$
binary16 pairs, <strong>0 violations</strong>.
</p>
<div class="figure">
<img src="fig4_thm32.png" data-source="paper" data-asset-id="fig-thm32" class="w-45"
alt="Log-log plot of AD gradient magnitude against the input gradient, against the proportional line">
<div class="caption fs-2">
One network, $f \equiv 0$, gradient prescribed on 24 bands of $y$: <strong>48/48 exact</strong>. $|D/y|$
moves <strong>8.5 orders of magnitude</strong>; exact arithmetic gives a constant ratio. Scope: 24 bands,
not all $2^{32}$ floats &mdash; the paper's full-range construction needs 569 layers.
</div>
</div>
</div>
</div>
</div>
<!-- ============================ FOOTER ============================ -->
<div class="footer" data-measure-role="footer">
<div>
<strong class="method-name">Floating-point networks with AD represent almost all functions and their gradients</strong> &middot;
Independent reproduction, ICML 2026 Reproducibility Challenge &middot; all six claims verified on CPU, $0.
</div>
<div>
Paper: <span class="repo">arXiv:2605.01702</span> &nbsp;&middot;&nbsp;
OpenReview: <span class="repo">g89qqA6qmD</span> &nbsp;&middot;&nbsp;
Reproduction: <span class="repo">SabaPivot</span>
</div>
</div>
</div>
</body>
</html>

Xet Storage Details

Size:
58.7 kB
·
Xet hash:
948bfdaee9a2005df2189160c1e5f8f9073f1ae3b5aa32d8c3bd66c5efb9556d

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.