Download index.html from NASK-PIB/RefusEU: direct link, hf CLI and curl.
- Browser
- Download file 29.6 kB
-
https://huggingface.co/spaces/NASK-PIB/RefusEU/resolve/main/index.html
- Command line
-
hf download hf://spaces/NASK-PIB/RefusEU/index.html
-
curl -L -o index.html https://huggingface.co/spaces/NASK-PIB/RefusEU/resolve/main/index.html
29.6 kB
| <html lang="en"> | |
| <head> | |
| <meta charset="UTF-8" /> | |
| <meta name="viewport" content="width=device-width, initial-scale=1.0"/> | |
| <title>Multilingual Refusal Alignment for Safer LLMs</title> | |
| <link rel="preconnect" href="https://fonts.googleapis.com"> | |
| <link href="https://fonts.googleapis.com/css2?family=Source+Serif+4:ital,opsz,wght@0,8..60,300;0,8..60,400;0,8..60,600;0,8..60,700;1,8..60,300;1,8..60,400&family=DM+Sans:ital,wght@0,300;0,400;0,500;0,600;1,300&family=JetBrains+Mono:wght@400;500&display=swap" rel="stylesheet"> | |
| <style> | |
| :root { | |
| --bg: #faf9f7; | |
| --surface: #ffffff; | |
| --border: #e8e4de; | |
| --text-primary: #1a1714; | |
| --text-secondary: #5c5650; | |
| --text-muted: #8c867e; | |
| --accent: #c0392b; | |
| --accent-light: #fdf0ef; | |
| --accent2: #2c5f8a; | |
| --accent2-light: #edf3f9; | |
| --tag-bg: #f0ede8; | |
| --shadow: 0 2px 16px rgba(26,23,20,0.07); | |
| --shadow-lg: 0 8px 40px rgba(26,23,20,0.11); | |
| --radius: 8px; | |
| --serif: 'Source Serif 4', Georgia, serif; | |
| --sans: 'DM Sans', system-ui, sans-serif; | |
| --mono: 'JetBrains Mono', monospace; | |
| } | |
| *, *::before, *::after { box-sizing: border-box; margin: 0; padding: 0; } | |
| html { scroll-behavior: smooth; } | |
| body { | |
| font-family: var(--sans); | |
| background: var(--bg); | |
| color: var(--text-primary); | |
| line-height: 1.7; | |
| -webkit-font-smoothing: antialiased; | |
| } | |
| /* ββ HERO βββββββββββββββββββββββββββββββββββββββββββ */ | |
| .hero { | |
| background: var(--surface); | |
| border-bottom: 1px solid var(--border); | |
| padding: 80px 24px 64px; | |
| text-align: center; | |
| } | |
| .venue-badge { | |
| display: inline-flex; | |
| align-items: center; | |
| gap: 6px; | |
| background: var(--accent-light); | |
| color: var(--accent); | |
| border: 1px solid rgba(192,57,43,0.2); | |
| border-radius: 20px; | |
| padding: 5px 14px; | |
| font-size: 0.78rem; | |
| font-weight: 600; | |
| letter-spacing: 0.04em; | |
| text-transform: uppercase; | |
| margin-bottom: 28px; | |
| } | |
| .hero h1 { | |
| font-family: var(--serif); | |
| font-size: clamp(1.7rem, 4vw, 2.6rem); | |
| font-weight: 600; | |
| line-height: 1.25; | |
| max-width: 780px; | |
| margin: 0 auto 8px; | |
| color: var(--text-primary); | |
| letter-spacing: -0.01em; | |
| } | |
| .hero h1 em { | |
| font-style: italic; | |
| color: var(--accent); | |
| } | |
| .authors { | |
| margin: 28px auto 0; | |
| max-width: 660px; | |
| font-size: 0.95rem; | |
| color: var(--text-secondary); | |
| } | |
| .authors .name-list { | |
| display: flex; | |
| flex-wrap: wrap; | |
| justify-content: center; | |
| gap: 4px 18px; | |
| font-weight: 500; | |
| color: var(--text-primary); | |
| margin-bottom: 6px; | |
| } | |
| .authors .name-list sup { | |
| font-size: 0.65em; | |
| color: var(--accent); | |
| font-weight: 600; | |
| } | |
| .affil { | |
| font-size: 0.82rem; | |
| color: var(--text-muted); | |
| line-height: 1.6; | |
| } | |
| /* ββ LINKS ROW βββββββββββββββββββββββββββββββββββββ */ | |
| .links-row { | |
| display: flex; | |
| flex-wrap: wrap; | |
| justify-content: center; | |
| gap: 10px; | |
| margin-top: 32px; | |
| } | |
| .btn { | |
| display: inline-flex; | |
| align-items: center; | |
| gap: 7px; | |
| padding: 9px 18px; | |
| border-radius: 6px; | |
| font-family: var(--sans); | |
| font-size: 0.84rem; | |
| font-weight: 500; | |
| text-decoration: none; | |
| transition: all 0.18s ease; | |
| border: 1px solid transparent; | |
| cursor: pointer; | |
| } | |
| .btn-primary { | |
| background: var(--text-primary); | |
| color: #fff; | |
| } | |
| .btn-primary:hover { background: #333; transform: translateY(-1px); box-shadow: var(--shadow); } | |
| .btn-outline { | |
| background: var(--surface); | |
| color: var(--text-primary); | |
| border-color: var(--border); | |
| } | |
| .btn-outline:hover { border-color: var(--text-primary); transform: translateY(-1px); box-shadow: var(--shadow); } | |
| .btn-dataset { | |
| background: var(--accent2-light); | |
| color: var(--accent2); | |
| border-color: rgba(44,95,138,0.2); | |
| } | |
| .btn-dataset:hover { background: #dbe9f5; transform: translateY(-1px); } | |
| .btn svg { width: 15px; height: 15px; flex-shrink: 0; } | |
| /* ββ LAYOUT ββββββββββββββββββββββββββββββββββββββββ */ | |
| .container { | |
| max-width: 860px; | |
| margin: 0 auto; | |
| padding: 0 24px; | |
| } | |
| section { | |
| padding: 64px 24px; | |
| max-width: 860px; | |
| margin: 0 auto; | |
| } | |
| .section-label { | |
| font-size: 0.72rem; | |
| font-weight: 600; | |
| letter-spacing: 0.1em; | |
| text-transform: uppercase; | |
| color: var(--accent); | |
| margin-bottom: 12px; | |
| } | |
| h2 { | |
| font-family: var(--serif); | |
| font-size: clamp(1.25rem, 2.5vw, 1.6rem); | |
| font-weight: 600; | |
| margin-bottom: 20px; | |
| color: var(--text-primary); | |
| letter-spacing: -0.01em; | |
| } | |
| p { | |
| color: var(--text-secondary); | |
| font-size: 0.97rem; | |
| line-height: 1.8; | |
| margin-bottom: 14px; | |
| } | |
| p:last-child { margin-bottom: 0; } | |
| /* ββ ABSTRACT ββββββββββββββββββββββββββββββββββββββ */ | |
| .abstract-section { | |
| border-bottom: 1px solid var(--border); | |
| } | |
| .abstract-card { | |
| background: var(--surface); | |
| border: 1px solid var(--border); | |
| border-radius: 12px; | |
| padding: 32px 36px; | |
| box-shadow: var(--shadow); | |
| } | |
| /* ββ HIGHLIGHTS ββββββββββββββββββββββββββββββββββββ */ | |
| .highlights-grid { | |
| display: grid; | |
| grid-template-columns: repeat(auto-fit, minmax(240px, 1fr)); | |
| gap: 16px; | |
| margin-top: 8px; | |
| } | |
| .highlight-card { | |
| background: var(--surface); | |
| border: 1px solid var(--border); | |
| border-radius: 10px; | |
| padding: 22px 24px; | |
| box-shadow: var(--shadow); | |
| transition: box-shadow 0.2s, transform 0.2s; | |
| position: relative; | |
| overflow: hidden; | |
| } | |
| .highlight-card::before { | |
| content: ''; | |
| position: absolute; | |
| top: 0; left: 0; right: 0; | |
| height: 3px; | |
| } | |
| .highlight-card:nth-child(1)::before { background: var(--accent); } | |
| .highlight-card:nth-child(2)::before { background: var(--accent2); } | |
| .highlight-card:nth-child(3)::before { background: #2a7a4e; } | |
| .highlight-card:nth-child(4)::before { background: #7a4a2a; } | |
| .highlight-card:hover { box-shadow: var(--shadow-lg); transform: translateY(-2px); } | |
| .highlight-icon { | |
| font-size: 1.5rem; | |
| margin-bottom: 10px; | |
| display: block; | |
| } | |
| .highlight-title { | |
| font-weight: 600; | |
| font-size: 0.9rem; | |
| color: var(--text-primary); | |
| margin-bottom: 6px; | |
| } | |
| .highlight-card p { | |
| font-size: 0.85rem; | |
| line-height: 1.65; | |
| margin: 0; | |
| } | |
| /* ββ DATASET βββββββββββββββββββββββββββββββββββββββ */ | |
| .dataset-section { | |
| background: var(--surface); | |
| border-top: 1px solid var(--border); | |
| border-bottom: 1px solid var(--border); | |
| padding: 64px 24px; | |
| } | |
| .dataset-inner { | |
| max-width: 860px; | |
| margin: 0 auto; | |
| } | |
| .dataset-stats { | |
| display: grid; | |
| grid-template-columns: repeat(auto-fit, minmax(150px, 1fr)); | |
| gap: 16px; | |
| margin: 28px 0; | |
| } | |
| .stat-box { | |
| background: var(--bg); | |
| border: 1px solid var(--border); | |
| border-radius: 10px; | |
| padding: 20px; | |
| text-align: center; | |
| } | |
| .stat-number { | |
| font-family: var(--serif); | |
| font-size: 2rem; | |
| font-weight: 600; | |
| color: var(--accent); | |
| display: block; | |
| line-height: 1.1; | |
| } | |
| .stat-label { | |
| font-size: 0.78rem; | |
| color: var(--text-muted); | |
| margin-top: 4px; | |
| display: block; | |
| font-weight: 500; | |
| } | |
| .lang-pills { | |
| display: flex; | |
| flex-wrap: wrap; | |
| gap: 7px; | |
| margin-top: 16px; | |
| } | |
| .lang-pill { | |
| background: var(--tag-bg); | |
| color: var(--text-secondary); | |
| border-radius: 5px; | |
| padding: 3px 10px; | |
| font-size: 0.78rem; | |
| font-weight: 500; | |
| font-family: var(--mono); | |
| } | |
| /* ββ RESEARCH QUESTIONS ββββββββββββββββββββββββββββ */ | |
| .rq-list { | |
| list-style: none; | |
| display: flex; | |
| flex-direction: column; | |
| gap: 14px; | |
| margin-top: 8px; | |
| } | |
| .rq-item { | |
| display: flex; | |
| gap: 16px; | |
| background: var(--surface); | |
| border: 1px solid var(--border); | |
| border-radius: 10px; | |
| padding: 18px 20px; | |
| box-shadow: var(--shadow); | |
| } | |
| .rq-tag { | |
| flex-shrink: 0; | |
| background: var(--text-primary); | |
| color: #fff; | |
| border-radius: 6px; | |
| padding: 2px 9px; | |
| font-size: 0.72rem; | |
| font-weight: 700; | |
| letter-spacing: 0.04em; | |
| height: fit-content; | |
| margin-top: 2px; | |
| } | |
| .rq-text { | |
| font-size: 0.92rem; | |
| color: var(--text-secondary); | |
| line-height: 1.7; | |
| margin: 0; | |
| } | |
| /* ββ IMAGE PLACEHOLDER βββββββββββββββββββββββββββββ */ | |
| .figure-block { | |
| background: var(--surface); | |
| border: 1.5px dashed var(--border); | |
| border-radius: 10px; | |
| padding: 0; | |
| overflow: hidden; | |
| margin: 24px 0; | |
| } | |
| .figure-block img { | |
| width: 100%; | |
| height: auto; | |
| display: block; | |
| } | |
| .figure-caption { | |
| padding: 10px 16px 12px; | |
| font-size: 0.8rem; | |
| color: var(--text-muted); | |
| border-top: 1px solid var(--border); | |
| background: var(--bg); | |
| font-style: italic; | |
| } | |
| /* ββ FINDINGS ββββββββββββββββββββββββββββββββββββββ */ | |
| .finding-list { | |
| list-style: none; | |
| display: flex; | |
| flex-direction: column; | |
| gap: 12px; | |
| margin-top: 8px; | |
| } | |
| .finding-item { | |
| display: flex; | |
| gap: 14px; | |
| align-items: flex-start; | |
| padding: 16px 20px; | |
| background: var(--surface); | |
| border: 1px solid var(--border); | |
| border-radius: 9px; | |
| box-shadow: var(--shadow); | |
| } | |
| .finding-dot { | |
| width: 8px; | |
| height: 8px; | |
| border-radius: 50%; | |
| background: var(--accent); | |
| flex-shrink: 0; | |
| margin-top: 8px; | |
| } | |
| .finding-item:nth-child(2) .finding-dot { background: var(--accent2); } | |
| .finding-item:nth-child(3) .finding-dot { background: #2a7a4e; } | |
| .finding-item:nth-child(4) .finding-dot { background: #7a5f2a; } | |
| .finding-item:nth-child(5) .finding-dot { background: #6a2a7a; } | |
| .finding-text { | |
| font-size: 0.9rem; | |
| color: var(--text-secondary); | |
| line-height: 1.7; | |
| margin: 0; | |
| } | |
| .finding-text strong { color: var(--text-primary); } | |
| /* ββ BIBTEX ββββββββββββββββββββββββββββββββββββββββ */ | |
| .bibtex-section { | |
| background: var(--surface); | |
| border-top: 1px solid var(--border); | |
| padding: 64px 24px; | |
| } | |
| .bibtex-inner { | |
| max-width: 860px; | |
| margin: 0 auto; | |
| } | |
| .code-block { | |
| background: #1d1b19; | |
| color: #d4cfc9; | |
| border-radius: 10px; | |
| padding: 24px 28px; | |
| font-family: var(--mono); | |
| font-size: 0.78rem; | |
| line-height: 1.8; | |
| overflow-x: auto; | |
| margin-top: 16px; | |
| position: relative; | |
| } | |
| .code-key { color: #e09a6a; } | |
| .code-val { color: #8bbf7a; } | |
| .code-brace { color: #888; } | |
| .copy-btn { | |
| position: absolute; | |
| top: 12px; right: 14px; | |
| background: rgba(255,255,255,0.08); | |
| border: 1px solid rgba(255,255,255,0.12); | |
| color: #aaa; | |
| padding: 4px 11px; | |
| border-radius: 5px; | |
| font-size: 0.72rem; | |
| cursor: pointer; | |
| font-family: var(--sans); | |
| transition: all 0.15s; | |
| } | |
| .copy-btn:hover { background: rgba(255,255,255,0.14); color: #fff; } | |
| /* ββ FOOTER ββββββββββββββββββββββββββββββββββββββββ */ | |
| footer { | |
| background: var(--bg); | |
| border-top: 1px solid var(--border); | |
| text-align: center; | |
| padding: 28px 24px; | |
| font-size: 0.78rem; | |
| color: var(--text-muted); | |
| } | |
| footer a { color: var(--accent2); text-decoration: none; } | |
| footer a:hover { text-decoration: underline; } | |
| /* ββ DIVIDER βββββββββββββββββββββββββββββββββββββββ */ | |
| .divider { | |
| height: 1px; | |
| background: var(--border); | |
| max-width: 860px; | |
| margin: 0 auto; | |
| } | |
| /* ββ RESPONSIVE ββββββββββββββββββββββββββββββββββββ */ | |
| @media (max-width: 600px) { | |
| .abstract-card { padding: 22px 20px; } | |
| .links-row { gap: 8px; } | |
| section { padding: 48px 20px; } | |
| .dataset-section { padding: 48px 20px; } | |
| .bibtex-section { padding: 48px 20px; } | |
| } | |
| /* ββ ANIMATIONS ββββββββββββββββββββββββββββββββββββ */ | |
| @keyframes fadeUp { | |
| from { opacity: 0; transform: translateY(18px); } | |
| to { opacity: 1; transform: translateY(0); } | |
| } | |
| .hero > * { | |
| animation: fadeUp 0.55s ease both; | |
| } | |
| .hero .venue-badge { animation-delay: 0.05s; } | |
| .hero h1 { animation-delay: 0.12s; } | |
| .hero .authors { animation-delay: 0.2s; } | |
| .hero .links-row { animation-delay: 0.28s; } | |
| </style> | |
| </head> | |
| <body> | |
| <!-- βββ HERO ββββββββββββββββββββββββββββββββββββββββββ --> | |
| <header class="hero"> | |
| <div class="venue-badge">Findings of ACL 2026</div> | |
| <h1>Multilingual Refusal Alignment<br>for Safer Large Language Models</h1> | |
| <div class="authors"> | |
| <div class="name-list"> | |
| <span>Aleksandra KrasnodΔbska<sup>β </sup></span> | |
| <span>Wojciech Kusa<sup>β </sup></span> | |
| <span>Aldo Lipani<sup>β‘</sup></span> | |
| </div> | |
| <div class="affil"> | |
| <sup>β </sup> NASK National Research Institute, Warsaw, Poland Β· | |
| <sup>β‘</sup> University College London, London, UK<br> | |
| </div> | |
| </div> | |
| <div class="links-row"> | |
| <a href="https://wojciechkusa.github.io/papers/RefusEU-2026.pdf" class="btn btn-primary"> | |
| <svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><path d="M14 2H6a2 2 0 0 0-2 2v16a2 2 0 0 0 2 2h12a2 2 0 0 0 2-2V8z"/><polyline points="14 2 14 8 20 8"/></svg> | |
| Paper | |
| </a> | |
| <a href="https://huggingface.co/datasets/NASK-PIB/RefusEU" class="btn btn-dataset" target="_blank"> | |
| <svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><ellipse cx="12" cy="5" rx="9" ry="3"/><path d="M21 12c0 1.66-4 3-9 3s-9-1.34-9-3"/><path d="M3 5v14c0 1.66 4 3 9 3s9-1.34 9-3V5"/></svg> | |
| RefusEU Dataset | |
| </a> | |
| <a href="#bibtex" class="btn btn-outline"> | |
| <svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><path d="M4 7V4h16v3"/><path d="M9 20h6"/><path d="M12 4v16"/></svg> | |
| BibTeX | |
| </a> | |
| </div> | |
| </header> | |
| <!-- βββ ABSTRACT ββββββββββββββββββββββββββββββββββββββ --> | |
| <section class="abstract-section"> | |
| <div class="section-label">Abstract</div> | |
| <div class="abstract-card"> | |
| <p>As Large Language Models are deployed globally, ensuring consistent safety across languages becomes paramount. Yet safety behaviors vary unpredictably between languages, and most alignment research remains English-centric β creating a critical gap for non-English speakers worldwide.</p> | |
| <p>We systematically investigate the dynamics of multilingual alignment: whether single-language alignment transfers cross-lingually, how language consistency is preserved during training, and the resulting trade-offs with general capabilities. We introduce <strong>RefusEU</strong>, a novel refusal alignment dataset covering <strong>12 European languages</strong>, including a held-out test set for evaluating state-of-the-art models.</p> | |
| <p>Our controlled Direct Preference Optimization (DPO) experiments reveal two key insights: aligning models exclusively in English is <em>insufficient</em> to ensure cross-lingual safety β even for identical harm categories. Conversely, training on multilingual datasets can improve safety <em>without degrading general performance</em>, as measured by the Global MMLU benchmark.</p> | |
| </div> | |
| </section> | |
| <div class="divider"></div> | |
| <!-- βββ RESEARCH QUESTIONS ββββββββββββββββββββββββββββ --> | |
| <section> | |
| <div class="section-label">Research Questions</div> | |
| <h2>What We Set Out to Answer</h2> | |
| <ul class="rq-list"> | |
| <li class="rq-item"> | |
| <span class="rq-tag">RQ1</span> | |
| <p class="rq-text">Do we need to perform multilingual alignment for each language on the same groups of prompts, or is training in a single language (English) sufficient to achieve cross-lingual safety?</p> | |
| </li> | |
| <li class="rq-item"> | |
| <span class="rq-tag">RQ2</span> | |
| <p class="rq-text">How well is cross-lingual consistency β the ability to respond in the prompt's language β preserved during multilingual training, and how does it interact with safety?</p> | |
| </li> | |
| <li class="rq-item"> | |
| <span class="rq-tag">RQ3</span> | |
| <p class="rq-text">How does multilingual safety alignment influence general multilingual capabilities, including factual knowledge, fluency, and linguistic correctness?</p> | |
| </li> | |
| </ul> | |
| </section> | |
| <div class="divider"></div> | |
| <!-- βββ DATASET βββββββββββββββββββββββββββββββββββββββ --> | |
| <div class="dataset-section"> | |
| <div class="dataset-inner"> | |
| <div class="section-label">Dataset</div> | |
| <h2>Introducing RefusEU</h2> | |
| <p>RefusEU is the first European dataset designed for alignment training as DPO-ready triples β (question, chosen, rejected) β and includes a separate, contamination-free test split. Each chosen response is a high-quality refusal; each rejected response was generated by a safety-abliterated model.</p> | |
| <div class="dataset-stats"> | |
| <div class="stat-box"> | |
| <span class="stat-number">12</span> | |
| <span class="stat-label">European languages</span> | |
| </div> | |
| <div class="stat-box"> | |
| <span class="stat-number">4k+</span> | |
| <span class="stat-label">pairs per language</span> | |
| </div> | |
| <div class="stat-box"> | |
| <span class="stat-number">14</span> | |
| <span class="stat-label">harm categories</span> | |
| </div> | |
| <div class="stat-box"> | |
| <span class="stat-number">1,400</span> | |
| <span class="stat-label">test samples / lang.</span> | |
| </div> | |
| </div> | |
| <p style="margin-bottom:10px;"><strong style="color: var(--text-primary);">Languages covered:</strong></p> | |
| <div class="lang-pills"> | |
| <span class="lang-pill">English</span> | |
| <span class="lang-pill">German</span> | |
| <span class="lang-pill">French</span> | |
| <span class="lang-pill">Italian</span> | |
| <span class="lang-pill">Spanish</span> | |
| <span class="lang-pill">Portuguese</span> | |
| <span class="lang-pill">Polish</span> | |
| <span class="lang-pill">Czech</span> | |
| <span class="lang-pill">Slovak</span> | |
| <span class="lang-pill">Slovenian</span> | |
| <span class="lang-pill">Lithuanian</span> | |
| <span class="lang-pill">Latvian</span> | |
| </div> | |
| <p style="margin-top: 20px;">Questions are generated using an adversarial pipeline based on Rainbow Teaming across 10 attack styles and 14 crime categories (Llama-Guard taxonomy). A multi-model labelling protocol (Llama-Guard-3-8B, PolyGuard-Qwen, GPT-4o-mini) ensures label quality, with a manual audit confirming 100% accuracy across 1,200 sampled pairs.</p> | |
| <div class="figure-block" id="fig-dataset"> | |
| <img src="img/dataset.png" alt="Dataset construction pipeline" class="figure-image" style="width: 50%; margin: 0 auto;" /> | |
| <div class="figure-caption">Figure 1 β Dataset construction process: adversarial prompt generation β multilingual translation β dual-model safety labelling β DPO triple curation.</div> | |
| </div> | |
| </div> | |
| </div> | |
| <div class="divider"></div> | |
| <!-- βββ METHODOLOGY βββββββββββββββββββββββββββββββββββ --> | |
| <section> | |
| <div class="section-label">Methodology</div> | |
| <h2>Experimental Design</h2> | |
| <p>To isolate alignment dynamics, we start from <em>abliterated</em> Llama-3.1-8B and 70B models β versions where safety mechanisms have been deliberately removed via refusal direction ablation β then realign them using DPO under four dataset configurations:</p> | |
| <div class="highlights-grid" style="margin: 20px 0 24px;"> | |
| <div class="highlight-card"> | |
| <span class="highlight-icon">βοΈ</span> | |
| <div class="highlight-title">Balanced</div> | |
| <p>All 12 languages with equal representation (34,668 samples total).</p> | |
| </div> | |
| <div class="highlight-card"> | |
| <span class="highlight-icon">π</span> | |
| <div class="highlight-title">High-Resource Only</div> | |
| <p>English, German, Italian, French, Spanish, Portuguese (17,334 samples).</p> | |
| </div> | |
| <div class="highlight-card"> | |
| <span class="highlight-icon">π¬π§</span> | |
| <div class="highlight-title">English Only</div> | |
| <p>Baseline to test whether English alignment is sufficient (2,889 samples).</p> | |
| </div> | |
| <div class="highlight-card"> | |
| <span class="highlight-icon">π</span> | |
| <div class="highlight-title">No English</div> | |
| <p>All 11 non-English languages β tests transfer to English from others.</p> | |
| </div> | |
| </div> | |
| <p>Additionally, 11 individual single-language DPO runs were performed to measure language-specific transfer. Evaluation uses Attack Success Rate (ASR) on RefusEU-test, language consistency, Global MMLU, and an LLM-as-a-Judge fluency/correctness protocol.</p> | |
| </section> | |
| <div class="divider"></div> | |
| <!-- βββ RESULTS ββββββββββββββββββββββββββββββββββββββββ --> | |
| <section> | |
| <div class="section-label">Results</div> | |
| <h2>Key Findings</h2> | |
| <div class="figure-block" id="fig-asr"> | |
| <img src="img/asr-table.png" alt="ASR comparison table results" style="width: 100%; display: block; margin: 0 auto 16px;" /> | |
| <img src="img/asr.png" alt="ASR comparison figure results" style="width: 100%; display: block; margin: 0 auto;" /> | |
| <div class="figure-caption">Table 2 and Figure 2 β Attack Success Rate (ASR %) on RefusEU-test. Lower is better. Balanced multilingual training achieves the lowest ASR across both model sizes.</div> | |
| </div> | |
| <ul class="finding-list"> | |
| <li class="finding-item"> | |
| <div class="finding-dot"></div> | |
| <p class="finding-text"><strong>English-only alignment is insufficient.</strong> Training exclusively on English safety preferences leads to notably higher ASR for low-resource languages, particularly with Llama-70B β demonstrating that cross-lingual safety transfer from English alone cannot be relied upon.</p> | |
| </li> | |
| <li class="finding-item"> | |
| <div class="finding-dot"></div> | |
| <p class="finding-text"><strong>Balanced multilingual training works best.</strong> The lowest average ASR across all languages is consistently achieved by the balanced 12-language configuration for both the 8B and 70B models, with high-resource-only training as a strong second choice.</p> | |
| </li> | |
| <li class="finding-item"> | |
| <div class="finding-dot"></div> | |
| <p class="finding-text"><strong>Linguistic proximity enables transfer.</strong> Closely related language pairs β PolishβCzech and PortugueseβSpanish β exhibit strongly correlated ASR values across training configurations, suggesting that structural similarity facilitates cross-lingual safety generalization.</p> | |
| </li> | |
| <li class="finding-item"> | |
| <div class="finding-dot"></div> | |
| <p class="finding-text"><strong>Language consistency and safety interact non-trivially.</strong> While high language consistency is generally desirable, explicitly enforcing it can reduce safety in smaller models like Llama-8B. Llama-70B achieves near-100% consistency across all configurations; smaller models degrade under English-only setups.</p> | |
| </li> | |
| <li class="finding-item"> | |
| <div class="finding-dot"></div> | |
| <p class="finding-text"><strong>General capabilities are largely preserved.</strong> Performance degradation on Global MMLU stays below 0.006 for both model sizes. For low-resource languages on the 8B model, translation-based pipelines (translate β answer in English β translate back) outperform native-language generation even for the unmodified Instruct baseline.</p> | |
| </li> | |
| </ul> | |
| <div class="figure-block" id="fig-scatter" style="margin-top: 24px;"> | |
| <img src="img/language.png" alt="ASR vs. language consistency scatter plot" style="width: 90%; display: block; margin: 0 auto;" /> | |
| <div class="figure-caption">Figure 3 β ASR vs. language consistency across training setups. Llama-70B with high-resource training achieves the best combined performance.</div> | |
| </div> | |
| </section> | |
| <div class="divider"></div> | |
| <!-- βββ CONTRIBUTIONS βββββββββββββββββββββββββββββββββ --> | |
| <section> | |
| <div class="section-label">Contributions</div> | |
| <h2>Summary</h2> | |
| <div class="highlights-grid"> | |
| <div class="highlight-card"> | |
| <span class="highlight-icon">ποΈ</span> | |
| <div class="highlight-title">RefusEU Dataset</div> | |
| <p>The first DPO-ready multilingual refusal dataset covering 12 European languages, with a fixed contamination-free evaluation split and fully audited safety labels.</p> | |
| </div> | |
| <div class="highlight-card"> | |
| <span class="highlight-icon">π¬</span> | |
| <div class="highlight-title">Controlled Experiments</div> | |
| <p>Systematic ablation across 4 training configurations + 11 single-language runs on deliberately de-safety-aligned base models for clean measurement.</p> | |
| </div> | |
| <div class="highlight-card"> | |
| <span class="highlight-icon">π</span> | |
| <div class="highlight-title">Multidimensional Evaluation</div> | |
| <p>ASR, language consistency, Global MMLU, and fluency/correctness measured across all 12 languages, revealing trade-offs invisible under single-metric reporting.</p> | |
| </div> | |
| </div> | |
| </section> | |
| <!-- βββ BIBTEX ββββββββββββββββββββββββββββββββββββββββ --> | |
| <div class="bibtex-section" id="bibtex"> | |
| <div class="bibtex-inner"> | |
| <div class="section-label">Citation</div> | |
| <h2>BibTeX</h2> | |
| <div class="code-block" id="bibtex-block"> | |
| <button class="copy-btn" onclick="copyBibtex()">Copy</button> | |
| <span class="code-brace">@inproceedings{</span><span class="code-val">krasnodebska2026refuseu</span><span class="code-brace">,</span><br> | |
| <span class="code-key">title</span> = <span class="code-val">{Multilingual Refusal Alignment for Safer Large Language Models}</span>,<br> | |
| <span class="code-key">author</span> = <span class="code-val">{KrasnodΔbska, Aleksandra and Kusa, Wojciech and Lipani, Aldo}</span>,<br> | |
| <span class="code-key">booktitle</span> = <span class="code-val">{Findings of the Association for Computational Linguistics: ACL 2026}</span>,<br> | |
| <span class="code-key">year</span> = <span class="code-val">{2026}</span>,<br> | |
| <span class="code-key">address</span> = <span class="code-val">{San Diego, California, United States}</span>,<br> | |
| <span class="code-key">publisher</span> = <span class="code-val">{Association for Computational Linguistics}</span><br> | |
| <span class="code-brace">}</span> | |
| </div> | |
| </div> | |
| </div> | |
| <footer> | |
| <p style="margin-top:6px;">RefusEU dataset available at <a href="https://huggingface.co/datasets/NASK-PIB/RefusEU" target="_blank">huggingface.co/datasets/NASK-PIB/RefusEU</a></p> | |
| </footer> | |
| <script> | |
| function copyBibtex() { | |
| const raw = `@inproceedings{krasnodebska2026refuseu, | |
| title = {Multilingual Refusal Alignment for Safer Large Language Models}, | |
| author = {KrasnodΔbska, Aleksandra and Kusa, Wojciech and Lipani, Aldo}, | |
| booktitle = {Findings of the Association for Computational Linguistics: ACL 2026}, | |
| year = {2026}, | |
| address = {San Diego, California, United States}, | |
| publisher = {Association for Computational Linguistics} | |
| }`; | |
| navigator.clipboard.writeText(raw).then(() => { | |
| const btn = document.querySelector('.copy-btn'); | |
| btn.textContent = 'Copied!'; | |
| setTimeout(() => btn.textContent = 'Copy', 2000); | |
| }); | |
| } | |
| </script> | |
| </body> | |
| </html> | |