Spaces:

CompactAI-O
/

CompactAI-Papers

Running

File size: 22,361 Bytes

<!DOCTYPE html>
<html lang="en">
<head>
  <meta charset="UTF-8">
  <meta name="viewport" content="width=device-width, initial-scale=1.0">
  <title>CompactAI Papers</title>
  <link rel="preconnect" href="https://fonts.googleapis.com">
  <link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
  <link href="https://fonts.googleapis.com/css2?family=Geist:wght@300..700&family=Geist+Mono:wght@400;500&family=Fredoka:wght@700&display=swap" rel="stylesheet">
  <link rel="stylesheet" href="everywhere.css">
</head>
<body>
  <svg width="0" height="0" style="position:absolute;pointer-events:none" aria-hidden="true">
    <defs>
      <filter id="puffy-3d-gold" x="-20%" y="-20%" width="140%" height="140%" color-interpolation-filters="sRGB">
        <feGaussianBlur in="SourceAlpha" stdDeviation="5" result="blur" />
        <feDiffuseLighting in="blur" surfaceScale="10" diffuseConstant="1.2" lighting-color="#ffeaa7" result="diffuse">
          <feDistantLight azimuth="225" elevation="50" />
        </feDiffuseLighting>
        <feComposite in="diffuse" in2="SourceAlpha" operator="in" result="diffuse-clipped" />
        <feFlood flood-color="#f1b51c" result="base-gold" />
        <feBlend mode="multiply" in="base-gold" in2="diffuse-clipped" result="shaded-body" />
        <feSpecularLighting in="blur" surfaceScale="10" specularConstant="1.9" specularExponent="40" lighting-color="#ffffff" result="specular">
          <feDistantLight azimuth="225" elevation="50" />
        </feSpecularLighting>
        <feComposite in="specular" in2="SourceAlpha" operator="in" result="specular-clipped" />
        <feComposite in="specular-clipped" in2="shaded-body" operator="arithmetic" k2="1" k3="1" result="lit-gold" />
        <feComposite in="lit-gold" in2="SourceAlpha" operator="in" />
      </filter>
    </defs>
  </svg>
  <a href="#mainContent" class="skip-link">Skip to papers</a>
  <div class="sr-only" aria-live="polite" aria-atomic="true" id="announcer"></div>

  <header class="site-header">
    <div class="container">
      <h1 class="site-title">Compact<span class="gold-shadow"><span class="gold">AI</span></span> Papers</h1>
      <p class="site-subtitle">Research &amp; ideas from the community</p>
      <div class="header-divider"></div>

      <div class="toolbar" aria-label="Paper tools">
        <label class="toolbar-search">
          <span class="sr-only">Search papers</span>
          <input id="paperSearch" type="search" placeholder="Search title, author, tags" autocomplete="off">
        </label>
        <label class="toolbar-select">
          <span class="sr-only">Sort papers</span>
          <select id="paperSort">
            <option value="featured">Featured first</option>
            <option value="newest">Newest first</option>
            <option value="title">Title A to Z</option>
            <option value="author">Author A to Z</option>
          </select>
        </label>
        <button class="toolbar-chip is-active" type="button" data-filter="all">All</button>
        <button class="toolbar-chip" type="button" data-filter="proved">Proved</button>
        <button class="toolbar-chip" type="button" data-filter="featured">Featured</button>
        <button class="toolbar-reset" type="button" id="resetFilters">Reset</button>
      </div>
    </div>
  </header>

  <main class="container" id="mainContent" aria-label="Papers collection">
    <h2 class="section-label">Papers</h2>
    <ul class="papers-grid" id="papersGrid" role="list"></ul>
    <p class="collection-status" id="collectionStatus" aria-live="polite"></p>
  </main>

  <section class="papermaker-callout" aria-label="Papermaker role">
    <div class="container">
      <p>Have an idea worth writing up? If a mod spots it and likes it, you might earn the <span class="role-name">@Papermaker</span> role &mdash; a nod that you helped push CompactAI forward.</p>
    </div>
  </section>

  <footer class="site-footer">
    <div class="container">
      <span>CompactAI</span>
    </div>
  </footer>

  <script>
    var PAPERS = [
      {
        id: "FewbutLong-md",
        title: "Long context, fewer samples",
        date: "2026-05",
        author: "Costikoooo",
        tags: ["dataset", "pre-processing", "training"],
        proved: false,
        featured: false,
        superFeatured: false,
      },
      {
        id: "STM_paper-md",
        title: "STM and the Circle Thing",
        date: "2026-05",
        author: "Dragonoid",
        tags: ["training", "experimental"],
        proved: true,
        featured: true,
        superFeatured: true,
        proof_script_content: "import torch\nimport torch.nn as nn\nimport torch.optim as optim\nimport random\nimport numpy as np\n\n# STM (Subtractive Training Method) - Definitive Proof\n# Fix: Added Mastery Floor to prevent premature hard-task removal.\n\nN_EASY, N_HARD = 1000, 250\nMASTERY_THRESHOLD = 0.15\nEPOCHS = 25\n\ndef generate_math(difficulty=\"easy\"):\n    if difficulty == \"easy\":\n        a, b = random.randint(0, 9), random.randint(0, 9)\n    else:\n        a, b = random.randint(10, 99), random.randint(10, 99)\n    return f\"{a}+{b}={a+b}\", difficulty\n\ntrain_raw = [generate_math(\"easy\") for _ in range(N_EASY)] + [generate_math(\"hard\") for _ in range(N_HARD)]\ntest_easy = [generate_math(\"easy\") for _ in range(200)]\ntest_hard = [generate_math(\"hard\") for _ in range(200)]\n\nchars = list(\"0123456789+=\") + [\"<PAD>\"]\nchar2idx = {c: i for i, c in enumerate(chars)}\nPAD_IDX, EQ_IDX = char2idx[\"<PAD>\"], char2idx[\"=\"]\n\ndef encode(data_list):\n    max_l = max(len(d[0]) for d in data_list)\n    X, Y = [], []\n    for text, _ in data_list:\n        enc = [char2idx[c] for c in text] + [PAD_IDX] * (max_l - len(text) + 1)\n        X.append(enc[:-1]); Y.append(enc[1:])\n    return torch.tensor(X), torch.tensor(Y)\n\nclass STM_Model(nn.Module):\n    def __init__(self, vocab_size, hidden_size=128):\n        super().__init__()\n        self.embedding = nn.Embedding(vocab_size, hidden_size)\n        self.rnn = nn.GRU(hidden_size, hidden_size, batch_first=True)\n        self.fc = nn.Linear(hidden_size, vocab_size)\n    def forward(self, x):\n        out, _ = self.rnn(self.embedding(x))\n        return self.fc(out)\n\nmodel = STM_Model(len(chars))\noptimizer = optim.Adam(model.parameters(), lr=0.003)\ncriterion = nn.CrossEntropyLoss(ignore_index=PAD_IDX, reduction='none')\n\ndef eval_acc(model, data_list):\n    model.eval()\n    X, Y = encode(data_list)\n    with torch.no_grad():\n        out = model(X).argmax(dim=-1)\n        correct = 0\n        for i in range(len(data_list)):\n            eq_pos = (X[i] == EQ_IDX).nonzero(as_tuple=True)[0].item()\n            pred = out[i, eq_pos:eq_pos+4]\n            target = Y[i, eq_pos:eq_pos+4]\n            mask = (target != PAD_IDX)\n            if torch.equal(pred[mask], target[mask]): correct += 1\n        return correct / len(data_list)\n\ndef get_metrics(model, data_list):\n    model.eval()\n    X, Y = encode(data_list)\n    with torch.no_grad():\n        out = model(X).view(-1, len(chars))\n        loss = criterion(out, Y.view(-1)).view(X.size(0), X.size(1))\n        ans_mask = ((X == EQ_IDX).cumsum(1) - (X == EQ_IDX).float()) * (Y != PAD_IDX).float()\n        return ((loss * ans_mask).sum(1) / (ans_mask.sum(1) + 1e-8)).numpy()\n\ncurrent_dataset = list(train_raw)\nprev_losses = get_metrics(model, current_dataset)\ntotal_tokens = 0\ninitial_tokens = len(train_raw) * 12\n\nfor epoch in range(1, EPOCHS + 1):\n    model.train()\n    X, Y = encode(current_dataset)\n    total_tokens += X.numel()\n    indices = np.random.permutation(len(current_dataset))\n    for i in range(0, len(current_dataset), 64):\n        idx = indices[i:i+64]\n        optimizer.zero_grad()\n        loss = nn.CrossEntropyLoss(ignore_index=PAD_IDX)(model(X[idx]).view(-1, len(chars)), Y[idx].view(-1))\n        loss.backward(); optimizer.step()\n\n    curr_losses = get_metrics(model, current_dataset)\n    drops = prev_losses - curr_losses\n    drop_cutoff = drops.mean() + (drops.std() * 0.5)\n\n    keep_indices = []\n    for i in range(len(current_dataset)):\n        if (drops[i] > drop_cutoff and curr_losses[i] < 0.5) or curr_losses[i] < MASTERY_THRESHOLD:\n            pass\n        else:\n            keep_indices.append(i)\n\n    if not keep_indices: break\n    current_dataset = [current_dataset[i] for i in keep_indices]\n    prev_losses = curr_losses[keep_indices]\n\nprint(\"STM proof complete.\")"
      },
      {
        id: "trueact-md",
        title: "TrueACT: A Different Kind of Neuron",
        date: "2026-05",
        author: "CompactAI",
        tags: ["architecture", "neurons", "math", "experimental"],
        proved: false,
        featured: false
      },
      {
        id: "apollonian_gasket-md",
        title: "Token Embeddings Inside an Integer Apollonian Gasket",
        date: "2026-05",
        author: "Mage",
        tags: ["embeddings", "geometry", "number-theory", "experimental"],
        proved: false,
        featured: false
      },
      {
        id: "overta_hypothesis-md",
        title: "The Overta Hypothesis: Knowledge-Free Foundation Models",
        date: "2026-05",
        author: "Amy",
        tags: ["training", "alignment", "reasoning", "experimental"],
        proved: false,
        featured: false
      },
      {
        id: "attention_experiment-md",
        title: "An Experiment With Attention",
        date: "2026-05",
        author: "Costikoooo",
        tags: ["attention", "benchmark", "experimental"],
        proved: false,
        featured: false
      },
      {
        id: "sparrow_fant-md",
        title: "Sparrow, FANT, and the Weird Stuff That Works",
        date: "2026-05",
        author: "Crownelius",
        tags: ["math", "small-models", "experimental"],
        proved: false,
        featured: false
      }
    ];

    (function() {
      var grid = document.getElementById('papersGrid');
      var searchInput = document.getElementById('paperSearch');
      var sortSelect = document.getElementById('paperSort');
      var resetButton = document.getElementById('resetFilters');
      var filterButtons = document.querySelectorAll('[data-filter]');
      var collectionStatus = document.getElementById('collectionStatus');
      var currentFilter = 'all';
      var viewerContainer = null;
      var modalState = null;

      function announce(msg) {
        var el = document.getElementById('announcer');
        if (el) {
          el.textContent = '';
          setTimeout(function() { el.textContent = msg; }, 50);
        }
      }

      function escapeHtml(str) {
        var div = document.createElement('div');
        div.appendChild(document.createTextNode(str));
        return div.innerHTML;
      }

      function getPaperById(paperId) {
        for (var i = 0; i < PAPERS.length; i++) {
          if (PAPERS[i].id === paperId) return PAPERS[i];
        }
        return null;
      }

      function matchesFilter(paper) {
        if (currentFilter === 'proved' && !paper.proved) return false;
        if (currentFilter === 'featured' && !paper.featured) return false;
        return true;
      }

      function matchesSearch(paper, query) {
        if (!query) return true;
        return [paper.title, paper.author, paper.date, (paper.tags || []).join(' ')].join(' ').toLowerCase().indexOf(query) !== -1;
      }

      function sortPapers(items) {
        var sorted = items.slice();
        var sortValue = sortSelect.value;
        sorted.sort(function(a, b) {
          if (sortValue === 'title') return a.title.localeCompare(b.title);
          if (sortValue === 'author') return (a.author || '').localeCompare(b.author || '');
          if (sortValue === 'newest') return (b.date || '').localeCompare(a.date || '');
          if (b.featured !== a.featured) return (b.featured ? 1 : 0) - (a.featured ? 1 : 0);
          return 0;
        });
        return sorted;
      }

      function renderGrid() {
        var query = searchInput.value.trim().toLowerCase();
        var visible = sortPapers(PAPERS.filter(function(paper) {
          return matchesFilter(paper) && matchesSearch(paper, query);
        }));
        grid.innerHTML = '';

        if (visible.length === 0) {
          grid.innerHTML = '<li class="empty-state" role="listitem"><p>No papers match these filters.</p><button type="button" id="clearEmptyState">Clear filters</button></li>';
          document.getElementById('clearEmptyState').addEventListener('click', resetFilters);
          collectionStatus.textContent = '0 papers shown.';
          return;
        }

        visible.forEach(function(paper, index) {
          grid.appendChild(createCard(paper, index));
        });

        collectionStatus.textContent = visible.length + ' papers shown.';
        requestAnimationFrame(function() {
          observeCards();
        });
      }

      function setFilter(nextFilter) {
        currentFilter = nextFilter;
        filterButtons.forEach(function(button) {
          button.classList.toggle('is-active', button.getAttribute('data-filter') === nextFilter);
        });
        renderGrid();
      }

      function resetFilters() {
        currentFilter = 'all';
        searchInput.value = '';
        sortSelect.value = 'featured';
        filterButtons.forEach(function(button) {
          button.classList.toggle('is-active', button.getAttribute('data-filter') === 'all');
        });
        renderGrid();
        announce('Filters cleared.');
      }

      function syncUrl(paperId) {
        var url = new URL(window.location.href);
        if (paperId) {
          url.searchParams.set('paper', paperId);
        } else {
          url.searchParams.delete('paper');
        }
        window.history.pushState({ paper: paperId }, '', url.toString());
      }

      function createCard(paper, index) {
        var li = document.createElement('li');
        li.setAttribute('role', 'listitem');

        var card = document.createElement('a');
        card.className = 'paper-card' + (paper.featured ? ' featured' : '') + (paper.superFeatured ? ' super-featured' : '');
        card.style.setProperty('--card-index', index);
        card.href = '#paper-' + paper.id;
        card.setAttribute('aria-label', 'Read paper: ' + paper.title);

        card.innerHTML =
          '<h3 class="card-title">' + escapeHtml(paper.title) + '</h3>' +
          (paper.author ? '<div class="card-author">' + escapeHtml(paper.author) + '</div>' : '') +
          '<div class="card-meta" id="meta-' + paper.id + '">' +
            (paper.date ? '<span>' + escapeHtml(paper.date) + '</span>' : '') +
            ((paper.tags || []).map(function(tag) {
              return '<span class="card-tag">' + escapeHtml(tag) + '</span>';
            }).join(' ')) +
            (paper.proved ? '<span class="card-proved">&check; PROVED</span>' : '') +
          '</div>' +
          '<span class="card-arrow" aria-hidden="true">&rarr;</span>';

        card.addEventListener('click', function(e) {
          e.preventDefault();
          window.open('Papers/' + paper.id + '.html', '_blank');
        });

        li.appendChild(card);
        return li;
      }

      var cardObserver = null;

      function observeCards() {
        var cards = grid.querySelectorAll('.paper-card');
        if (!cards.length) return;

        if (cardObserver) {
          cardObserver.disconnect();
        }

        if (!('IntersectionObserver' in window)) {
          cards.forEach(function(card) {
            card.classList.add('is-visible');
          });
          return;
        }

        cardObserver = new IntersectionObserver(function(entries, observer) {
          entries.forEach(function(entry) {
            if (entry.isIntersecting) {
              entry.target.classList.add('is-visible');
              observer.unobserve(entry.target);
            }
          });
        }, {
          threshold: 0.18,
          rootMargin: '0px 0px -8% 0px'
        });

        cards.forEach(function(card) {
          card.classList.remove('is-visible');
          cardObserver.observe(card);
        });
      }

      function openPaper(paper, triggerEl, pushState) {
        closeViewer(false);

        modalState = {
          previousFocus: triggerEl || document.activeElement,
          paper: paper
        };

        if (pushState !== false) syncUrl(paper.id);

        viewerContainer = document.createElement('div');
        viewerContainer.className = 'paper-overlay';
        viewerContainer.setAttribute('role', 'dialog');
        viewerContainer.setAttribute('aria-modal', 'true');
        viewerContainer.setAttribute('aria-labelledby', 'paperDialogTitle');

        viewerContainer.innerHTML =
          '<div class="paper-view" role="document">' +
            '<button class="paper-view-close" type="button" aria-label="Close paper viewer">&times;</button>' +
            '<div class="paper-page">' +
              '<div class="paper-page-shell">' +
                '<h2 id="paperDialogTitle" class="paper-page-title">' + escapeHtml(paper.title) + '</h2>' +
                '<div class="paper-page-meta">' + escapeHtml(paper.author || 'CompactAI') + (paper.date ? ' · ' + escapeHtml(paper.date) : '') + '</div>' +
                '<div class="paper-page-actions">' +
                  '<a class="paper-view-link" href="Papers/' + escapeHtml(paper.id) + '.html" target="_blank" rel="noopener">Open in new tab</a>' +
                  '<button class="paper-view-link" type="button" data-action="copy">Copy link</button>' +
                '</div>' +
                '<iframe class="paper-view-frame" title="' + escapeHtml(paper.title) + '" src="Papers/' + paper.id + '.html"></iframe>' +
              '</div>' +
            '</div>' +
          '</div>';

        document.body.appendChild(viewerContainer);
        document.body.style.overflow = 'hidden';
        document.querySelector('main').setAttribute('inert', '');

        var closeBtn = viewerContainer.querySelector('.paper-view-close');
        var copyBtn = viewerContainer.querySelector('[data-action="copy"]');
        var iframe = viewerContainer.querySelector('.paper-view-frame');

        function close() {
          closeViewer(true);
        }

        closeBtn.addEventListener('click', close);

        copyBtn.addEventListener('click', function() {
          var link = window.location.origin + window.location.pathname + '?paper=' + encodeURIComponent(paper.id);
          if (navigator.clipboard && navigator.clipboard.writeText) {
            navigator.clipboard.writeText(link).then(function() {
              announce('Paper link copied.');
            });
          }
        });

        viewerContainer.addEventListener('click', function(e) {
          if (e.target === viewerContainer) close();
        });

        iframe.addEventListener('load', function() {
          try {
            var doc = iframe.contentDocument;
            var shell = doc.querySelector('.paper-page-shell');
            if (shell) {
              var titleEl = shell.querySelector('.paper-page-title');
              var metaEl = shell.querySelector('.paper-page-meta');
              if (titleEl) titleEl.style.display = 'none';
              if (metaEl) metaEl.style.display = 'none';
            }
            var body = doc.body;
            var html = doc.documentElement;
            if (body) {
              iframe.style.height = Math.max(body.scrollHeight, html.scrollHeight) + 'px';
              iframe.style.overflow = 'hidden';
            }
          } catch(e) {}
          announce('Opened paper viewer: ' + paper.title);
        });

        document.addEventListener('keydown', function handler(e) {
          if (!viewerContainer) {
            document.removeEventListener('keydown', handler);
            return;
          }
          if (e.key === 'Escape') {
            close();
            document.removeEventListener('keydown', handler);
          }
          if (e.key === 'Tab') {
            var focusable = viewerContainer.querySelectorAll('button, a');
            var first = focusable[0];
            var last = focusable[focusable.length - 1];
            if (e.shiftKey && document.activeElement === first) {
              last.focus();
              e.preventDefault();
            } else if (!e.shiftKey && document.activeElement === last) {
              first.focus();
              e.preventDefault();
            }
          }
        });

        closeBtn.focus();
      }

      function closeViewer(syncHistory) {
        if (!viewerContainer) return;
        var previousFocus = modalState && modalState.previousFocus;
        if (viewerContainer.classList.contains('closing')) return;
        viewerContainer.classList.add('closing');
        setTimeout(function() {
          if (!viewerContainer) return;
          viewerContainer.remove();
          viewerContainer = null;
          modalState = null;
          document.body.style.overflow = '';
          document.querySelector('main').removeAttribute('inert');
          if (syncHistory !== false) syncUrl(null);
          if (previousFocus && typeof previousFocus.focus === 'function') previousFocus.focus();
          announce('Paper viewer closed.');
        }, 320);
      }

      searchInput.addEventListener('input', renderGrid);
      sortSelect.addEventListener('change', renderGrid);
      resetButton.addEventListener('click', resetFilters);
      filterButtons.forEach(function(button) {
        button.addEventListener('click', function() {
          setFilter(button.getAttribute('data-filter'));
        });
      });

      window.addEventListener('popstate', function() {
        var params = new URLSearchParams(window.location.search);
        var paperId = params.get('paper');
        if (!paperId) {
          closeViewer(false);
          return;
        }
        var paper = getPaperById(paperId);
        if (paper) openPaper(paper, null, false);
      });

      renderGrid();
      observeCards();

      var initialPaperId = new URLSearchParams(window.location.search).get('paper');
      if (initialPaperId) {
        var initialPaper = getPaperById(initialPaperId);
        if (initialPaper) openPaper(initialPaper, null, false);
      }

    })();
  </script>

</body>
</html>