1<!DOCTYPE html> 2<html lang="en" dir="ltr"><head><title>Paper2Project - A Tool to Generate Project Ideas from Scientific Papers</title><meta charset="utf-8"/><link rel="preload" href="../index-47d9ba67.css" as="style"/><link rel="preload" href="../prescript-2bfc6315.js" as="script"/><style type="text/css">@font-face {font-family:'Schibsted Grotesk';font-style:normal;font-weight:400;src:url(/cf-fonts/v/schibsted-grotesk/5.2.8/latin/wght/normal.woff2);unicode-range:U+0000-00FF,U+0131,U+0152-0153,U+02BB-02BC,U+02C6,U+02DA,U+02DC,U+0304,U+0308,U+0329,U+2000-206F,U+20AC,U+2122,U+2191,U+2193,U+2212,U+2215,U+FEFF,U+FFFD;font-display:swap;}@font-face {font-family:'Schibsted Grotesk';font-style:normal;font-weight:400;src:url(/cf-fonts/v/schibsted-grotesk/5.2.8/latin-ext/wght/normal.woff2);unicode-range:U+0100-02BA,U+02BD-02C5,U+02C7-02CC,U+02CE-02D7,U+02DD-02FF,U+0304,U+0308,U+0329,U+1D00-1DBF,U+1E00-1E9F,U+1EF2-1EFF,U+2020,U+20A0-20AB,U+20AD-20C0,U+2113,U+2C60-2C7F,U+A720-A7FF;font-display:swap;}@font-face {font-family:'Schibsted Grotesk';font-style:normal;font-weight:700;src:url(/cf-fonts/v/schibsted-grotesk/5.2.8/latin-ext/wght/normal.woff2);unicode-range:U+0100-02BA,U+02BD-02C5,U+02C7-02CC,U+02CE-02D7,U+02DD-02FF,U+0304,U+0308,U+0329,U+1D00-1DBF,U+1E00-1E9F,U+1EF2-1EFF,U+2020,U+20A0-20AB,U+20AD-20C0,U+2113,U+2C60-2C7F,U+A720-A7FF;font-display:swap;}@font-face {font-family:'Schibsted Grotesk';font-style:normal;font-weight:700;src:url(/cf-fonts/v/schibsted-grotesk/5.2.8/latin/wght/normal.woff2);unicode-range:U+0000-00FF,U+0131,U+0152-0153,U+02BB-02BC,U+02C6,U+02DA,U+02DC,U+0304,U+0308,U+0329,U+2000-206F,U+20AC,U+2122,U+2191,U+2193,U+2212,U+2215,U+FEFF,U+FFFD;font-display:swap;}</style><link rel="preconnect" href="https://cdnjs.cloudflare.com" crossorigin="anonymous"/><meta name="viewport" content="width=device-width, initial-scale=1.0"/><meta name="og:site_name" content="Dmitrii's personal blog"/><meta property="og:title" content="Paper2Project - A Tool to Generate Project Ideas from Scientific Papers"/><meta property="og:type" content="website"/><meta name="twitter:card" content="summary_large_image"/><meta name="twitter:title" content="Paper2Project - A Tool to Generate Project Ideas from Scientific Papers"/><meta name="twitter:description" content="TL;DR Iâm building âPaper2projectâ tool, that aims to identify useful/relevant projects from the latest scientific papers."/><meta property="og:description" content="TL;DR Iâm building âPaper2projectâ tool, that aims to identify useful/relevant projects from the latest scientific papers."/><meta property="og:image:alt" content="TL;DR Iâm building âPaper2projectâ tool, that aims to identify useful/relevant projects from the latest scientific papers."/><meta property="twitter:domain" content="eamag.me"/><meta property="og:url" content="https://eamag.me/2025/paper-to-project"/><meta property="twitter:url" content="https://eamag.me/2025/paper-to-project"/><link rel="icon" href="../static/icon.png"/><meta name="description" content="TL;DR Iâm building âPaper2projectâ tool, that aims to identify useful/relevant projects from the latest scientific papers."/><meta name="generator" content="Quartz"/><link href="../index-47d9ba67.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-735924e0.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-589c1cea.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-bbc3f7c0.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-2911f5d1.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-e427079e.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-53b25b2b.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-c93c6a44.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-8a06014f.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-77d6a441.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-6ee21cee.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-1ea6ad18.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-34bdfded.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-caf7c009.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-788c9ca3.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-642859cd.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-4abc06bd.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-274a3dfe.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-94cb6c84.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-58d5ff56.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../component-7262b5a7.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../static/resource-style-a5c806d2.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="../static/resource-style-6bd36952.css" rel="stylesheet" type="text/css" data-persist="true"/><link href="https://cdn.jsdelivr.net/npm/[email protected]/dist/katex.min.css" rel="stylesheet" type="text/css" data-persist="true"/>
2<script src="../prescript-2bfc6315.js" type="4870083964258e09b4948d00-application/javascript" data-persist="true"></script>
2<script type="4870083964258e09b4948d00-application/javascript" data-persist="true">const fetchData = fetch("../static/contentIndex.json").then(data => data.json())</script>
2<meta property="og:image:width" content="1200"/><meta property="og:image:height" content="630"/><meta property="og:image" content="https://eamag.me/2025/paper-to-project-og-image.webp"/><meta property="og:image:url" content="https://eamag.me/2025/paper-to-project-og-image.webp"/><meta name="twitter:image" content="https://eamag.me/2025/paper-to-project-og-image.webp"/><meta property="og:image:type" content="image/webp"/>
2<script data-cfasync="false" nonce="adac23c6-2496-4bb1-bb33-838c76822df4">try{(function(w,d){!function(j,k,l,m){if(j.zaraz)console.error("zaraz is loaded twice");else{j[l]=j[l]||{};j[l].executed=[];j.zaraz={deferred:[],listeners:[]};j.zaraz._v="20";j.zaraz._n="adac23c6-2496-4bb1-bb33-838c76822df4";j.zaraz.q=[];j.zaraz._f=function(n){return async function(){var o=Array.prototype.slice.call(arguments);j.zaraz.q.push({m:n,a:o})}};for(const p of["track","set","debug"])j.zaraz[p]=j.zaraz._f(p);j.zaraz.init=()=>{var q=k.getElementsByTagName(m)[0],r=k.createElement(m),s=k.getElementsByTagName("title")[0];s&&(j[l].t=k.getElementsByTagName("title")[0].text);j[l].x=Math.random();j[l].w=j.screen.width;j[l].h=j.screen.height;j[l].j=j.innerHeight;j[l].e=j.innerWidth;j[l].l=j.location.href;j[l].r=k.referrer;j[l].k=j.screen.colorDepth;j[l].n=k.characterSet;j[l].o=(new Date).getTimezoneOffset();if(j.dataLayer)for(const t of Object.entries(Object.entries(dataLayer).reduce((u,v)=>({...u[1],...v[1]}),{})))zaraz.set(t[0],t[1],{scope:"page"});j[l].q=[];for(;j.zaraz.q.length;){const w=j.zaraz.q.shift();j[l].q.push(w)}r.defer=!0;for(const x of[localStorage,sessionStorage])Object.keys(x||{}).filter(z=>z.startsWith("_zaraz_")).forEach(y=>{try{j[l]["z_"+y.slice(7)]=JSON.parse(x.getItem(y))}catch{j[l]["z_"+y.slice(7)]=x.getItem(y)}});r.referrerPolicy="origin";r.src="/cdn-cgi/zaraz/s.js?z="+btoa(encodeURIComponent(JSON.stringify(j[l])));q.parentNode.insertBefore(r,q)};["complete","interactive"].includes(k.readyState)?zaraz.init():j.addEventListener("DOMContentLoaded",zaraz.init)}}(w,d,"zarazData","script");window.zaraz._p=async cc=>new Promise(cd=>{if(cc){cc.e&&cc.e.forEach(ce=>{try{const cf=d.querySelector("script[nonce]"),cg=cf?.nonce||cf?.getAttribute("nonce"),ch=d.createElement("script");cg&&(ch.nonce=cg);ch.innerHTML=ce;ch.onload=()=>{d.head.removeChild(ch)};d.head.appendChild(ch)}catch(ci){console.error(`Error executing script: ${ce}\n`,ci)}});Promise.allSettled((cc.f||[]).map(cj=>fetch(cj[0],cj[1])))}cd()});zaraz._p({"e":["(function(w,d){})(window,document)"]});})(window,document)}catch(e){throw fetch("/cdn-cgi/zaraz/t"),e;};</script>
2</head><body data-slug="2025/paper-to-project" data-basepath><div id="quartz-root" class="page" data-frame="default"><div id="quartz-body"><div class="left sidebar"><h2 class="page-title"><a href="..">Dmitrii's personal blog</a></h2><div class="mobile-only"><div class="spacer"></div></div><div class="flex-component" style="flex-direction: row; flex-wrap: nowrap; gap: 0.5rem;"><div style="flex-grow: 1; flex-shrink: 1; flex-basis: auto; order: 0; align-self: center; justify-self: center;"><div class="search"><button class="search-button" aria-label="Search" aria-expanded="false"><svg role="img" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 19.9 19.7"><title>Search</title><g class="search-path" fill="none"><path stroke-linecap="square" d="M18.5 18.3l-5.4-5.4"></path><circle cx="8" cy="8" r="7"></circle></g></svg><p>Search</p></button><div class="search-container"><div class="search-space"><input autocomplete="off" class="search-bar" name="search" type="text" aria-label="Search for something..." placeholder="Search for something..."/><div class="search-layout" data-preview="true" data-field-priority="["title","content","tags"]"></div></div></div></div></div><div style="flex-grow: 0; flex-shrink: 1; flex-basis: auto; order: 0; align-self: center; justify-self: center;"><button class="darkmode" aria-label="Dark mode"><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" version="1.1" class="dayIcon" x="0px" y="0px" viewBox="0 0 35 35" style="enable-background:new 0 0 35 35" xml:space="preserve" aria-label="Dark mode"><title>Dark mode</title><path d="M6,17.5C6,16.672,5.328,16,4.5,16h-3C0.672,16,0,16.672,0,17.5 S0.672,19,1.5,19h3C5.328,19,6,18.328,6,17.5z M7.5,26c-0.414,0-0.789,0.168-1.061,0.439l-2,2C4.168,28.711,4,29.086,4,29.5 C4,30.328,4.671,31,5.5,31c0.414,0,0.789-0.168,1.06-0.44l2-2C8.832,28.289,9,27.914,9,27.5C9,26.672,8.329,26,7.5,26z M17.5,6 C18.329,6,19,5.328,19,4.5v-3C19,0.672,18.329,0,17.5,0S16,0.672,16,1.5v3C16,5.328,16.671,6,17.5,6z M27.5,9 c0.414,0,0.789-0.168,1.06-0.439l2-2C30.832,6.289,31,5.914,31,5.5C31,4.672,30.329,4,29.5,4c-0.414,0-0.789,0.168-1.061,0.44 l-2,2C26.168,6.711,26,7.086,26,7.5C26,8.328,26.671,9,27.5,9z M6.439,8.561C6.711,8.832,7.086,9,7.5,9C8.328,9,9,8.328,9,7.5 c0-0.414-0.168-0.789-0.439-1.061l-2-2C6.289,4.168,5.914,4,5.5,4C4.672,4,4,4.672,4,5.5c0,0.414,0.168,0.789,0.439,1.06 L6.439,8.561z M33.5,16h-3c-0.828,0-1.5,0.672-1.5,1.5s0.672,1.5,1.5,1.5h3c0.828,0,1.5-0.672,1.5-1.5S34.328,16,33.5,16z M28.561,26.439C28.289,26.168,27.914,26,27.5,26c-0.828,0-1.5,0.672-1.5,1.5c0,0.414,0.168,0.789,0.439,1.06l2,2 C28.711,30.832,29.086,31,29.5,31c0.828,0,1.5-0.672,1.5-1.5c0-0.414-0.168-0.789-0.439-1.061L28.561,26.439z M17.5,29 c-0.829,0-1.5,0.672-1.5,1.5v3c0,0.828,0.671,1.5,1.5,1.5s1.5-0.672,1.5-1.5v-3C19,29.672,18.329,29,17.5,29z M17.5,7 C11.71,7,7,11.71,7,17.5S11.71,28,17.5,28S28,23.29,28,17.5S23.29,7,17.5,7z M17.5,25c-4.136,0-7.5-3.364-7.5-7.5 c0-4.136,3.364-7.5,7.5-7.5c4.136,0,7.5,3.364,7.5,7.5C25,21.636,21.636,25,17.5,25z"></path></svg><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" version="1.1" class="nightIcon" x="0px" y="0px" viewBox="0 0 100 100" style="enable-background:new 0 0 100 100" xml:space="preserve" aria-label="Light mode"><title>Light mode</title><path d="M96.76,66.458c-0.853-0.852-2.15-1.064-3.23-0.534c-6.063,2.991-12.858,4.571-19.655,4.571 C62.022,70.495,50.88,65.88,42.5,57.5C29.043,44.043,25.658,23.536,34.076,6.47c0.532-1.08,0.318-2.379-0.534-3.23 c-0.851-0.852-2.15-1.064-3.23-0.534c-4.918,2.427-9.375,5.619-13.246,9.491c-9.447,9.447-14.65,22.008-14.65,35.369 c0,13.36,5.203,25.921,14.65,35.368s22.008,14.65,35.368,14.65c13.361,0,25.921-5.203,35.369-14.65 c3.872-3.871,7.064-8.328,9.491-13.246C97.826,68.608,97.611,67.309,96.76,66.458z"></path></svg></button></div><div style="flex-grow: 0; flex-shrink: 1; flex-basis: auto; order: 0; align-self: center; justify-self: center;"><button class="readermode" aria-label="Reader mode"><svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" version="1.1" class="readerIcon" fill="currentColor" stroke="currentColor" stroke-width="0.2" stroke-linecap="round" stroke-linejoin="round" width="64px" height="64px" viewBox="0 0 24 24" aria-label="Reader mode"><title>Reader mode</title><g transform="translate(-1.8, -1.8) scale(1.15, 1.2)"><path d="M8.9891247,2.5 C10.1384702,2.5 11.2209868,2.96705384 12.0049645,3.76669482 C12.7883914,2.96705384 13.8709081,2.5 15.0202536,2.5 L18.7549359,2.5 C19.1691495,2.5 19.5049359,2.83578644 19.5049359,3.25 L19.5046891,4.004 L21.2546891,4.0045
27396 C21.6343849,4.00457396 21.9481801,4.28672784 21.9978425,4.6528034 L22.0046891,4.75457396 L22.0046891,20.25 C22.0046891,20.6296958 21.7225353,20.943491 21.3564597,20.9931534 L21.2546891,21 L2.75468914,21 C2.37499337,21 2.06119817,20.7178461 2.01153575,20.3517706 L2.00468914,20.25 L2.00468914,4.75457396 C2.00468914,4.37487819 2.28684302,4.061083 2.65291858,4.01142057 L2.75468914,4.00457396 L4.50368914,4.004 L4.50444233,3.25 C4.50444233,2.87030423 4.78659621,2.55650904 5.15267177,2.50684662 L5.25444233,2.5 L8.9891247,2.5 Z M4.50368914,5.504 L3.50468914,5.504 L3.50468914,19.5 L10.9478955,19.4998273 C10.4513189,18.9207296 9.73864328,18.5588115 8.96709342,18.5065584 L8.77307039,18.5 L5.25444233,18.5 C4.87474657,18.5 4.56095137,18.2178461 4.51128895,17.8517706 L4.50444233,17.75 L4.50368914,5.504 Z M19.5049359,17.75 C19.5049359,18.1642136 19.1691495,18.5 18.7549359,18.5 L15.2363079,18.5 C14.3910149,18.5 13.5994408,18.8724714 13.0614828,19.4998273 L20.5046891,19.5 L20.5046891,5.504 L19.5046891,5.504 L19.5049359,17.75 Z M18.0059359,3.999 L15.0202536,4 L14.8259077,4.00692283 C13.9889509,4.06666544 13.2254227,4.50975805 12.7549359,5.212 L12.7549359,17.777 L12.7782651,17.7601316 C13.4923805,17.2719483 14.3447024,17 15.2363079,17 L18.0059359,16.999 L18.0056891,4.798 L18.0033792,4.75457396 L18.0056891,4.71 L18.0059359,3.999 Z M8.9891247,4 L6.00368914,3.999 L6.00599909,4.75457396 L6.00599909,4.75457396 L6.00368914,4.783 L6.00368914,16.999 L8.77307039,17 C9.57551536,17 10.3461406,17.2202781 11.0128313,17.6202194 L11.2536891,17.776 L11.2536891,5.211 C10.8200889,4.56369974 10.1361548,4.13636104 9.37521067,4.02745763 L9.18347055,4.00692283 L8.9891247,4 Z"></path></g></svg></button></div></div><div class="explorer nav-files-container" data-behavior="link" data-collapsed="collapsed" data-savestate="true" data-data-fns="{"order":["filter","map","sort"],"sortFn":"(a2, b2) => {\n if (!a2.isFolder && !b2.isFolder || a2.isFolder && b2.isFolder) {\n return (a2.displayName || \"\").localeCompare(b2.displayName || \"\", void 0, {\n numeric: true,\n sensitivity: \"base\"\n });\n }\n if (!a2.isFolder && b2.isFolder) {\n return 1;\n }\n return -1;\n }","filterFn":"(node) => node.slugSegment !== \"tags\"","mapFn":"(node) => {\n return node;\n }"}"><button type="button" class="explorer-toggle mobile-explorer hide-until-loaded" data-mobile="true" aria-controls="explorer-38" aria-label="Explorer"><svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide-menu"><line x1="4" x2="20" y1="12" y2="12"></line><line x1="4" x2="20" y1="6" y2="6"></line><line x1="4" x2="20" y1="18" y2="18"></line></svg></button><button type="button" class="title-button explorer-toggle desktop-explorer" data-mobile="false" aria-expanded="true"><h2>Explorer</h2><svg xmlns="http://www.w3.org/2000/svg" width="14" height="14" viewBox="5 8 14 8" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="fold"><polyline points="6 9 12 15 18 9"></polyline></svg></button><div id="explorer-38" class="explorer-content" aria-expanded="false" role="group"><ul class="explorer-ul overflow" id="list-0"><li class="overflow-end"></li></ul></div><template id="template-file"><li><a href="#" class="nav-file-title tree-item-self"></a></li></template><template id="template-folder"><li><div class="folder-container nav-folder-title tree-item-self"><svg xmlns="http://www.w3.org/2000/svg" width="12" height="12" viewBox="5 8 14 8" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="folder-icon nav-folder-collapse-indicator collapse-icon"><polyline points="6 9 12 15 18 9"></polyline></svg><div><button class="folder-button"><span class="folder-title"></span></button></div></div><div class="folder-outer"><ul class="content tree-item-children"></ul></div></li></template></div></div><div class="center"><div class="page-header"><div class="popover-hint"><nav class="breadcrumb-container" aria-label="breadcrumbs"><div class="breadcrumb-element"><a href="../">Home</a><p> ⯠</p></div><div class="breadcrumb-element"><a href="../2025/">2025</a><p> ⯠</p></div><div class="breadcrumb-element"><a href>Paper2Project - A Tool to Generate Project Ideas from Scientific Papers</a></div></nav><h1 class="article-title">Paper2Project - A Tool to Generate Project Ideas from Scientific Papers</h1>
2<details class="note-properties metadata-container" open data-collapsed="false"><summary class="note-properties-header"><span class="note-properties-title">Properties</span><span class="note-properties-count">1</span></summary><table class="note-properties-table"><tbody><tr class="note-properties-row metadata-property"><td class="note-properties-key metadata-property-key">tags</td><td class="note-properties-value metadata-property-value"><span class="note-properties-tags"><a href="../tags/research" class="internal internal-link tag-link">research</a><span class="note-properties-separator">, </span><a href="../tags/llm" class="internal internal-link tag-link">llm</a><span class="note-properties-separator">, </span><a href="../tags/project" class="internal internal-link tag-link">project</a><span class="note-properties-separator">, </span><a href="../tags/ai" class="internal internal-link tag-link">ai</a><span class="note-properties-separator">, </span><a href="../tags/engineering" class="internal internal-link tag-link">engineering</a></span></td></tr></tbody></table></details><p show-comma="true" class="content-meta"><time datetime="2025-04-03T00:00:00.000Z">Apr 03, 2025</time><span>9 min read</span></p></div></div><article class="popover-hint"><div class="markdown-preview-view markdown-rendered"><h2 id="tldr">TL;DR<a role="anchor" aria-hidden="true" tabindex="-1" data-no-popover="true" href="#tldr" class="internal internal-link"><svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></a></h2> 3<p>Iâm building âPaper2projectâ tool, that aims to identify useful/relevant projects from the latest scientific papers. You can find the first ranked projects here at <a href="https://openreview-copilot.eamag.me/projects" class="external external-link">https://openreview-copilot.eamag.me/projects<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a>. I would like to get more feedback and contributions!</p> 4<h2 id="motivation">Motivation<a role="anchor" aria-hidden="true" tabindex="-1" data-no-popover="true" href="#motivation" class="internal internal-link"><svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></a></h2> 5<p>Some people have lots of time and donât want to waste it scrolling social media. Others want to contribute to scientific progress, but donât know how to apply their skills, or donât know what skills they should develop. This project aims to simplify the barrier of entry to scientific contribution by personalizing potential projects based on a personâs capabilities and interests. Iâve already tried to build a similar thing in <a href="../2024/automated-paper-classification" class="internal internal-link alias" data-slug="2024/automated-paper-classification">Automated Paper Classification</a>, but that was mostly focused on paper categories to keep track of new things in Sparse Autoencoders, for example.</p> 6<h3 id="vision">Vision<a role="anchor" aria-hidden="true" tabindex="-1" data-no-popover="true" href="#vision" class="internal internal-link"><svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></a></h3> 7<p>Based on a personâs interests we suggest scientific projects that bring the most impact to the world. We can collect interests based on their profile from devices or through a chat. We have projects stored and query them on demand, each project is rated in several dimensions. This can be extended to a company-level research to integrate the most relevant updates and patents.</p> 8<h3 id="current-version-scope">Current version scope<a role="anchor" aria-hidden="true" tabindex="-1" data-no-popover="true" href="#current-version-scope" class="internal internal-link"><svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></a></h3> 9<ul> 10<li>Only use papers from ICLR 2025</li> 11<li>Main focus is on top rated papers, and <code>interpretability and explainable AI</code> area</li> 12<li>For each paper projects are extracted based on complexity 13<ul> 14<li>No coding</li> 15<li>
15High school student</li> 16<li><strong>Undergraduate</strong> (main focus)</li> 17<li>PhD in the same field</li> 18<li>PhD in the another field</li> 19<li>Entrepreneur (converting project to a real world usage)</li> 20</ul> 21</li> 22<li>Projects are filtered and ranked through an LLM agent tournament</li> 23<li>Top projects get a handholding step-by-step how-to guide</li> 24</ul> 25<blockquote> 26<p>Example from <a href="https://openreview-copilot.eamag.me/projects" class="external external-link">https://openreview-copilot.eamag.me/projects<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a>: <img src="../2025/paper2project_example.png" alt width="auto" height="auto"/></p> 27</blockquote> 28<h2 id="use-cases">Use cases<a role="anchor" aria-hidden="true" tabindex="-1" data-no-popover="true" href="#use-cases" class="internal internal-link"><svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></a></h2> 29<ul> 30<li>Student: instead of a beginner project <code>Code a simple CNN in Python</code> they get a project for their dissertation about Sparse Autoencoders usage in explainable AI</li> 31<li>Researcher: top rated project is relevant to their initial idea, collaboration with authors</li> 32<li>Entrepreneur: takes a blender procedural generation <a href="https://github.com/mit-gfx/VLMaterial" class="external external-link">paper<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a> and creates an easy to use SaaS-like plugin</li> 33<li>LLM: uses latest attention improvement paper and adds it to an open source repo</li> 34</ul> 35<h2 id="methodology">Methodology<a role="anchor" aria-hidden="true" tabindex="-1" data-no-popover="true" href="#methodology" class="internal internal-link"><svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></a></h2> 36<h3 id="paper-extraction">Paper extraction<a role="anchor" aria-hidden="true" tabindex="-1" data-no-popover="true" href="#paper-extraction" class="internal internal-link"><svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></a></h3> 37<p>Get (mostly anonymous) papers from <a href="https://openreview.net/" class="external external-link">openreview<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a> similar to what I did in <a href="../2025/openreview-for-paperqa" class="internal internal-link alias" data-slug="2025/openreview-for-paperqa">Openreview For PaperQA</a>. Use <a href="https://blog.google/technology/google-deepmind/gemini-model-thinking-updates-march-2025/" class="external external-link">Gemini 2.5 Pro<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a> to extract projects using a structured output like</p> 38<figure data-rehype-pretty-code-figure><pre tabindex="0" data-language="python" data-theme="github-light github-dark"><code data-language="python" data-theme="github-light github-dark" style="display:grid;"><span data-line><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string">"a_reasoning"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">: content.Schema(</span></span> 39<span data-line><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;" data-token-type="important"> type</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;" data-token-type="operator">=</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="function">content.Type.</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="value">STRING</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">,</span></span> 40<span data-line><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="function"> ),</span></span> 41<span data-line><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string"> "project"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">: content.Schema(</span></span> 42<span data-line><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;" data-token-type="important"> type</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;" data-token-type="operator">=</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="function">content.Type.</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="value">OBJECT</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">,</span></span> 43<span data-line><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;" data-token-type="important"> enum</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;" data-token-type="operator">=</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">[],</span></span> 44<span data-line><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;" data-token-type="important"> required</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;" data-token-type="operator">=</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">[</span></span> 45<span data-line><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string"> "high_school"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">,</span></span> 46<span data-line><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string"> "no_coding_experience"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">,</span></span> 47<span data-line><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string"> "entrepreneur"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">,</span></span> 48<span data-line><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string"> "undergraduate"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">,</span></span> 49<span data-line><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string"> "phd_same_field"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">,</span></span> 50<span data-line><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string"> "phd_other_field"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">,</span></span> 51<span data-line><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="function"> ],</span></span> 52<span data-line><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;" data-token-type="important"> properties</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;" data-token-type="operator">=</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">{</span></span> 53<span data-line><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string"> "high_school"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">: content.Schema(</span></span> 54<span data-line><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;" data-token-type="important"> type</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;" data-token-type="operator">=</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="function">content.Type.</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="value">OBJECT</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">,</span></span> 55<span data-line><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;" data-token-type="important"> enum</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;" data-token-type="operator">=</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">[],</span></span> 56<span data-line><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;" data-token-type="important"> required</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;" data-token-type="operator">=</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">[</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string">"project"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">, </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string">"a_reasoning"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">],</span></span> 57<span data-line><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;" data-token-type="important"> properties</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;" data-token-type="operator">=</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">{</span></span> 58<span data-line><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string"> "a_reasoning"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">: content.Schema(</span></span> 59<span data-line><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;" data-token-type="important"> type</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;" data-token-type="operator">=</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="function">content.Type.</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="value">STRING</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">,</span></span> 60<span data-line><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="function"> ),</span></span> 61<span data-line><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string"> "project"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">: content.Schema(</span></span> 62<span data-line><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;" data-token-type="important"> type</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;" data-token-type="operator">=</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="function">content.Type.</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="value">STRING</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">,</span></span> 63<span data-line><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="function"> ),</span></span> 64<span data-line><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="function"> },</span></span> 65<span data-line><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="function"> ),</span></span></code></pre></figure> 66<p>Iâm using <code>a_reasoning</code> so thinking traces are created for each project before the project is created, I noticed it helps with a project ideas.</p> 67<h3 id="embed-and-deduplicate">Embed and deduplicate<a role="anchor" aria-hidden="true" tabindex="-1" data-no-popover="true" href="#embed-and-deduplicate" class="internal internal-link"><svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></a></h3> 68<p>This is an optional step, as I wanted to replicate a <code>Proximity Agent</code> from <a href="https://arxiv.org/pdf/2502.18864v1" class="external external-link">a recent paper from Google<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a>
68. Iâve created embedding using <a href="https://x.com/OfficialLoganK/status/1898081744886087774" class="external external-link">gemini-embedding-exp-03-07<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a> and ran some comparison using cosine similarity, clustering and graph creation via <a href="https://opencollective.com/networkx" class="external external-link">NetworkX<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a>, but these clusters are mostly similar to pre-defined areas from openreview.</p> 69<h3 id="rank-projects-via-agentic-glicko-tournament">Rank projects via agentic Glicko Tournament<a role="anchor" aria-hidden="true" tabindex="-1" data-no-popover="true" href="#rank-projects-via-agentic-glicko-tournament" class="internal internal-link"><svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></a></h3> 70<p>For a category <code>undergraduate</code> I put all projects into a 1v1 <a href="https://en.wikipedia.org/wiki/Swiss-system_tournament" class="external external-link">Swiss-system tournament<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a>. Iâve used gemini models again with a structured output to have a debate, improve an idea and find out what idea is better. Iâve also tried Gemma3:27b. I will put an example output for two papers here.</p> 71<h2 id="next-steps">Next steps<a role="anchor" aria-hidden="true" tabindex="-1" data-no-popover="true" href="#next-steps" class="internal internal-link"><svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></a></h2> 72<p>Please share your thoughts! You can directly text me anywhere, for</p> 73<h3 id="how-to-improve-projects">How to improve projects<a role="anchor" aria-hidden="true" tabindex="-1" data-no-popover="true" href="#how-to-improve-projects" class="internal internal-link"><svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></a></h3> 74<p>If projects are:</p> 75<ul> 76<li>Too simple or not important: add more (negative) examples to the prompt</li> 77<li>Already done by someone: add another agent doing literature search in the project generation step</li> 78<li>Too big or complex: split into smaller sub-projects in the <code>handholding guide</code> stage</li> 79</ul> 80<h3 id="how-to-improve-ranking">How to improve ranking<a role="anchor" aria-hidden="true" tabindex="-1" data-no-popover="true" href="#how-to-improve-ranking" class="internal internal-link"><svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></a></h3> 81<p>I think this is a difficult problem. Right now Iâm relying mostly on LLM improvements, but if I can somehow create a dataset with a ranked ideas and their inspiration, we can RL an LLM on it. This dataset can be created for example by checking the most important citations in papers, so the dataset is <code>{previous_paper: new_paper_title}</code></p> 82<h3 id="how-to-make-this-project-more-useful">How to make this project more useful<a role="anchor" aria-hidden="true" tabindex="-1" data-no-popover="true" href="#how-to-make-this-project-more-useful" class="internal internal-link"><svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></a></h3> 83<p>I think itâs mostly sharing the project, involving more people and gathering feedback to improve. Hence this post :)</p> 84<h2 id="links-and-inspirations">Links and inspirations<a role="anchor" aria-hidden="true" tabindex="-1" data-no-popover="true" href="#links-and-inspirations" class="internal internal-link"><svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></a></h2> 85<ul> 86<li>Neel Nandaâs <a href="https://www.alignmentforum.org/posts/LbrPTJ4fmABEdEnLf/200-concrete-open-problems-in-mechanistic-interpretability" class="external external-link">post about open problem<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a></li> 87<li><a href="https://research.google/blog/accelerating-scientific-breakthroughs-with-an-ai-co-scientist/" class="external external-link">
87AI scientist blog post<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a> from Google</li> 88<li><a href="https://github.com/Just-Curieous/Curie" class="external external-link">Curie<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a> AI agents for science</li> 89<li><a href="https://www.futurehouse.org/" class="external external-link">Future House<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a> automatic scientific discovery</li> 90<li>Sakanaâs <a href="https://github.com/SakanaAI/AI-Scientist" class="external external-link">AI scientist<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a></li> 91<li><a href="https://github.com/ulab-uiuc/research-town" class="external external-link">Research Town simulator<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a></li> 92<li><a href="https://allenai.org/blog/codescientist" class="external external-link">CodeScientist<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a> by Ai2</li> 93</ul> 94<h3 id="footnotes">Footnotes<a role="anchor" aria-hidden="true" tabindex="-1" data-no-popover="true" href="#footnotes" class="internal internal-link"><svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"></path><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"></path></svg></a></h3> 95<ul> 96<li>Idea 1 is based on <a href="https://openreview.net/forum?id=tcsZt9ZNKD" class="external external-link">this<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a>.</li> 97<li>Idea 2 is based on <a href="https://openreview.net/forum?id=SctfBCLmWo" class="external external-link">this<svg aria-hidden="true" class="external-icon" style="max-width:0.8em;max-height:0.8em;" viewBox="0 0 512 512"><path d="M320 0H288V64h32 82.7L201.4 265.4 178.7 288 224 333.3l22.6-22.6L448 109.3V192v32h64V192 32 0H480 320zM32 32H0V64 480v32H32 456h32V480 352 320H424v32 96H64V96h96 32V32H160 32z"></path></svg></a>.</li> 98<li>Ranking agent output:</li> 99</ul> 100<figure data-rehype-pretty-code-figure><pre tabindex="0" data-language="json" data-theme="github-light github-dark"><code data-language="json" data-theme="github-light github-dark" style="display:grid;"><span data-line><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">{</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="punctuation">"debate"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">:{</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="punctuation">"turn_1"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">:</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string">"Expert A: Alright, let's review these proposals. Idea 1, based on the Sparse Autoencoder (SAE) paper, involves implementing mean feature dictionaries for a simplified task. It seems quite aligned with the source paper's core concepts but scales it down significantly. Focuses on interpretability techniques. Expert B: Idea 2 tackles dataset bias, replicating a key experiment from the 'Decade's Battle' paper using standard architectures like ResNet, and then extending it slightly by varying one parameter. This is more in the classic deep learning training/evaluation paradigm. Expert C: Initial thoughts: Idea 1 seems potentially more novel in *topic* for an undergrad, touching on cutting-edge interpretability, but might require more specific guidance on the theoretical underpinnings. Idea 2 feels more standard in *methodology* (train, test, analyze)
100which might be easier for an undergrad to grasp procedurally, but the novelty lies in the specific variable exploration."</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">,</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="punctuation">"turn2"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">:</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string">"Expert C: Let's talk feasibility. Idea 1 requires understanding activation patching and logit differences, using a pre-trained small transformer. Finding or setting up a *suitably small* model and defining a truly simple task is key. The coding itself might not be overly complex for the core logic, but debugging the interpretability setup could be tricky. Access to pre-trained models like GPT-2 Small is feasible. Expert B: Idea 2's feasibility hinges on accessing the datasets (YFCC, CC, DataComp can be large) and compute resources for training ResNets, even smaller ones. Standard deep learning libraries simplify implementation, but dataset handling and managing training runs require infrastructure and time. The variable exploration part seems feasible if kept constrained (e.g., preprocessing variations vs. complex architectural changes). Expert A: Learning potential: Idea 1 offers deep dives into mechanistic interpretability concepts â a valuable, emerging area. Skills involve specific analysis techniques (patching, logit diff), Python, and potentially PyTorch/TensorFlow for model interaction. Idea 2 provides solid experience in the end-to-end deep learning workflow: data handling, model training, evaluation, hyperparameter sensitivity analysis â broadly applicable skills."</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">,</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="punctuation">"turn3"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">:</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string">"Expert B: Novelty and scope: Idea 2's replication is standard, but the specific investigation (e.g., effect of specific object categories on dataset classification) could yield genuinely interesting, novel results within its narrow scope. The scope seems manageable if the variable chosen is well-defined. Expert A: Idea 1's novelty is high in terms of *applying* these specific interpretability methods (mean dictionaries, sufficiency) at an undergrad level. Even on a simplified task, it's less common than standard DL projects. The scope seems appropriate â implement two specific techniques from the paper. Testability is clear via the sufficiency metric. Expert C: Alignment: Both align well. Idea 1 directly implements simplified versions of techniques discussed (Sections 4.2, 4.3). Idea 2 replicates the core experiment and adds a standard sensitivity analysis, a natural extension. Both require careful reading of the source papers."</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">,</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="punctuation">"turn4"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">:</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string">"Expert C: Considering the typical undergrad timeframe (1-2 semesters) and resources, Idea 1 might be slightly less demanding computationally and in terms of data management once the initial conceptual hurdle of interpretability techniques is overcome. The simplified task definition is crucial here. Expert B: However, the conceptual barrier for Idea 1 *could* be higher for students without prior exposure to transformer internals or interpretability. Idea 2 leverages more standard DL workflows, potentially making it easier to get started, even if compute/data requirements are larger. Expert A: I lean slightly towards Idea 1 for its unique learning opportunity in interpretability, assuming adequate mentorship is available. The technical implementation seems less resource-intensive than potentially large-scale training in Idea 2. The 'sufficiency' metric provides a clear quantitative goal. Expert B: I see the appeal of Idea 1's topic, but Idea 2 offers more transferable, standard DL skills. Perhaps it depends on the student's specific interests and the course context. Expert C: Let's weigh feasibility and learning potential. Idea 1 offers specialized learning, potentially higher conceptual difficulty, but lower resource needs. Idea 2 offers broader DL skills, potentially lower conceptual difficulty (for the replication part), but higher resource needs. Both are testable and reasonably scoped if managed well. Given the goal of providing a feasible *and* potentially novel experience, Idea 1 might have a slight edge due to the uniqueness of the interpretability angle for undergrads, provided the task simplification is done effectively."</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">,</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="punctuation">"turn5"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">:</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="value">null</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">,</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="punctuation">"turn6"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">:</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="value">null</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">,</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="punctuation">"turn7"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">:</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="value">null</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">,</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="punctuation">"turn8"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">:</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="value">null</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">,</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="punctuation">"turn9"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">:</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="value">null</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">,</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="punctuation">"turn10"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">:</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="value">null</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">},</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="punctuation">"improvements"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">:{</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="punctuation">"idea_1"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">:</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string">"* Clearly define the 'simplified task' upfront (e.g., specific logic gate simulation, string manipulation).</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="string">\n</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string">* Provide starter code for loading the pre-trained model and accessing activations.</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="string">\n</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string">* Suggest specific layers/heads to analyze based on prior work or preliminary probing.</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="string">\n</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string">* Recommend using a very small model (e.g., distilled versions if available) to speed up iteration.</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="string">\n</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string">* Add a qualitative analysis component: visualize some features or activation patterns."</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">,</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="punctuation">"idea_2"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">:</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string">"* Specify *which* datasets to use and recommend using smaller subsets initially (e.g., 10k images per dataset).</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="string">\n</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string">* Provide clear baseline code or point to a specific repository for the replication part.</sp
100an><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="string">\n</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string">* Strongly constrain the 'single variable' exploration to ensure feasibility (e.g., focus *only* on data augmentation types, or *only* on training set size ratios).</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="string">\n</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string">* Recommend specific compute resources or platforms (e.g., Google Colab Pro, university HPC access).</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="string">\n</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string">* Encourage rigorous statistical testing of the results comparing different variable settings."</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">},</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="punctuation">"judgment"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">:{</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="punctuation">"explanation"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">:</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string">"Both project ideas are derived well from their source papers and offer valuable learning experiences. Idea 1 focuses on cutting-edge interpretability techniques applied to a simplified problem, offering high novelty and specific skill development (activation patching, mechanistic analysis) with potentially lower computational demands but higher conceptual complexity initially. Idea 2 involves a more standard deep learning workflow (replication, sensitivity analysis) focused on dataset bias, offering broader skill development (training pipelines, data handling) but requiring significant data and compute resources. The panel leans towards Idea 1 primarily due to its higher novelty for an undergraduate project and potentially better feasibility regarding computational resources and data management, assuming the task simplification is effective and adequate mentorship is provided for the interpretability concepts. While Idea 2 builds valuable standard skills, its resource requirements and the potentially less novel core replication task make it slightly less ideal than the unique learning opportunity presented by Idea 1."</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">,</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;" data-token-type="punctuation">"result"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">:</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;" data-token-type="string">"idea_1"</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;" data-token-type="punctuation">}}</span></span></code></pre></figure></div></article><hr/><div class="page-footer"></div></div><div class="right sidebar"><div class="graph"><h3>Graph View</h3><div class="graph-outer"><div class="graph-container" data-cfg="{"drag":true,"zoom":true,"depth":1,"scale":1.1,"repelForce":0.5,"centerForce":0.3,"linkDistance":30,"fontSize":0.6,"opacityScale":1,"showTags":true,"removeTags":[],"focusOnHover":false,"enableRadial":false}"></div><button class="global-graph-icon" aria-label="Global Graph"><svg version="1.1" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" x="0px" y="0px" viewBox="0 0 55 55" fill="currentColor" xml:space="preserve"><path d="M49,0c-3.309,0-6,2.691-6,6c0,1.035,0.263,2.009,0.726,2.86l-9.829,9.829C32.542,17.634,30.846,17,29,17 101 s-3.542,0.634-4.898,1.688l-7.669-7.669C16.785,10.424,17,9.74,17,9c0-2.206-1.794-4-4-4S9,6.794,9,9s1.794,4,4,4 102 c0.74,0,1.424-0.215,2.019-0.567l7.669,7.669C21.634,21.458,21,23.154,21,25s0.
102634,3.542,1.688,4.897L10.024,42.562 103 C8.958,41.595,7.549,41,6,41c-3.309,0-6,2.691-6,6s2.691,6,6,6s6-2.691,6-6c0-1.035-0.263-2.009-0.726-2.86l12.829-12.829 104 c1.106,0.86,2.44,1.436,3.898,1.619v10.16c-2.833,0.478-5,2.942-5,5.91c0,3.309,2.691,6,6,6s6-2.691,6-6c0-2.967-2.167-5.431-5-5.91 105 v-10.16c1.458-0.183,2.792-0.759,3.898-1.619l7.669,7.669C41.215,39.576,41,40.26,41,41c0,2.206,1.794,4,4,4s4-1.794,4-4 106 s-1.794-4-4-4c-0.74,0-1.424,0.215-2.019,0.567l-7.669-7.669C36.366,28.542,37,26.846,37,25s-0.634-3.542-1.688-4.897l9.665-9.665 107 C46.042,11.405,47.451,12,49,12c3.309,0,6-2.691,6-6S52.309,0,49,0z M11,9c0-1.103,0.897-2,2-2s2,0.897,2,2s-0.897,2-2,2 108 S11,10.103,11,9z M6,51c-2.206,0-4-1.794-4-4s1.794-4,4-4s4,1.794,4,4S8.206,51,6,51z M33,49c0,2.206-1.794,4-4,4s-4-1.794-4-4 109 s1.794-4,4-4S33,46.794,33,49z M29,31c-3.309,0-6-2.691-6-6s2.691-6,6-6s6,2.691,6,6S32.309,31,29,31z M47,41c0,1.103-0.897,2-2,2 110 s-2-0.897-2-2s0.897-2,2-2S47,39.897,47,41z M49,10c-2.206,0-4-1.794-4-4s1.794-4,4-4s4,1.794,4,4S51.206,10,49,10z"></path></svg></button></div><div class="global-graph-outer"><div class="global-graph-container" data-cfg="{"drag":true,"zoom":true,"depth":-1,"scale":0.9,"repelForce":0.5,"centerForce":0.2,"linkDistance":30,"fontSize":0.6,"opacityScale":1,"showTags":true,"removeTags":[],"focusOnHover":true,"enableRadial":true}"></div></div></div><div class="recent-notes"><h3>Recent Notes</h3><ul class="recent-ul"><li class="recent-li"><div class="section"><div class="desc"><h3><a href="../2026/ai-calibration" class="internal">When an AI Says It's 90% Sure, How Would You Check?</a></h3></div><p class="meta"><time datetime="2026-09-17T00:00:00.000Z">Sep 17, 2026</time></p><ul class="tags"><li><a class="internal tag-link" href="../tags/ai">ai</a></li><li><a class="internal tag-link" href="../tags/llm">llm</a></li><li><a class="internal tag-link" href="../tags/discovery">discovery</a></li><li><a class="internal tag-link" href="../tags/research">research</a></li><li><a class="internal tag-link" href="../tags/ai-safety">ai-safety</a></li><li><a class="internal tag-link" href="../tags/weekend-project">weekend-project</a></li></ul></div></li><li class="recent-li"><div class="section"><div class="desc"><h3><a href="../2026/backgammon-bot" class="internal">I Asked Meta Spark 1.3 to Build an AI Backgammon Bot. Why Did Learning Stall?</a></h3></div><p class="meta"><time datetime="2026-09-15T00:00:00.000Z">Sep 15, 2026</time></p><ul class="tags"><li><a class="internal tag-link" href="../tags/ai">ai</a></li><li><a class="internal tag-link" href="../tags/project">project</a></li><li><a class="internal tag-link" href="../tags/weekend-project">weekend-project</a></li><li><a class="internal tag-link" href="../tags/agents">agents</a></li></ul></div></li><li class="recent-li"><div class="section"><div class="desc"><h3><a href="../" class="internal">Home</a></h3></div><p class="meta"><time datetime="2026-09-15T00:00:00.000Z">Sep 15, 2026</time></p><ul class="tags"></ul></div></li></ul></div><div class="toc"><button type="button" class="toc-header" aria-controls="toc-30" aria-expanded="true"><h3>Table of Contents</h3><svg xmlns="http://www.w3.org/2000/svg" width="24" height="24" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="fold"><polyline points="6 9 12 15 18 9"></polyline></svg></button><ul id="list-0" class="toc-content overflow"><li class="depth-0"><a href="#tldr" data-for="tldr">TL;DR</a></li><li class="depth-0"><a href="#motivation" data-for="motivation">Motivation</a></li><li class="depth-1"><a href="#vision" data-for="vision">Vision</a></li><li class="depth-1"><a href="#current-version-scope" data-for="current-version-scope">Current version scope</a></li><li class="depth-0"><a href="#use-cases" data-for="use-cases">Use cases</a></li><li class="depth-0"><a href="#methodology" data-for="methodology">Methodology</a></li><li class="depth-1"><a href="#paper-extraction" data-for="paper-extraction">Paper extraction</a></li><li class="depth-1"><a href="#embed-and-deduplicate" data-for="embed-and-deduplicate">Embed and deduplicate</a></li><li class="depth-1"><a href="#rank-projects-via-agentic-glicko-tournament" data-for="rank-projects-via-agentic-glicko-tournament">Rank projects via agentic Glicko Tournament</a></li><li class="depth-0"><a href="#next-steps" data-for="next-steps">Next steps</a></li><li class="depth-1"><a href="#how-to-improve-projects" data-for="how-to-improve-projects">How to improve projects</a></li><li class="depth-1"><a href="#how-to-improve-ranking" data-for="how-to-improve-ranking">How to improve ranking</a></li><li class="depth-1"><a href="#how-to-make-this-project-more-useful" data-for="how-to-make-this-project-more-useful">How to make this project more useful</a></li><li class="depth-0"><a href="#links-and-inspirations" data-for="links-and-inspirations">Links and inspirations</a></li><li class="depth-1"><a href="#footnotes" data-for="footnotes">Footnotes</a></li><li class="overflow-end"></li></ul></div><div class="backlinks"><h3>Backlinks</h3><ul id="list-0" class="overflow"><li><a href="../2025/ai-4-life-science-learnings" class="internal">Learnings From 2025 AI For Life Science Conference (AI engineer view)</a></li><li><a href="../2025/categories-of-ai-research-ideas" class="internal">I analyzed 300 ICLR 2025 papers, here's how top AI scientists come up with ideas for their papers, and what AI can learn from it</a></li><li><a href="../2025/neobert-fine-tuning" class="internal">Fine-tuning NeoBERT for a multi-class text classification with a dataset created by LLM</a></li><li><a href="../2025/openreview-for-paperqa" class="internal">Ask ICLR 2025 any question!</a></li><li><a href="../2025/project-ideas" class="internal">Project ideas</a></li><li><a href="../2025/ranking-with-llms" class="internal">Semantically rank everything using LLMs (companies, candidates, ideas etc)</a></li><li><a href="../2025/research-assistant-ai-tools" class="internal">AI Tools To Get Started In Research</a></li><li><a href="../2025/links/links-and-a-retrospective-of-2025" class="internal">Links for 2025 (And A Retrospective)</a></li><li><a href="../2025/links/links-for-april-2025" class="internal">Links for April 2025</a></li><li><a href="../2025/links/links-for-june-2025" class="internal">Links for June 2025</a></li><li><a href="../list-of-all-projects" class="internal">Links To Deployed Projects</a></li><li><a href="../" class="internal">Home</a></li><li class="overflow-end"></li></ul></div></div><footer class><p>Created with <a href="https://quartz.jzhao.xyz/">Quartz v5.0.0</a> © 2026</p><ul><li><a href="https://github.com/eamag">GitHub</a></li><li><a href="https://x.com/EamagAI">Twitter/X</a></li><li><a href="https://www.linkedin.com/in/eamag/">LinkedIn</a></li><li><a href="https://eamag.me/index.xml">RSS</a></li><li><a href="http://eepurl.com/i35ils">Newsletter</a></li></ul></footer></div></div><!-- Cloudflare Pages Analytics -->
vendor: 99 bytes, line 110
110<script defer src='https://static.cloudflareinsights.com/beacon.min.js' data-cf-beacon='{"token": "
1103cc480702b614857b7b14f0e3ea77531
vendor: 61 bytes, line 110
110"}' type="4870083964258e09b4948d00-text/javascript"></script>
110<!-- Cloudflare Pages Analytics -->
vendor: 142 bytes, line 110
110<script src="/cdn-cgi/scripts/7d0fa10a/cloudflare-static/rocket-loader.min.js" data-cf-settings="4870083964258e09b4948d00-|49" defer></script>
110<script type="module" src="https://static.cloudflareinsights.com/beacon.min.js/v31edd6df95cf4e85bb4c19e7a9bdbcba1788362987495" integrity="sha512-iIg7k2xntmwu6/uSb5tpc/hySgZc4eoL31yB29W6tJFo2akwjPWcEqnCEdJvGexCL0KEQwVYv5BlowfhVz26hg==" data-cf-beacon='{"version":"2024.11.0","token":"57213ca53c61439ab0cebc0fe6ef211a","r":1,"spa":2}' crossorigin="anonymous"></script>
110 111</body>
111<script src="../static/resource-after-e13fffa1.js" type="4870083964258e09b4948d00-application/javascript" data-persist="true"></script>
111<script src="../static/resource-after-8b336f9a.js" type="4870083964258e09b4948d00-application/javascript" data-persist="true"></script>
111<script src="../static/resource-after-baba652b.js" type="4870083964258e09b4948d00-application/javascript" data-persist="true"></script>
111<script src="../static/resource-after-5eb75950.js" type="4870083964258e09b4948d00-module" data-persist="true"></script>
111<script src="https://cdn.jsdelivr.net/npm/[email protected]/dist/contrib/copy-tex.min.js" type="4870083964258e09b4948d00-application/javascript" data-persist="true"></script>
111<script src="../postscript-237007ac.js" type="4870083964258e09b4948d00-module" data-persist="true"></script>
111</html>
Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.