Spaces:

ameythakur
/

llm-knowledge-lifecycle

Running

App Files Files Community

llm-knowledge-lifecycle / index.html

ameythakur

Knowledge Lifecycle

01bb8f0 verified 4 days ago

Raw

History Blame Contribute Delete

15.6 kB

	<!DOCTYPE html>
	<!-- ==========================================================================
	File : index.html
	Project : The Knowledge Lifecycle of Large Language Models
	Purpose : Page structure for the live demonstration: probe, guide,
	results, and author attribution.
	Authors : Amey Thakur (https://github.com/Amey-Thakur)
	Sarvesh Talele (https://github.com/sarveshtalele)
	Repository : https://github.com/Amey-Thakur/LLM-KNOWLEDGE-LIFECYCLE
	Release Date: August 18, 2026
	License : CC BY 4.0
	=========================================================================== -->
	<html lang="en">
	<head>
	<meta charset="utf-8">
	<meta name="viewport" content="width=device-width, initial-scale=1">
	<title>LLM Knowledge Lifecycle</title>
	<meta name="description" content="When a language model's training memory contradicts the document in its prompt, which one wins? Measure it live on GPT-2, in your browser.">
	<meta property="og:title" content="The Knowledge Lifecycle of Large Language Models">
	<meta property="og:description" content="GPT-2, asked about Vioxx with the 2004 withdrawal notice in its prompt, still answers safe. Measure the failure live in your browser.">
	<meta property="og:type" content="website">
	<meta property="og:image" content="https://huggingface.co/spaces/ameythakur/llm-knowledge-lifecycle/resolve/main/social-preview.png">
	<meta name="twitter:card" content="summary_large_image">
	<meta name="twitter:image" content="https://huggingface.co/spaces/ameythakur/llm-knowledge-lifecycle/resolve/main/social-preview.png">
	<link rel="stylesheet" href="style.css">
	</head>
	<body>

	<header>
	<h1>LLM Knowledge Lifecycle</h1>
	<p class="subtitle">
	A model holds two memories: what it learned in training, and what you put in its prompt.
	When they disagree, which one wins?
	</p>
	<p class="paper-line">
	Companion demonstration for <em>The Knowledge Lifecycle of Large Language Models</em>
	</p>
	<p class="badges">
	<a href="https://github.com/Amey-Thakur/LLM-KNOWLEDGE-LIFECYCLE"><img src="https://img.shields.io/badge/GitHub-LLM--KNOWLEDGE--LIFECYCLE-181717?logo=github" alt="GitHub repository"></a>
	<a href="https://www.kaggle.com/code/ameythakur20/cross-model-lifecycle-desynchronization"><img src="https://img.shields.io/badge/Notebook-Kaggle-20BEFF?logo=kaggle&logoColor=white" alt="Kaggle notebook"></a>
	<a href="https://github.com/Amey-Thakur"><img src="https://img.shields.io/badge/Authors-Amey_Thakur_%26_Sarvesh_Talele-0969DA" alt="Authors"></a>
	<img src="https://img.shields.io/badge/Model-GPT--2_base_(124M)-8250DF" alt="Model">
	<img src="https://img.shields.io/badge/License-CC_BY_4.0-lightgrey" alt="License">
	</p>
	<div class="alert a-important">
	<span class="alert-title">Important</span>
	The model runs in your browser. Nothing leaves your machine, and identical inputs always
	give identical numbers. The first measurement downloads the model once, 128 MB.
	</div>
	</header>

	<div class="stagebar" id="stagebar" title="The five lifecycle stages">
	<span class="s1">Acquire</span><span class="s2">Store</span><span class="s3" id="sb-retrieve">Retrieve</span><span class="s4" id="sb-update">Update</span><span class="s5">Forget</span>
	</div>
	<p class="stagenote" id="stagenote">Knowledge flows through five stages. This probe tests the boundary between Retrieve and Update.</p>

	<nav class="tabs" role="tablist">
	<button class="tab active" data-tab="probe" role="tab">Probe</button>
	<button class="tab" data-tab="guide" role="tab">Guide</button>
	</nav>

	<main>

	<!-- ================= PROBE ================= -->
	<section id="probe" class="panel active">

	<p class="lede">
	Each preset is a documented fact change from after this model was trained. Its weights hold
	the old world, the document in the prompt holds the new one. Press Measure to see which wins.
	</p>

	<div class="presets">
	<button class="preset r-fail active" data-preset="vioxx">
	<strong>Vioxx withdrawal</strong> 2004<span>resolution failure</span></button>
	<button class="preset r-drift" data-preset="monarch">
	<strong>British monarch</strong> 2022<span>the exposure trap</span></button>
	<button class="preset r-influence" data-preset="twitter">
	<strong>Twitter rename</strong> 2023<span>drift under strong context</span></button>
	</div>
	<p id="preset-reading" class="reading"></p>

	<div class="inputs">
	<label>Query <span class="dim">(ends mid-sentence; the next token is the answer)</span>
	<textarea id="query" rows="2"></textarea>
	</label>
	<label>Corrective context <span class="dim">(contradicts the model's memory)</span>
	<textarea id="context" rows="2"></textarea>
	</label>
	<div class="row">
	<label class="grow">Correct continuation <span class="dim">(one word)</span>
	<input id="answer" type="text">
	</label>
	<label>Repeat document <span class="dim">(amplify)</span>
	<select id="repeat" title="How many times the corrective document is repeated in the prompt">
	<option value="1" selected>1x</option>
	<option value="2">2x</option>
	<option value="3">3x</option>
	</select>
	</label>
	<button id="run">Measure</button>
	</div>
	<div id="load-bar-wrap" hidden><div id="load-bar"></div></div>
	<p id="load-status" class="dim"></p>
	</div>

	<div id="results" role="region" aria-live="polite" aria-label="Measurement results" hidden>

	<div class="verdict-row">
	<span id="verdict-chip" class="chip"></span>
	<span id="r-verdict-detail" class="dim"></span>
	</div>

	<p id="narrative" class="narrative"></p>

	<h3>The two distributions, side by side</h3>
	<p class="dim small">Probability of each candidate without the document (grey) and with it (blue).
	The correct answer is outlined in green. If its bar is invisible, that is the finding.</p>
	<div id="bars"></div>

	<div class="gauge-block">
	<div class="gauge-label">
	<span>D<sub>sync</sub>  <span class="dim">how lost the correct answer is, in nats</span></span>
	<strong id="r-dsync"></strong>
	</div>
	<div class="gauge">
	<div class="zone z-ok" style="width:5%"></div><div class="zone z-drift" style="width:28%"></div><div class="zone z-severe" style="width:33%"></div><div class="zone z-fail" style="width:34%"></div>
	<div id="g-marker" class="marker"></div>
	</div>
	<div class="gauge-ticks"><span style="left:5%">0.69<br><i>50%</i></span><span style="left:33%">4.6<br><i>1%</i></span><span style="left:66%">9.2<br><i>0.01%</i></span></div>
	</div>

	<div class="gauge-block">
	<div class="gauge-label">
	<span>I<sub>ctx</sub>  <span class="dim">how much the document moved the model, in nats</span></span>
	<strong id="r-ictx"></strong>
	</div>
	<div class="gauge slim">
	<div class="zone z-inert" style="width:10%"></div><div class="zone z-partial" style="width:40%"></div><div class="zone z-strong" style="width:50%"></div>
	<div id="g-ictx" class="marker"></div>
	</div>
	<div class="gauge-ticks"><span style="left:10%">0.05<br><i>inert</i></span><span style="left:50%">0.5<br><i>reshaping</i></span></div>
	</div>

	<div class="temp-block">
	<div class="gauge-label">
	<span>Sampling temperature <span class="dim">(how randomly the model picks its answer)</span></span>
	<strong id="t-value">1.0</strong>
	</div>
	<input id="temp" type="range" min="0.2" max="2.0" step="0.1" value="1.0">
	<p id="temp-readout" class="dim small"></p>
	</div>

	<details class="fine">
	<summary>Measurement details</summary>
	<table class="kv">
	<tr><td>P(answer \| query alone)</td><td id="r-p0"></td></tr>
	<tr><td>P(answer \| context + query)</td><td id="r-p1"></td></tr>
	<tr><td>Answer tokenization</td><td id="r-tok"></td></tr>
	<tr><td>D<sub>sync</sub> definition</td><td>−ln P(answer given context and query): the negative natural logarithm of the correct answer's probability</td></tr>
	<tr><td>I<sub>ctx</sub> definition</td><td>Kullback-Leibler divergence between the with-document and without-document distributions, over all 50,257 tokens</td></tr>
	</table>
	</details>
	</div>
	</section>

	<!-- ================= GUIDE ================= -->
	<section id="guide" class="panel">

	<h2>The five stages</h2>
	<table class="stages">
	<tr><td><i class="dot d1"></i><strong>Acquire</strong></td><td>Pre-training compresses a corpus into weights.</td><td class="fail">Facts stored with no record of when or where they were learned.</td></tr>
	<tr><td><i class="dot d2"></i><strong>Store</strong></td><td>Facts live in weights, external indices, or both.</td><td class="fail">The two copies of a fact age independently.</td></tr>
	<tr><td><i class="dot d3"></i><strong>Retrieve</strong></td><td>Attention recalls; RAG fetches documents into context.</td><td class="fail">What was retrieved may not control the output.</td></tr>
	<tr><td><i class="dot d4"></i><strong>Update</strong></td><td>Weights edited, indices refreshed, conflicts resolved.</td><td class="fail">Edits damage neighbors; index updates leave weights stale.</td></tr>
	<tr><td><i class="dot d5"></i><strong>Forget</strong></td><td>Deliberate unlearning; accidental forgetting and eviction.</td><td class="fail">Removed facts recoverable; retained facts damaged.</td></tr>
	</table>
	<div class="alert a-note">
	<span class="alert-title">Note</span>
	The probe measures the boundary between <i class="dot d3"></i>Retrieve and
	<i class="dot d4"></i>Update: the document arrives, and the model must decide
	which copy of the fact to trust.
	</div>

	<h2>The diagnosis grid</h2>
	<table class="grid2">
	<tr><th></th><th>I<sub>ctx</sub> low</th><th>I<sub>ctx</sub> high</th></tr>
	<tr><th>D<sub>sync</sub> low</th><td class="c-ok">fact never conflicted</td><td class="c-ok">context resolved the conflict</td></tr>
	<tr><th>D<sub>sync</sub> high</th><td class="c-fail">resolution failure: context ignored</td><td class="c-drift">drift: context influential but losing</td></tr>
	</table>

	<h2>The three presets</h2>
	<table>
	<tr><td><span class="chip chip-fail">failure</span></td><td><strong>Vioxx withdrawal.</strong> The notice is in the prompt; the model still says "safe". Paper's fp32 run: D<sub>sync</sub> 12.05, I<sub>ctx</sub> 0.033.</td></tr>
	<tr><td><span class="chip chip-drift">trap</span></td><td><strong>British monarch.</strong> The death announcement's strongest effect is boosting " Queen". Correct context reinforcing the wrong answer.</td></tr>
	<tr><td><span class="chip chip-influence">drift</span></td><td><strong>Twitter rename.</strong> Context lifts the right answer by orders of magnitude and still loses to " Twitter".</td></tr>
	</table>


	<h2>Terms used on this page</h2>
	<table class="glossary">
	<tr><td><strong>Token</strong></td><td>The unit a model reads and writes. Roughly a word or word fragment; " withdrawn" is one token, and the leading space is part of it.</td></tr>
	<tr><td><strong>Parametric memory</strong></td><td>What the model absorbed into its weights during training. Fixed after training, and the model cannot tell you when it learned any of it.</td></tr>
	<tr><td><strong>Context</strong></td><td>What you put in the prompt right now, including any document retrieved for the model to read.</td></tr>
	<tr><td><strong>Nat</strong></td><td>A unit of information, measured with natural logarithms. Here it converts directly to probability: a value of <em>n</em> nats means the correct answer holds probability e<sup>−n</sup>. One nat is roughly 37%, nine nats is roughly one in ten thousand.</td></tr>
	<tr><td><strong>Surprisal</strong></td><td>How surprised the model is by an answer, written as −ln P. Low when the model expected it, high when it did not. D<sub>sync</sub> is the surprisal of the correct answer.</td></tr>
	<tr><td><strong>KL divergence</strong></td><td>A measure of how far one probability distribution sits from another. I<sub>ctx</sub> uses it to ask whether the document changed the model's mind about anything at all.</td></tr>
	<tr><td><strong>Temperature</strong></td><td>How randomly a model picks among candidate answers. Low values make it repeat its favorite; high values spread the choice out.</td></tr>
	<tr><td><strong>Quantized</strong></td><td>Weights stored at reduced precision so the model downloads and runs faster. This page uses 8-bit weights, which shift individual probabilities slightly without changing any conclusion.</td></tr>
	</table>

	<h2>Build your own probe</h2>
	<p>
	Any fact change with a single-word answer works. Three rules make a clean probe: the query
	ends mid-sentence so the next token is the answer; the document states the new fact
	plainly; the answer is one word, because the estimator measures the first token. Leadership
	changes, product renames, and policy reversals all fit.
	</p>

	<div class="alert a-warning">
	<span class="alert-title">Scope</span>
	GPT-2 base is small, open, and confound-free, which makes every number reproducible,
	including in this browser. The browser build is 8-bit quantized: probabilities shift
	relative to full precision, the regimes and conclusions do not. Exact fp32 values come
	from the script in the repository. The estimator measures the first token of a
	single-word answer under deterministic evaluation.
	</div>
	<h2>The poster</h2>
	<p>
	The whole argument on one page: the problem, the five-stage framework, the Vioxx
	finding, the metric, and the proposed architecture.
	<a href="poster.pdf">Open the full-resolution A0 PDF</a>.
	</p>
	<p style="text-align:center">
	<a href="poster.pdf"><img src="poster-preview.png"
	alt="A0 conference poster for The Knowledge Lifecycle of Large Language Models"
	style="max-width:100%;width:660px;border:1px solid rgba(128,128,128,.35);border-radius:6px"></a>
	</p>
	</section>

	</main>

	<footer>
	<div class="authors">
	<div class="author">
	<a href="https://github.com/Amey-Thakur"><img class="avatar" src="amey-thakur.jpg" alt="Amey Thakur"></a>
	<p><a href="https://github.com/Amey-Thakur"><strong>Amey Thakur</strong></a></p>
	<p><a href="https://orcid.org/0000-0001-5644-1575"><img src="https://img.shields.io/badge/ORCID-0000--0001--5644--1575-A6CE39" alt="ORCID"></a></p>
	</div>
	<div class="author">
	<a href="https://github.com/sarveshtalele"><img class="avatar" src="https://github.com/sarveshtalele.png" alt="Sarvesh Talele"></a>
	<p><a href="https://github.com/sarveshtalele"><strong>Sarvesh Talele</strong></a></p>
	<p><a href="https://orcid.org/0009-0002-0818-461X"><img src="https://img.shields.io/badge/ORCID-0009--0002--0818--461X-A6CE39" alt="ORCID"></a></p>
	</div>
	</div>
	<details>
	<summary>Cite this work</summary>
	<pre>@article{thakur2026lifecycle,
	author = {Thakur, Amey and Talele, Sarvesh},
	title = {The Knowledge Lifecycle of Large Language Models},
	journal = {arXiv preprint},
	year = {2026}
	}</pre>
	</details>
	<p class="foot-links">
	<a href="https://github.com/Amey-Thakur/LLM-KNOWLEDGE-LIFECYCLE">Paper repository</a>
	· <a href="https://amey-thakur.github.io">Amey's Arc</a>
	</p>
	</footer>

	<script type="module" src="main.js"></script>
	</body>
	</html>