<?xml version='1.0' encoding='UTF-8'?><?xml-stylesheet href='static/style.xsl' type='text/xsl'?><OAI-PMH xmlns="http://www.openarchives.org/OAI/2.0/" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xsi:schemaLocation="http://www.openarchives.org/OAI/2.0/ http://www.openarchives.org/OAI/2.0/OAI-PMH.xsd"><responseDate>2026-11-22T11:20:11Z</responseDate><request verb="GetRecord" identifier="oai:ecommons.cornell.edu:1813/126551" metadataPrefix="dim">https://ecommons.cornell.edu/server/oai/request</request><GetRecord><record><header><identifier>oai:ecommons.cornell.edu:1813/126551</identifier><datestamp>2026-08-27T03:16:40Z</datestamp><setSpec>com_1813_35</setSpec><setSpec>col_1813_47</setSpec></header><metadata><dim:dim xmlns:dim="http://www.dspace.org/xmlns/dspace/dim" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xmlns:doc="http://www.lyncode.com/xoai" xsi:schemaLocation="http://www.dspace.org/xmlns/dspace/dim http://www.dspace.org/schema/dim.xsd">
   <dim:field mdschema="dc" element="contributor" qualifier="author">Kim, Woojeong</dim:field>
   <dim:field mdschema="dc" element="contributor" qualifier="chair" lang="en_US">Rush, Alexander</dim:field>
   <dim:field mdschema="dc" element="contributor" qualifier="committeeMember" lang="en_US">Abdelfattah, Mohamed</dim:field>
   <dim:field mdschema="dc" element="contributor" qualifier="committeeMember" lang="en_US">Kuleshov, Volodymyr</dim:field>
   <dim:field mdschema="dc" element="date" qualifier="accessioned">2026-08-26T19:50:50Z</dim:field>
   <dim:field mdschema="dc" element="date" qualifier="available">2026-08-26T19:50:50Z</dim:field>
   <dim:field mdschema="dc" element="date" qualifier="issued">2026-05</dim:field>
   <dim:field mdschema="dc" element="identifier" qualifier="other">Kim_cornellgrad_0058F_15523</dim:field>
   <dim:field mdschema="dc" element="identifier" qualifier="other">https://www.proquest.com/LegacyDocView/DISSNUM/32673301</dim:field>
   <dim:field mdschema="dc" element="identifier" qualifier="uri">https://hdl.handle.net/1813/126551</dim:field>
   <dim:field mdschema="dc" element="identifier" qualifier="doi">https://doi.org/10.7298/x9r4-r685</dim:field>
   <dim:field mdschema="dc" element="description" lang="en_US">135 pages</dim:field>
   <dim:field mdschema="dc" element="description" qualifier="abstract" lang="en_US">Large language models apply compute uniformly by default, yet structural asymmetries at every stage of the model lifecycle make this uniformity systematically wasteful. This thesis identifies three such asymmetries and proposes a targeted solution to each. At the training stage, reinforcement learning post-training is dominated by rollout generation cost, yet existing methods allocate rollout budget uniformly across prompts regardless of their informativeness. Pilot-Commit addresses this with a two-stage allocation framework that estimates per-prompt reward variance with a small pilot budget and concentrates the remaining rollouts on high-variance prompts, matching baseline accuracy with significantly fewer total rollouts. At the inference stage, the prefill and decode phases have mismatched computational profiles: prefill is compute-bound while decode is memory-bound. OverFill exploits this by using the full model for prefill and a width-pruned model for decode, achieving Pareto-optimal accuracy-latency tradeoffs across multiple model scales while incurring minimal latency overhead. At the deployment stage, developers rely on expensive large-model API calls to implement fuzzy functions that resist symbolic specification, paying the cost of intent encoding at every invocation. Program-as-Weights introduces a compiler-interpreter paradigm in which a large neural compiler translates specifications into compact neural programs that a small fixed interpreter executes locally, amortizing compilation cost across all future invocations. Across training, inference, and deployment, the same principle holds: identifying where compute produces value and reallocating accordingly yields substantial gains without changing the underlying objectives.</dim:field>
   <dim:field mdschema="dc" element="language" qualifier="iso">en</dim:field>
   <dim:field mdschema="dc" element="subject" lang="en_US">Large language models</dim:field>
   <dim:field mdschema="dc" element="title" lang="en_US">Adaptive Allocation of Compute Capacity in Large Language Models</dim:field>
   <dim:field mdschema="dc" element="type" lang="en_US">dissertation or thesis</dim:field>
   <dim:field mdschema="dc" element="format" qualifier="mimetype">application/pdf</dim:field>
   <dim:field mdschema="thesis" element="degree" qualifier="discipline">Computer Science</dim:field>
   <dim:field mdschema="thesis" element="degree" qualifier="grantor">Cornell University</dim:field>
   <dim:field mdschema="thesis" element="degree" qualifier="level">Doctor of Philosophy</dim:field>
   <dim:field mdschema="thesis" element="degree" qualifier="name">Ph. D., Computer Science</dim:field>
   <dim:field mdschema="dcterms" element="license">https://hdl.handle.net/1813/59810.2</dim:field>
   <dim:field mdschema="dspace" element="entity" qualifier="type">Publication</dim:field>
   <dim:field mdschema="cris" element="virtual" qualifier="collection" authority="https://ecommons.cornell.edu/handle/1813/47" confidence="600">Cornell Theses and Dissertations</dim:field>
   <dim:field mdschema="cris" element="virtual" qualifier="author">Kim, Woojeong</dim:field>
   <dim:field mdschema="cris" element="virtualsource" qualifier="collection">5893a6ea-7af3-41d7-abc6-04bcd26ab5df</dim:field>
   <dim:field mdschema="others" element="access-status">open.access</dim:field>
   <dim:field mdschema="others" element="access-status">open.access</dim:field>
   <dim:field mdschema="cerif" element="openaire" authority="" confidence="-1">&lt;Publication xmlns="https://www.openaire.eu/cerif-profile/1.1/" id="6ebe9567-c958-4e2e-bff1-45d65047ce07">
	&lt;Type xmlns="https://www.openaire.eu/cerif-profile/vocab/COAR_Publication_Types">http://purl.org/coar/resource_type/c_1843&lt;/Type>
	&lt;Language>en&lt;/Language>
   	&lt;Title>Adaptive Allocation of Compute Capacity in Large Language Models&lt;/Title>
   	&lt;PublishedIn>
    	&lt;Publication>
      	&lt;/Publication>
   	&lt;/PublishedIn>
   	&lt;PublicationDate>2026-05&lt;/PublicationDate>
   	&lt;DOI>https://doi.org/10.7298/x9r4-r685&lt;/DOI>
   	&lt;Authors>
      	&lt;Author>
        	&lt;DisplayName>Kim, Woojeong&lt;/DisplayName>
         	&lt;Affiliation>
         		&lt;OrgUnit>
         		&lt;/OrgUnit>
         	&lt;/Affiliation>
      	&lt;/Author>
	&lt;/Authors>
   	&lt;Editors>
	&lt;/Editors>
    &lt;Publishers>
        &lt;Publisher>
            &lt;OrgUnit />
        &lt;/Publisher>
    &lt;/Publishers>
    &lt;Keyword>Large language models&lt;/Keyword>
   	&lt;Abstract>Large language models apply compute uniformly by default, yet structural asymmetries at every stage of the model lifecycle make this uniformity systematically wasteful. This thesis identifies three such asymmetries and proposes a targeted solution to each. At the training stage, reinforcement learning post-training is dominated by rollout generation cost, yet existing methods allocate rollout budget uniformly across prompts regardless of their informativeness. Pilot-Commit addresses this with a two-stage allocation framework that estimates per-prompt reward variance with a small pilot budget and concentrates the remaining rollouts on high-variance prompts, matching baseline accuracy with significantly fewer total rollouts. At the inference stage, the prefill and decode phases have mismatched computational profiles: prefill is compute-bound while decode is memory-bound. OverFill exploits this by using the full model for prefill and a width-pruned model for decode, achieving Pareto-optimal accuracy-latency tradeoffs across multiple model scales while incurring minimal latency overhead. At the deployment stage, developers rely on expensive large-model API calls to implement fuzzy functions that resist symbolic specification, paying the cost of intent encoding at every invocation. Program-as-Weights introduces a compiler-interpreter paradigm in which a large neural compiler translates specifications into compact neural programs that a small fixed interpreter executes locally, amortizing compilation cost across all future invocations. Across training, inference, and deployment, the same principle holds: identifying where compute produces value and reallocating accordingly yields substantial gains without changing the underlying objectives.&lt;/Abstract>
	&lt;Access xmlns="http://purl.org/coar/access_right" 
    >
    &lt;/Access>
&lt;/Publication>
</dim:field>
</dim:dim>
</metadata></record></GetRecord></OAI-PMH>