<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="sawant-2026-gradient">
    <titleInfo>
        <title>When Gradient Importance Lies: Adaptive LoRA Rank Allocation Fails Under GRPO</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Yash</namePart>
        <namePart type="family">Sawant</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <originInfo>
        <dateIssued>2026-10</dateIssued>
    </originInfo>
    <typeOfResource>text</typeOfResource>
    <relatedItem type="host">
        <titleInfo>
            <title>Proceedings of the Seventh Workshop on Insights from Negative Results in NLP</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">João</namePart>
            <namePart type="family">Sedoc</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Raphael</namePart>
            <namePart type="family">Shu</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Aleksandr</namePart>
            <namePart type="family">Drozd</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Arjun</namePart>
            <namePart type="family">Akula</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Anna</namePart>
            <namePart type="family">Rogers</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Salvatore</namePart>
            <namePart type="family">Giorgi</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Adrian</namePart>
            <namePart type="family">Benton</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <originInfo>
            <publisher>Association for Computational Linguistics</publisher>
            <place>
                <placeTerm type="text">Budapest, Hungary</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">conference publication</genre>
        <identifier type="isbn">979-8-89176-480-4</identifier>
    </relatedItem>
    <abstract>Adaptive rank allocation for LoRA—allocating more parameters to important layers and fewer to unimportant ones—consistently improves efficiency under supervised fine-tuning (SFT). We test whether this success transfers to reinforcement learning, specifically Group Relative Policy Optimization (GRPO). Using gradient-magnitude profiling on Qwen 2.5 1.5B with GSM8K, we find that, in our setting, it does not: proportional rank allocation degrades accuracy by 4.5 points compared to uniform allocation (70.0% vs. 74.5%), despite using identical parameter budgets. We identify two mechanisms behind this failure. First, the gradient landscape under GRPO is fundamentally flatter than under SFT—the max-to-min layer importance ratio is only 2.17×, whereas the layer concentration reported by Shi et al. (2024) for SFT (top 30% of layers carrying 80% of the gradient signal) implies a max/min ratio well above 10×. All layers carry meaningful gradient signal; none are truly idle. Second, we observe a gradient amplification effect: non-uniform allocation widens the importance spread from 2.17× to 3.00×, creating a positive feedback loop where high-rank layers absorb more gradient while low-rank layers are progressively silenced. A random-allocation control yields the same amplification (r=0.972 correlation between assigned rank and resulting gradient share), indicating that rank causally determines gradient importance rather than the reverse. The negative result is single-seed and single-task; we present it as preliminary evidence that gradient importance does not predict capacity requirements under RL, and that naïve transfer of SFT-era rank allocation strategies to alignment training should be evaluated cautiously.</abstract>
    <identifier type="citekey">sawant-2026-gradient</identifier>
    <location>
        <url>https://aclanthology.org/2026.insights-1.14/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>110</start>
            <end>114</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T When Gradient Importance Lies: Adaptive LoRA Rank Allocation Fails Under GRPO
%A Sawant, Yash
%Y Sedoc, João
%Y Shu, Raphael
%Y Drozd, Aleksandr
%Y Akula, Arjun
%Y Rogers, Anna
%Y Giorgi, Salvatore
%Y Benton, Adrian
%S Proceedings of the Seventh Workshop on Insights from Negative Results in NLP
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-480-4
%F sawant-2026-gradient
%X Adaptive rank allocation for LoRA—allocating more parameters to important layers and fewer to unimportant ones—consistently improves efficiency under supervised fine-tuning (SFT). We test whether this success transfers to reinforcement learning, specifically Group Relative Policy Optimization (GRPO). Using gradient-magnitude profiling on Qwen 2.5 1.5B with GSM8K, we find that, in our setting, it does not: proportional rank allocation degrades accuracy by 4.5 points compared to uniform allocation (70.0% vs. 74.5%), despite using identical parameter budgets. We identify two mechanisms behind this failure. First, the gradient landscape under GRPO is fundamentally flatter than under SFT—the max-to-min layer importance ratio is only 2.17×, whereas the layer concentration reported by Shi et al. (2024) for SFT (top 30% of layers carrying 80% of the gradient signal) implies a max/min ratio well above 10×. All layers carry meaningful gradient signal; none are truly idle. Second, we observe a gradient amplification effect: non-uniform allocation widens the importance spread from 2.17× to 3.00×, creating a positive feedback loop where high-rank layers absorb more gradient while low-rank layers are progressively silenced. A random-allocation control yields the same amplification (r=0.972 correlation between assigned rank and resulting gradient share), indicating that rank causally determines gradient importance rather than the reverse. The negative result is single-seed and single-task; we present it as preliminary evidence that gradient importance does not predict capacity requirements under RL, and that naïve transfer of SFT-era rank allocation strategies to alignment training should be evaluated cautiously.
%U https://aclanthology.org/2026.insights-1.14/
%P 110-114

Markdown (Informal)

[When Gradient Importance Lies: Adaptive LoRA Rank Allocation Fails Under GRPO](https://aclanthology.org/2026.insights-1.14/) (Sawant, insights 2026)

• When Gradient Importance Lies: Adaptive LoRA Rank Allocation Fails Under GRPO (Sawant, insights 2026)

ACL

• Yash Sawant. 2026. When Gradient Importance Lies: Adaptive LoRA Rank Allocation Fails Under GRPO. In Proceedings of the Seventh Workshop on Insights from Negative Results in NLP, pages 110–114, Budapest, Hungary. Association for Computational Linguistics.