@inproceedings{jang-etal-2026-back,
title = "Back to the Future: Forward-Only Committor Projection for Trajectory-Level Collapse Attribution in {RLVR} Reasoning",
author = "Jang, Jaeeun and
Lee, Hansle and
Kim, Sangmin",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1720/",
pages = "36769--36801",
ISBN = "979-8-89176-462-0",
abstract = "GRPO-trained LLMs sharpen single-sample reasoning but collapse in policy entropy: 7B models lose to their base at Pass@64 on AIME. Existing mitigations score tokens from current-step statistics, missing that collapse is a trajectory-level event accumulating through coherent batch drift. We propose FRACT, a per-rollout score projecting a learned T-step committor{'}s gradient onto the entropy direction in the dual Fisher metric, with a forward-only closed form requiring no extra backward through the policy. As a sparse, sign-aware selective KL atop GRPO, FRACT improves Acc@16 over KL-Cov by +.037/+.034/+.045 across Qwen2.5-Base 1.5B/3B/7B and simultaneously raises Pass@1 and Pass@64 on all four benchmarks at 7B{---}mitigating, not merely shifting, the accuracy{--}diversity trade-off. Random-matched-mass gains only +.009 and Anti-FRACT underperforms GRPO by {\ensuremath{-}}.007 at 1.5B{---}the score, not regularization mass, is the cause at {\ensuremath{\sim}}1.5{\%} structural overhead."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="jang-etal-2026-back">
<titleInfo>
<title>Back to the Future: Forward-Only Committor Projection for Trajectory-Level Collapse Attribution in RLVR Reasoning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jaeeun</namePart>
<namePart type="family">Jang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hansle</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sangmin</namePart>
<namePart type="family">Kim</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>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</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">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>GRPO-trained LLMs sharpen single-sample reasoning but collapse in policy entropy: 7B models lose to their base at Pass@64 on AIME. Existing mitigations score tokens from current-step statistics, missing that collapse is a trajectory-level event accumulating through coherent batch drift. We propose FRACT, a per-rollout score projecting a learned T-step committor’s gradient onto the entropy direction in the dual Fisher metric, with a forward-only closed form requiring no extra backward through the policy. As a sparse, sign-aware selective KL atop GRPO, FRACT improves Acc@16 over KL-Cov by +.037/+.034/+.045 across Qwen2.5-Base 1.5B/3B/7B and simultaneously raises Pass@1 and Pass@64 on all four benchmarks at 7B—mitigating, not merely shifting, the accuracy–diversity trade-off. Random-matched-mass gains only +.009 and Anti-FRACT underperforms GRPO by \ensuremath-.007 at 1.5B—the score, not regularization mass, is the cause at \ensuremath\sim1.5% structural overhead.</abstract>
<identifier type="citekey">jang-etal-2026-back</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1720/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>36769</start>
<end>36801</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Back to the Future: Forward-Only Committor Projection for Trajectory-Level Collapse Attribution in RLVR Reasoning
%A Jang, Jaeeun
%A Lee, Hansle
%A Kim, Sangmin
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F jang-etal-2026-back
%X GRPO-trained LLMs sharpen single-sample reasoning but collapse in policy entropy: 7B models lose to their base at Pass@64 on AIME. Existing mitigations score tokens from current-step statistics, missing that collapse is a trajectory-level event accumulating through coherent batch drift. We propose FRACT, a per-rollout score projecting a learned T-step committor’s gradient onto the entropy direction in the dual Fisher metric, with a forward-only closed form requiring no extra backward through the policy. As a sparse, sign-aware selective KL atop GRPO, FRACT improves Acc@16 over KL-Cov by +.037/+.034/+.045 across Qwen2.5-Base 1.5B/3B/7B and simultaneously raises Pass@1 and Pass@64 on all four benchmarks at 7B—mitigating, not merely shifting, the accuracy–diversity trade-off. Random-matched-mass gains only +.009 and Anti-FRACT underperforms GRPO by \ensuremath-.007 at 1.5B—the score, not regularization mass, is the cause at \ensuremath\sim1.5% structural overhead.
%U https://aclanthology.org/2026.findings-emnlp.1720/
%P 36769-36801
Markdown (Informal)
[Back to the Future: Forward-Only Committor Projection for Trajectory-Level Collapse Attribution in RLVR Reasoning](https://aclanthology.org/2026.findings-emnlp.1720/) (Jang et al., Findings 2026)
ACL