@inproceedings{chen-etal-2026-warm,
title = "Warm Up with a Rubric, Then Merge: Decoupling Grounding Gains from Reasoning Costs in Vision-Language Models",
author = "Chen, Jingfan and
Wang, Lin and
Su, Jingran and
Qing, Li",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.2148/",
pages = "48839--48855",
ISBN = "979-8-89176-459-0",
abstract = "Process supervision improves perceptual grounding in RLVR-trained vision-language models, but it also incurs repeated verifier queries and reshapes the policy by suppressing exploratory entropy. We show that its benefit and cost are empirically separable: grounding gains emerge early, whereas the exploration cost accumulates with continued supervision, holding the policy{'}s entropy below that of outcome-only RLVR. This hidden behavioral cost shows up as skill tradeoffs on static benchmarks and, more importantly, as a degraded initialization for subsequent outcome-only RLVR that prematurely limits its reachable ceiling. Motivated by this asymmetry, we propose warmup-and-merge: early-stop the process-supervised warmup, merge only a low-rank component of the warmup task vector into the base model, and continue with outcome-only RLVR under an adaptive on-policy reward that targets the model{'}s current perceptual failures. Across multiple benchmarks and backbones, our method improves perception while preserving exploration, achieving stronger final performance at a fraction of the verifier cost of continuous process supervision."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="chen-etal-2026-warm">
<titleInfo>
<title>Warm Up with a Rubric, Then Merge: Decoupling Grounding Gains from Reasoning Costs in Vision-Language Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jingfan</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lin</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jingran</namePart>
<namePart type="family">Su</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Li</namePart>
<namePart type="family">Qing</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 2026 Conference on Empirical Methods in Natural Language Processing</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">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>Process supervision improves perceptual grounding in RLVR-trained vision-language models, but it also incurs repeated verifier queries and reshapes the policy by suppressing exploratory entropy. We show that its benefit and cost are empirically separable: grounding gains emerge early, whereas the exploration cost accumulates with continued supervision, holding the policy’s entropy below that of outcome-only RLVR. This hidden behavioral cost shows up as skill tradeoffs on static benchmarks and, more importantly, as a degraded initialization for subsequent outcome-only RLVR that prematurely limits its reachable ceiling. Motivated by this asymmetry, we propose warmup-and-merge: early-stop the process-supervised warmup, merge only a low-rank component of the warmup task vector into the base model, and continue with outcome-only RLVR under an adaptive on-policy reward that targets the model’s current perceptual failures. Across multiple benchmarks and backbones, our method improves perception while preserving exploration, achieving stronger final performance at a fraction of the verifier cost of continuous process supervision.</abstract>
<identifier type="citekey">chen-etal-2026-warm</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.2148/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>48839</start>
<end>48855</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Warm Up with a Rubric, Then Merge: Decoupling Grounding Gains from Reasoning Costs in Vision-Language Models
%A Chen, Jingfan
%A Wang, Lin
%A Su, Jingran
%A Qing, Li
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F chen-etal-2026-warm
%X Process supervision improves perceptual grounding in RLVR-trained vision-language models, but it also incurs repeated verifier queries and reshapes the policy by suppressing exploratory entropy. We show that its benefit and cost are empirically separable: grounding gains emerge early, whereas the exploration cost accumulates with continued supervision, holding the policy’s entropy below that of outcome-only RLVR. This hidden behavioral cost shows up as skill tradeoffs on static benchmarks and, more importantly, as a degraded initialization for subsequent outcome-only RLVR that prematurely limits its reachable ceiling. Motivated by this asymmetry, we propose warmup-and-merge: early-stop the process-supervised warmup, merge only a low-rank component of the warmup task vector into the base model, and continue with outcome-only RLVR under an adaptive on-policy reward that targets the model’s current perceptual failures. Across multiple benchmarks and backbones, our method improves perception while preserving exploration, achieving stronger final performance at a fraction of the verifier cost of continuous process supervision.
%U https://aclanthology.org/2026.emnlp-main.2148/
%P 48839-48855
Markdown (Informal)
[Warm Up with a Rubric, Then Merge: Decoupling Grounding Gains from Reasoning Costs in Vision-Language Models](https://aclanthology.org/2026.emnlp-main.2148/) (Chen et al., EMNLP 2026)
ACL