@inproceedings{alisyam-etal-2026-tokens,
title = "Not All Tokens Earn the Reward: Contrastive Credit Assignment for Production Detection",
author = "Alisyam, Tejas and
Sircar, Prateek and
Gupta, Deepak",
editor = "Atanasova, Pepa and
Haffari, Gholamreza and
Wang, Minghan",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-industry.198/",
pages = "3322--3333",
ISBN = "979-8-89176-460-6",
abstract = "Seller-authored product titles often contradict their own bullet points (a title reading ``128GB'' against a bullet claiming ``512GB''), which misleads shoppers and drives returns. At catalog scale, detection is precision-binding: even a small false-positive rate flags a huge absolute volume of good listings. We deploy a two-stage cascade on one national storefront of a large e-commerce marketplace: a compact 4B detector flags candidates and a strong reasoning model (DeepSeek{~}R1) verifies them, reaching {\ensuremath{>}}99{\%} precision while scoring all 36M listings. Verification is the expensive stage, so our goal is to fold it into the cheap model and run on the 4B detector alone; until then, every gain in the detector{'}s recall catches more defects per verifier call and lifts coverage. We improve that first stage with Contrastive Token Weighting (CTW), a per-token credit-assignment method for GRPO. GRPO spreads one sequence-level advantage uniformly across a completion, even though only a few pivotal tokens decide correctness; CTW instead reads credit from a signal the reward already computes, the split of each sampled group into correct and incorrect completions, and upweights positions where a frozen scorer tells the two apart. The scorer can be a separate stronger teacher (CTW-T) or, when none is available, a frozen copy of the student itself (CTW-S, teacher-free); the teacher-free variant is the one we deploy, and it lifts the detector to 85.9 F1 (+2.2 over SFT, with the gain concentrated in recall, +3.4). A controlled math study shows the reward-induced partition, not the reweighting itself, is what drives the gain."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="alisyam-etal-2026-tokens">
<titleInfo>
<title>Not All Tokens Earn the Reward: Contrastive Credit Assignment for Production Detection</title>
</titleInfo>
<name type="personal">
<namePart type="given">Tejas</namePart>
<namePart type="family">Alisyam</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Prateek</namePart>
<namePart type="family">Sircar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Deepak</namePart>
<namePart type="family">Gupta</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: Industry Track</title>
</titleInfo>
<name type="personal">
<namePart type="given">Pepa</namePart>
<namePart type="family">Atanasova</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Gholamreza</namePart>
<namePart type="family">Haffari</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Minghan</namePart>
<namePart type="family">Wang</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">979-8-89176-460-6</identifier>
</relatedItem>
<abstract>Seller-authored product titles often contradict their own bullet points (a title reading “128GB” against a bullet claiming “512GB”), which misleads shoppers and drives returns. At catalog scale, detection is precision-binding: even a small false-positive rate flags a huge absolute volume of good listings. We deploy a two-stage cascade on one national storefront of a large e-commerce marketplace: a compact 4B detector flags candidates and a strong reasoning model (DeepSeek R1) verifies them, reaching \ensuremath>99% precision while scoring all 36M listings. Verification is the expensive stage, so our goal is to fold it into the cheap model and run on the 4B detector alone; until then, every gain in the detector’s recall catches more defects per verifier call and lifts coverage. We improve that first stage with Contrastive Token Weighting (CTW), a per-token credit-assignment method for GRPO. GRPO spreads one sequence-level advantage uniformly across a completion, even though only a few pivotal tokens decide correctness; CTW instead reads credit from a signal the reward already computes, the split of each sampled group into correct and incorrect completions, and upweights positions where a frozen scorer tells the two apart. The scorer can be a separate stronger teacher (CTW-T) or, when none is available, a frozen copy of the student itself (CTW-S, teacher-free); the teacher-free variant is the one we deploy, and it lifts the detector to 85.9 F1 (+2.2 over SFT, with the gain concentrated in recall, +3.4). A controlled math study shows the reward-induced partition, not the reweighting itself, is what drives the gain.</abstract>
<identifier type="citekey">alisyam-etal-2026-tokens</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-industry.198/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>3322</start>
<end>3333</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Not All Tokens Earn the Reward: Contrastive Credit Assignment for Production Detection
%A Alisyam, Tejas
%A Sircar, Prateek
%A Gupta, Deepak
%Y Atanasova, Pepa
%Y Haffari, Gholamreza
%Y Wang, Minghan
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-460-6
%F alisyam-etal-2026-tokens
%X Seller-authored product titles often contradict their own bullet points (a title reading “128GB” against a bullet claiming “512GB”), which misleads shoppers and drives returns. At catalog scale, detection is precision-binding: even a small false-positive rate flags a huge absolute volume of good listings. We deploy a two-stage cascade on one national storefront of a large e-commerce marketplace: a compact 4B detector flags candidates and a strong reasoning model (DeepSeek R1) verifies them, reaching \ensuremath>99% precision while scoring all 36M listings. Verification is the expensive stage, so our goal is to fold it into the cheap model and run on the 4B detector alone; until then, every gain in the detector’s recall catches more defects per verifier call and lifts coverage. We improve that first stage with Contrastive Token Weighting (CTW), a per-token credit-assignment method for GRPO. GRPO spreads one sequence-level advantage uniformly across a completion, even though only a few pivotal tokens decide correctness; CTW instead reads credit from a signal the reward already computes, the split of each sampled group into correct and incorrect completions, and upweights positions where a frozen scorer tells the two apart. The scorer can be a separate stronger teacher (CTW-T) or, when none is available, a frozen copy of the student itself (CTW-S, teacher-free); the teacher-free variant is the one we deploy, and it lifts the detector to 85.9 F1 (+2.2 over SFT, with the gain concentrated in recall, +3.4). A controlled math study shows the reward-induced partition, not the reweighting itself, is what drives the gain.
%U https://aclanthology.org/2026.emnlp-industry.198/
%P 3322-3333
Markdown (Informal)
[Not All Tokens Earn the Reward: Contrastive Credit Assignment for Production Detection](https://aclanthology.org/2026.emnlp-industry.198/) (Alisyam et al., EMNLP 2026)
ACL