@inproceedings{chen-etal-2026-intervention,
title = "Intervention-Certified Amplification: A Validity Gate for $\sigma$-Weighted Post-Training",
author = "Chen, Leiyue and
Wei, Meng and
Hu, Xuefeng and
Liu, Wei",
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.2090/",
pages = "44510--44522",
ISBN = "979-8-89176-462-0",
abstract = "Post-training methods use training-side signals to decide which supervision to emphasize, but those signals need not identify behavior that is capability-bearing in the target environment. We probe this with DIAA (Dual-Implicit Adversarial Amplification), a per-sample {\ensuremath{\sigma}} = sigmoid(-{\ensuremath{\beta}}{\ensuremath{\Delta}}) SFT/RL hybrid: after every literal Python code fence is deleted from the anchor data, the model still emits unexecuted Tool-Integrated Reasoning (TIR) form on 29{\%} of MATH-500 prompts, three times vanilla GRPO, and suppressing the form recovers accuracy. A paired within-prompt intervention decomposes the removable drag along GRPO {\textrightarrow} SFT {\textrightarrow} DIAA as +2.2 {\textrightarrow} +8.5 {\textrightarrow} +10.6 pp, identifying SFT imitation and {\ensuremath{\sigma}}-amplification as comparable sources; a teacher RFT with 0{\%} TIR likewise trains a student emitting TIR at 45.8{\%}. We propose Intervention-Certified Amplification (ICA), which gates the per-sample weight by a certificate derived from paired rollouts, {\ensuremath{\sigma}}{\_}eff(x) = {\ensuremath{\sigma}}(x){\textperiodcentered}c(x; {\ensuremath{\theta}}{\_}c). Frozen and online instantiations cut TIR to 13.8{\%} and 11.2{\%}, more than halve residual drag while improving accuracy, and carry the joint improvement to held-out GSM8K without retuning the certificate."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="chen-etal-2026-intervention">
<titleInfo>
<title>Intervention-Certified Amplification: A Validity Gate for σ-Weighted Post-Training</title>
</titleInfo>
<name type="personal">
<namePart type="given">Leiyue</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Meng</namePart>
<namePart type="family">Wei</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xuefeng</namePart>
<namePart type="family">Hu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wei</namePart>
<namePart type="family">Liu</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>Post-training methods use training-side signals to decide which supervision to emphasize, but those signals need not identify behavior that is capability-bearing in the target environment. We probe this with DIAA (Dual-Implicit Adversarial Amplification), a per-sample \ensuremathσ = sigmoid(-\ensuremathβ\ensuremathΔ) SFT/RL hybrid: after every literal Python code fence is deleted from the anchor data, the model still emits unexecuted Tool-Integrated Reasoning (TIR) form on 29% of MATH-500 prompts, three times vanilla GRPO, and suppressing the form recovers accuracy. A paired within-prompt intervention decomposes the removable drag along GRPO → SFT → DIAA as +2.2 → +8.5 → +10.6 pp, identifying SFT imitation and \ensuremathσ-amplification as comparable sources; a teacher RFT with 0% TIR likewise trains a student emitting TIR at 45.8%. We propose Intervention-Certified Amplification (ICA), which gates the per-sample weight by a certificate derived from paired rollouts, \ensuremathσ_eff(x) = \ensuremathσ(x)·c(x; \ensuremathθ_c). Frozen and online instantiations cut TIR to 13.8% and 11.2%, more than halve residual drag while improving accuracy, and carry the joint improvement to held-out GSM8K without retuning the certificate.</abstract>
<identifier type="citekey">chen-etal-2026-intervention</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2090/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>44510</start>
<end>44522</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Intervention-Certified Amplification: A Validity Gate for σ-Weighted Post-Training
%A Chen, Leiyue
%A Wei, Meng
%A Hu, Xuefeng
%A Liu, Wei
%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 chen-etal-2026-intervention
%X Post-training methods use training-side signals to decide which supervision to emphasize, but those signals need not identify behavior that is capability-bearing in the target environment. We probe this with DIAA (Dual-Implicit Adversarial Amplification), a per-sample \ensuremathσ = sigmoid(-\ensuremathβ\ensuremathΔ) SFT/RL hybrid: after every literal Python code fence is deleted from the anchor data, the model still emits unexecuted Tool-Integrated Reasoning (TIR) form on 29% of MATH-500 prompts, three times vanilla GRPO, and suppressing the form recovers accuracy. A paired within-prompt intervention decomposes the removable drag along GRPO → SFT → DIAA as +2.2 → +8.5 → +10.6 pp, identifying SFT imitation and \ensuremathσ-amplification as comparable sources; a teacher RFT with 0% TIR likewise trains a student emitting TIR at 45.8%. We propose Intervention-Certified Amplification (ICA), which gates the per-sample weight by a certificate derived from paired rollouts, \ensuremathσ_eff(x) = \ensuremathσ(x)·c(x; \ensuremathθ_c). Frozen and online instantiations cut TIR to 13.8% and 11.2%, more than halve residual drag while improving accuracy, and carry the joint improvement to held-out GSM8K without retuning the certificate.
%U https://aclanthology.org/2026.findings-emnlp.2090/
%P 44510-44522
Markdown (Informal)
[Intervention-Certified Amplification: A Validity Gate for 𝜎-Weighted Post-Training](https://aclanthology.org/2026.findings-emnlp.2090/) (Chen et al., Findings 2026)
ACL