@inproceedings{savkin-etal-2026-trust,
title = "Trust the Context, Not the Weights: Faithful Multi-Hop {QA} via Synthetic Mid-Training",
author = "Savkin, Maksim and
Goncharov, Mikhail and
Gambashidze, Alexander and
Chepurova, Alla and
Tarasov, Dmitrii and
Ionov, Timur and
Andriianov, Nikita and
Pugacheva, Daria and
Oseledets, Ivan and
Konovalov, Vasily and
Galichin, Andrey V.",
editor = "Agrawal, Sweta and
Abad, Alberto",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: System Demonstrations",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-demos.47/",
pages = "488--497",
ISBN = "979-8-89176-461-3",
abstract = "Recent progress in language models has been defined by scale, yet many practical applications benefit more from robust reasoning than extensive parametric knowledge. We introduce OCC-RAG, a family of models for faithful context-based QA, which prioritizes task-specific reasoning capabilities over knowledge capacity. To train OCC-RAG, we implement a novel pipeline synthesizing over three million multi-context, multi-hop QA examples that target multi-hop reasoning, strict context faithfulness, and calibrated abstention. Mid-trained on this corpus, the models produce structured reasoning traces with source citations grounded in literal quotes from the context. Despite their compact size, OCC-RAG models outperform their same-size general-purpose counterparts on every benchmark and exceed models 2{--}$6\times$ larger on financial reasoning (TAT-QA) and faithfulness (ConFiQA), attaining the best results among all evaluated open-weight models up to 32B parameters. On ConFiQA they rank behind only a frontier proprietary models; on multi-hop reasoning (HotpotQA, MuSiQue) and refusal (MuSiQue-Un) they are on par with 8B-class models and close much of the remaining gap to the largest evaluated systems. We release both our \href{https://github.com/optimal-cognitive-core}{code} and \href{https://huggingface.co/collections/occ-ai/occ-rag}{family of models}.[{\ensuremath{<}}https://hf.co/collections/occ-ai/occ-rag{\ensuremath{>}}]"
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="savkin-etal-2026-trust">
<titleInfo>
<title>Trust the Context, Not the Weights: Faithful Multi-Hop QA via Synthetic Mid-Training</title>
</titleInfo>
<name type="personal">
<namePart type="given">Maksim</namePart>
<namePart type="family">Savkin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mikhail</namePart>
<namePart type="family">Goncharov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alexander</namePart>
<namePart type="family">Gambashidze</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alla</namePart>
<namePart type="family">Chepurova</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dmitrii</namePart>
<namePart type="family">Tarasov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Timur</namePart>
<namePart type="family">Ionov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nikita</namePart>
<namePart type="family">Andriianov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Daria</namePart>
<namePart type="family">Pugacheva</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ivan</namePart>
<namePart type="family">Oseledets</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Vasily</namePart>
<namePart type="family">Konovalov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Andrey</namePart>
<namePart type="given">V</namePart>
<namePart type="family">Galichin</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: System Demonstrations</title>
</titleInfo>
<name type="personal">
<namePart type="given">Sweta</namePart>
<namePart type="family">Agrawal</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alberto</namePart>
<namePart type="family">Abad</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">979-8-89176-461-3</identifier>
</relatedItem>
<abstract>Recent progress in language models has been defined by scale, yet many practical applications benefit more from robust reasoning than extensive parametric knowledge. We introduce OCC-RAG, a family of models for faithful context-based QA, which prioritizes task-specific reasoning capabilities over knowledge capacity. To train OCC-RAG, we implement a novel pipeline synthesizing over three million multi-context, multi-hop QA examples that target multi-hop reasoning, strict context faithfulness, and calibrated abstention. Mid-trained on this corpus, the models produce structured reasoning traces with source citations grounded in literal quotes from the context. Despite their compact size, OCC-RAG models outperform their same-size general-purpose counterparts on every benchmark and exceed models 2–6\times larger on financial reasoning (TAT-QA) and faithfulness (ConFiQA), attaining the best results among all evaluated open-weight models up to 32B parameters. On ConFiQA they rank behind only a frontier proprietary models; on multi-hop reasoning (HotpotQA, MuSiQue) and refusal (MuSiQue-Un) they are on par with 8B-class models and close much of the remaining gap to the largest evaluated systems. We release both our \hrefhttps://github.com/optimal-cognitive-corecode and \hrefhttps://huggingface.co/collections/occ-ai/occ-ragfamily of models.[\ensuremath<https://hf.co/collections/occ-ai/occ-rag\ensuremath>]</abstract>
<identifier type="citekey">savkin-etal-2026-trust</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-demos.47/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>488</start>
<end>497</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Trust the Context, Not the Weights: Faithful Multi-Hop QA via Synthetic Mid-Training
%A Savkin, Maksim
%A Goncharov, Mikhail
%A Gambashidze, Alexander
%A Chepurova, Alla
%A Tarasov, Dmitrii
%A Ionov, Timur
%A Andriianov, Nikita
%A Pugacheva, Daria
%A Oseledets, Ivan
%A Konovalov, Vasily
%A Galichin, Andrey V.
%Y Agrawal, Sweta
%Y Abad, Alberto
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: System Demonstrations
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-461-3
%F savkin-etal-2026-trust
%X Recent progress in language models has been defined by scale, yet many practical applications benefit more from robust reasoning than extensive parametric knowledge. We introduce OCC-RAG, a family of models for faithful context-based QA, which prioritizes task-specific reasoning capabilities over knowledge capacity. To train OCC-RAG, we implement a novel pipeline synthesizing over three million multi-context, multi-hop QA examples that target multi-hop reasoning, strict context faithfulness, and calibrated abstention. Mid-trained on this corpus, the models produce structured reasoning traces with source citations grounded in literal quotes from the context. Despite their compact size, OCC-RAG models outperform their same-size general-purpose counterparts on every benchmark and exceed models 2–6\times larger on financial reasoning (TAT-QA) and faithfulness (ConFiQA), attaining the best results among all evaluated open-weight models up to 32B parameters. On ConFiQA they rank behind only a frontier proprietary models; on multi-hop reasoning (HotpotQA, MuSiQue) and refusal (MuSiQue-Un) they are on par with 8B-class models and close much of the remaining gap to the largest evaluated systems. We release both our \hrefhttps://github.com/optimal-cognitive-corecode and \hrefhttps://huggingface.co/collections/occ-ai/occ-ragfamily of models.[\ensuremath<https://hf.co/collections/occ-ai/occ-rag\ensuremath>]
%U https://aclanthology.org/2026.emnlp-demos.47/
%P 488-497
Markdown (Informal)
[Trust the Context, Not the Weights: Faithful Multi-Hop QA via Synthetic Mid-Training](https://aclanthology.org/2026.emnlp-demos.47/) (Savkin et al., EMNLP 2026)
ACL
- Maksim Savkin, Mikhail Goncharov, Alexander Gambashidze, Alla Chepurova, Dmitrii Tarasov, Timur Ionov, Nikita Andriianov, Daria Pugacheva, Ivan Oseledets, Vasily Konovalov, and Andrey V. Galichin. 2026. Trust the Context, Not the Weights: Faithful Multi-Hop QA via Synthetic Mid-Training. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 488–497, Budapest, Hungary. Association for Computational Linguistics.