@inproceedings{kundu-etal-2026-forge,
title = "{FORGE}: Forensic Reasoning with Grounded Evidence",
author = "Kundu, Rohit and
Jia, Shan and
Mohanty, Vishal and
Balachandran, Athula and
Roy-Chowdhury, Amit",
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.2054/",
doi = "10.18653/2026.findings-emnlp.2054",
pages = "43745--43767",
ISBN = "979-8-89176-462-0",
abstract = "Forensic deepfake analysis demands more than binary classification: investigators need region-grounded natural language explanations they can verify against the image. Multimodal large language models (MLLMs) are a natural fit, but pretrained MLLMs fail systematically, producing globally coherent text that misses the small localized cues defining manipulations. We argue this is an inductive bias problem rather than a capacity issue: the image-text contrastive objective training MLLM visual encoders optimizes for whole-image semantic summaries, not patch-level forensic detail. The same mismatch explains why prior deepfake reasoning methods target either face manipulation or fully AI-generated content, never both.\par We propose $\textcolor{blue}{\texttt{FORGE}}$, which addresses the mismatch by routing a second visual stream into the language model from a Vision-Only Model (VOM) trained on dense patch prediction rather than image-text alignment. The MLLM{'}s native encoder and the VOM operate on a shared patch grid, which lets us interleave their tokens with preserved spatial correspondence; we show this beats naive concatenation. A two-stage adapter training protocol (generic image-caption alignment, then joint task-specific optimization) prevents the localized stream from overfitting to training-domain manipulations. Across face-manipulated and fully synthetic content, $\textcolor{blue}{\texttt{FORGE}}$ produces region-referential explanations answering fine-grained attribute queries ({``}Does the $\texttt{eyes}$/$\texttt{nose}$/$\texttt{mouth}$ look real or fake?'') and substantially outperforms in-domain baselines on cross-domain evaluations; region-specific evaluation and human studies confirm explanation faithfulness."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="kundu-etal-2026-forge">
<titleInfo>
<title>FORGE: Forensic Reasoning with Grounded Evidence</title>
</titleInfo>
<name type="personal">
<namePart type="given">Rohit</namePart>
<namePart type="family">Kundu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shan</namePart>
<namePart type="family">Jia</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Vishal</namePart>
<namePart type="family">Mohanty</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Athula</namePart>
<namePart type="family">Balachandran</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Amit</namePart>
<namePart type="family">Roy-Chowdhury</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>Forensic deepfake analysis demands more than binary classification: investigators need region-grounded natural language explanations they can verify against the image. Multimodal large language models (MLLMs) are a natural fit, but pretrained MLLMs fail systematically, producing globally coherent text that misses the small localized cues defining manipulations. We argue this is an inductive bias problem rather than a capacity issue: the image-text contrastive objective training MLLM visual encoders optimizes for whole-image semantic summaries, not patch-level forensic detail. The same mismatch explains why prior deepfake reasoning methods target either face manipulation or fully AI-generated content, never both.\par We propose \textcolorblueFORGE, which addresses the mismatch by routing a second visual stream into the language model from a Vision-Only Model (VOM) trained on dense patch prediction rather than image-text alignment. The MLLM’s native encoder and the VOM operate on a shared patch grid, which lets us interleave their tokens with preserved spatial correspondence; we show this beats naive concatenation. A two-stage adapter training protocol (generic image-caption alignment, then joint task-specific optimization) prevents the localized stream from overfitting to training-domain manipulations. Across face-manipulated and fully synthetic content, \textcolorblueFORGE produces region-referential explanations answering fine-grained attribute queries (“Does the eyes/nose/mouth look real or fake?”) and substantially outperforms in-domain baselines on cross-domain evaluations; region-specific evaluation and human studies confirm explanation faithfulness.</abstract>
<identifier type="citekey">kundu-etal-2026-forge</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.2054</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2054/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>43745</start>
<end>43767</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T FORGE: Forensic Reasoning with Grounded Evidence
%A Kundu, Rohit
%A Jia, Shan
%A Mohanty, Vishal
%A Balachandran, Athula
%A Roy-Chowdhury, Amit
%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 kundu-etal-2026-forge
%X Forensic deepfake analysis demands more than binary classification: investigators need region-grounded natural language explanations they can verify against the image. Multimodal large language models (MLLMs) are a natural fit, but pretrained MLLMs fail systematically, producing globally coherent text that misses the small localized cues defining manipulations. We argue this is an inductive bias problem rather than a capacity issue: the image-text contrastive objective training MLLM visual encoders optimizes for whole-image semantic summaries, not patch-level forensic detail. The same mismatch explains why prior deepfake reasoning methods target either face manipulation or fully AI-generated content, never both.\par We propose \textcolorblueFORGE, which addresses the mismatch by routing a second visual stream into the language model from a Vision-Only Model (VOM) trained on dense patch prediction rather than image-text alignment. The MLLM’s native encoder and the VOM operate on a shared patch grid, which lets us interleave their tokens with preserved spatial correspondence; we show this beats naive concatenation. A two-stage adapter training protocol (generic image-caption alignment, then joint task-specific optimization) prevents the localized stream from overfitting to training-domain manipulations. Across face-manipulated and fully synthetic content, \textcolorblueFORGE produces region-referential explanations answering fine-grained attribute queries (“Does the eyes/nose/mouth look real or fake?”) and substantially outperforms in-domain baselines on cross-domain evaluations; region-specific evaluation and human studies confirm explanation faithfulness.
%R 10.18653/2026.findings-emnlp.2054
%U https://aclanthology.org/2026.findings-emnlp.2054/
%U https://doi.org/10.18653/2026.findings-emnlp.2054
%P 43745-43767
Markdown (Informal)
[FORGE: Forensic Reasoning with Grounded Evidence](https://aclanthology.org/2026.findings-emnlp.2054/) (Kundu et al., Findings 2026)
ACL
- Rohit Kundu, Shan Jia, Vishal Mohanty, Athula Balachandran, and Amit Roy-Chowdhury. 2026. FORGE: Forensic Reasoning with Grounded Evidence. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 43745–43767, Budapest, Hungary. Association for Computational Linguistics.