OmniTrace: A Unified Framework for Generation-Time Source Attribution in Omni-Modal LLMs

Qianqi Yan; Yichen Guo; Ching-Chen Kuo; Shan Jiang; Hang Yin; Yang Zhao; Xin Eric Wang

cs.CLMar 20, 2026

OmniTrace: A Unified Framework for Generation-Time Source Attribution in Omni-Modal LLMs

Authors: Qianqi Yan, Yichen Guo, Ching-Chen Kuo, Shan Jiang, Hang Yin, Yang Zhao, Xin Eric Wang

Organizations: University of California, Santa Barbara · eBay

Abstract

An omni-modal language model can answer a question about a multimodal document, a recording, or a video, but it cannot show which part of the input the answer came from. Asking the model to cite its evidence yields what it says it used; measuring what it used has so far required rerunning it many times with parts of the input removed, a cost that grows with every second of audio and video. We show that the generation pass already holds most of this information. OmniTrace attributes each generated sentence to source units (text chunks, images, and seconds of audio or video) by reading late-layer attention as mass over units with no extra forward pass and no per-modality tuning; head weights can be trained where ablation data are available. With fixed causal test setting, the units OmniTrace selects are causally important for responses and for individual sentences in video, audio, and image-text question answering on Qwen3-Omni-30B, MiniCPM-o 4.5 and Qwen2.5-Omni-7B. With one generation pass, OmniTrace recovers 77-96% of the effect of leave-one-out attribution, which needs up to 41 extra passes. OmniTrace also agrees best with human-marked evidence: its response-level rankings reach an AUROC of 0.83 on documents and 0.64 on video, against 0.76 and 0.60 for the strongest baseline.

Figures & tables

Figure 1: OmniTrace reads what a model used from the pass that generates the answer. Left: video, audio, and documents enter one input sequence as source units (seconds, passages, images). While the decoder writes a sentence, OmniTrace sums the attention mass on each unit (bars) and labels the units that stand out within their modality (orange). Top right: every sentence comes with its sources, as time intervals or source ids; here most labelled seconds fall inside the span that annotators marked, although no annotation was used. Bottom right: tested by hiding units and replaying the response, OmniTrace reaches 77–96% of the effect of leave-one-out with no extra forward pass, where perturbation needs 16 to 41 per question (Qwen2.5-Omni-7B).

Method	training	pass	Image	video
	No	No extra		Audio,
Self-citation (prompted)	✓	–	✓	✓
ContextCite [ 8 ]	✓	–	–	–
MIRAGE [ 11 ]	✓	–	–	–
AT2 [ 9 ]	–	✓	–	–
MultAttnAttrib [ 12 ]	✓	✓	✓	–

Table 1: Attribution methods for generated text. Image and audio, video: source types the method is defined and evaluated on beyond text.

	Video	Audio	Image–text
Dataset	NExT-GQA	MMAU	MMDocRAG
Task	multiple choice	multiple choice	long answer
Questions	300	249	400
Source unit	1 s of frames	1 s of audio	passage / image
Units per question	46	22	15
Input length	10–90 s	1.5–35 s	15 quotes

Table 2: Data. ∗ Open-ended answers graded as fully correct by an LLM judge Opus 5 [ 28 ] against the reference answer ( Appendix E ).

	Video				Audio				Image–text
Method	Fwd	Lift	S-lift	S-spec	Fwd	Lift	S-lift	S-spec	Fwd	Lift	S-lift	S-spec
Embedding similarity	0	0.008	0.009	0.005	0	0.034	0.031	0.005	0	0.216	0.216	0.156
Self-citation	1	0.013	0.012	0.009	1	0.040	0.036	0.024	1	0.184	0.349	0.317
Leave-one-out ∗	41	0.016	0.024	0.022	17	0.100	0.133	0.105	16	0.368	0.404	0.286
ContextCite ∗	33	0.016	0.027	0.023	28	0.099	0.139	0.119	33	0.389	0.397	0.285
OmniTrace on Qwen2.5-Omni-7B

Table 3: Causal faithfulness and cost. Fwd: extra forward passes per question. Lift: extra drop in the response’s log-probability when a method’s top units are hidden, over hiding the same number of random units; S-lift and S-spec: sentence lift and specificity. Bold / underline : largest / second largest value of a column. ∗ Defined for text in the original work; run on audio, video, and images with OmniTrace’s units and placeholders. ¶ Head weights fitted on ablation data; no pass added at test time.

Figure 2: Hiding what OmniTrace selects hurts the response more than hiding random units of the same count , in every modality, at every budget, and on every model. Each line is the drop in the log-probability of a model’s own response when a share of its input is hidden: orange when the hidden units are the ones OmniTrace ranks highest for that model, blue when they are drawn at random. Solid lines are Qwen2.5-Omni-7B, dashed lines MiniCPM-o 4.5, and dotted lines Qwen3-Omni-30B, each explaining its own responses with the same OmniTrace setting. Bands are 95% bootstrap intervals.

Figure 3: Faithfulness against cost . Lift of each method as a share of leave-one-out’s lift, against the extra forward passes it needs per question. Both OmniTrace variants sit in the shaded column: they read their scores from the generation pass. OmniTrace-trained needs ablation data from the same model beforehand.

Figure 4: The model’s own citations . Left: share of responses that cite at least one valid source when asked to; OmniTrace attributes every response. Right: lift of the cited units and of OmniTrace’s units of the same count on the same responses.

Figure 5: A response and its attribution . Rows are generated sentences and cells are seconds of the video; darker cells hold more attention mass (scaled per row), and orange boxes are the labelled units. The bar on top is the human-marked evidence: 18 of the 21 labelled cells fall inside it, although no annotation was used.

	Video		Image–text
Ranking	AUROC	Hit@1	AUROC	Hit@1
Random	0.500	0.215	0.500	0.194
Embedding similarity	0.604	0.288	0.763	0.527
Leave-one-out	0.549	0.309	0.721	0.657
ContextCite	0.566	0.342	0.743	0.642
OmniTrace	0.637	0.230	0.829	0.698

Table 4: Agreement of response-level rankings with human-marked evidence. AUROC: probability that an evidence unit is ranked above another unit. Hit@1: the top unit is evidence.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Modality ( $n$ )	Budget	Lift, empty	Lift, mean	Mean $-$ empty (paired)
Qwen2.5-Omni-7B	Video (246)	10%	0.007	0.007	0.000 [ $-$ 0.002, 0.003]
		20%	0.016	0.014	$-$ 0.002 [ $-$ 0.004, 0.000]
		30%	0.017	0.017	0.000 [ $-$ 0.001, 0.001]
		50%	0.027	0.025	$-$ 0.002 [ $-$ 0.008, 0.003]
	Audio (196)	10%	0.056	0.062	0.006 [ $-$ 0.002, 0.013]
		20%	0.069	0.078	0.009 [0.000, 0.018]

Appendix

Table 5: Lift of OmniTrace under two ways of hiding a unit, on the same questions: a content-free placeholder at the input ( empty , used throughout the paper) and replacing the unit’s embeddings by their mean ( mean ).

Video (

n

= 246)

Audio (

n

= 196)

Image–text (

n

= 400)

Method

Fwd

Lift

S-lift

S-spec

Fwd

Lift

S-lift

S-spec

Fwd

Lift

S-lift

S-spec

Baselines on MiniCPM-o 4.5

Embedding similarity

0.030

0.032

0.013

0.067

\downarrow

0.055

\downarrow

0.015

\downarrow

0.248

\uparrow

0.365

\uparrow

0.282

\uparrow

Self-citation †

1 ‡

0.026

\uparrow

0.037

\uparrow

0.034

\uparrow

1 ‡

0.064

\downarrow

0.093

\downarrow

0.114

\downarrow

1 ‡

0.267

\uparrow

0.907

\uparrow

0.903

\uparrow

Leave-one-out ∗

0.037

\uparrow

0.039

\uparrow

0.027

\uparrow

0.189

\uparrow

0.152

\uparrow

0.086

\uparrow

0.451

\uparrow

0.736

\uparrow

0.545

\uparrow

ContextCite ∗

0.050

\uparrow

0.058

\uparrow

0.032

\uparrow

0.193

\uparrow

0.157

\uparrow

0.101

\uparrow

0.453

\uparrow

0.722

\uparrow

0.542

\uparrow

Appendix

Table 6: The comparison of Table 3 repeated on MiniCPM-o 4.5: every baseline and both OmniTrace variants are run on this model. Columns as in Table 3 .

\downarrow

\uparrow

: reliably lower / higher than OmniTrace on this model on the same questions (paired 95% bootstrap interval). † Self-citation is a separate generation ( ‡ one extra pass), scored on the 210, 92, and 227 responses that cite; the document citation run covers 252 of the 400 questions.

Model	Modality ( $n$ )	Budget	Drop, top	Drop, random	Lift	Lift $>0$	Keep gain
Qwen2.5-Omni-7B	Video (246)	10%	0.009	0.002	0.007 [0.005, 0.009]	70%	0.022 [0.006, 0.048]
		20%	0.020	0.004	0.016 [0.011, 0.021]	73%	0.010 [ $-$ 0.002, 0.021]
		30%	0.025	0.009	0.017 [0.013, 0.021]	77%	0.006 [ $-$ 0.004, 0.013]
		50%	0.043	0.016	0.027 [0.021, 0.034]	88%	0.005 [ $-$ 0.000, 0.011]
	Audio (196)	10%	0.112	0.056	0.056 [0.043, 0.069]	72%	0.042 [0.033, 0.053]
		20%	0.144	0.075	0.069 [0.055, 0.086]	75%	0.066 [0.051, 0.080]

Appendix

Table 7: Response-level results of OmniTrace at every budget. Drops are in nats per token. Lift

>0

: share of the questions with a positive lift. Brackets: 95% bootstrap intervals.

			Lift				OmniTrace $-$ method (paired)
Modality	Budget	OmniTrace	Embedding	Leave-one-out	ContextCite	OT-trained	Embedding	Leave-one-out	ContextCite	OT-trained
Video	10%	0.007	0.002	0.014	0.013	0.006	0.005 [0.002, 0.008]	$-$ 0.007 [ $-$ 0.011, $-$ 0.005]	$-$ 0.006 [ $-$ 0.008, $-$ 0.004]	0.001 [ $-$ 0.001, 0.003]
	20%	0.016	0.008	0.016	0.016	0.014	0.008 [0.004, 0.012]	$-$ 0.001 [ $-$ 0.004, 0.004]	$-$ 0.000 [ $-$ 0.005, 0.007]	0.002 [ $-$ 0.003, 0.009]
	30%	0.020	0.012	0.019	0.022	0.018	0.008 [0.004, 0.013]	0.001 [ $-$ 0.003, 0.006]	$-$ 0.002 [ $-$ 0.006, 0.002]	0.002 [ $-$ 0.003, 0.009]
	50%	0.027	0.019	0.020	0.031	0.023	0.008 [0.002, 0.015]	0.007 [0.003, 0.011]	$-$ 0.004 [ $-$ 0.009, 0.000]	0.004 [ $-$ 0.003, 0.013]
Audio	10%	0.060	0.023	0.090	0.087	0.068	0.037 [0.025, 0.050]	$-$ 0.030 [ $-$ 0.038, $-$ 0.023]	$-$ 0.027 [ $-$ 0.035, $-$ 0.019]	$-$ 0.008 [ $-$ 0.013, $-$ 0.003]

Appendix

Table 8: Lift of every method at every budget on Qwen2.5-Omni-7B, and the paired difference between OmniTrace and each baseline with its 95% bootstrap interval. OT-trained: OmniTrace-trained ( Appendix A ).

Model	Modality ( $n$ )	Units labelled	No label	Lift	Keep gain	S-lift	S-spec
Qwen2.5-Omni-7B	Video (244)	4.0 (10%)	2%	0.006 [0.003, 0.010]	0.046 [0.011, 0.111]	0.012 [0.009, 0.014]	0.007 [0.004, 0.010]
	Audio (196)	2.6 (16%)	14%	0.079 [0.060, 0.098]	0.077 [0.060, 0.096]	0.125 [0.107, 0.143]	0.106 [0.082, 0.130]
	Image–text (400)	3.1 (21%)	0%	0.364 [0.329, 0.403]	0.393 [0.353, 0.438]	0.411 [0.387, 0.434]	0.240 [0.214, 0.267]
MiniCPM-o 4.5	Video (233)	5.2 (13%)	0%	0.006 [0.002, 0.010]	0.006 [ $-$ 0.004, 0.015]	0.014 [0.008, 0.020]	0.008 [0.001, 0.015]
	Audio (187)	1.6 (15%)	0%	0.073 [0.047, 0.100]	0.042 [0.024, 0.063]	0.080 [0.059, 0.100]	0.047 [0.011, 0.085]
	Image–text (400)	3.1 (21%)	0%	0.170 [0.140, 0.201]	0.149 [0.124, 0.178]	0.458 [0.437, 0.479]	0.344 [0.320, 0.369]

Appendix

Table 9: Causal test of the labels of Equation 3 at

\tau=1.5

. Units labelled: mean number per response and share of the candidate units. No label: responses whose response-level label is empty. Lift and keep gain compare the labelled units with random units of the same count.

Qwen2.5-Omni-7B ( $n$ = 246, 196, 400)
	Video				Audio				Image–text
	Lift	Keep	S-lift	S-spec	Lift	Keep	S-lift	S-spec	Lift	Keep	S-lift	S-spec
OmniTrace	0.016	0.010	0.015	0.007	0.069	0.066	0.103	0.084	0.348	0.412	0.375	0.245
argmax tracing + curation	0.009 $\downarrow$	0.008	0.010 $\downarrow$	0.004 $\downarrow$	0.067	0.064	0.096 $\downarrow$	0.081	0.346	0.420	–	–
last layer only	0.014	0.010	–	–	0.048 $\downarrow$	0.046 $\downarrow$	–	–	0.306 $\downarrow$	0.380 $\downarrow$	–	–
per-token mean	–	–	–	–	–	–	–	–	0.224 $\downarrow$	0.255 $\downarrow$	0.286 $\downarrow$	0.214 $\downarrow$

Appendix

Table 10: OmniTrace on three models, variants that change one design choice, and human-marked evidence. Lift and Keep: lift and keep gain at response level (20% of the duration; 2 quotes); S-lift and S-spec: sentence lift and specificity, in nats per token.

\downarrow

\uparrow

: reliably lower / higher than OmniTrace on the same questions or sentences. Argmax tracing + curation traces each token to its highest-attention position and filters the votes; last layer only and per-token mean change one choice in Equation 2 . Human-evidence units hides the units annotators marked (not available for audio); their number differs from the budget, so their lift is against random sets of their own size.

Figure 6: Lift of argmax tracing with curation as a share of OmniTrace’s lift. In each panel the groups of bars are, from left to right, Qwen2.5-Omni-7B, on which the variant was developed, MiniCPM-o 4.5, and Qwen3-Omni-30B.

Ranking	$\rho$ (score, size)	Top-1 size	Image share
OmniTrace (sum)	$+$ 0.51	1.94	0.59
Per-token mean	$-$ 0.60	0.44	0.11
Leave-one-out	$+$ 0.09	1.62	0.48
ContextCite	$+$ 0.14	1.65	0.52
OmniTrace-trained	$+$ 0.08	1.71	0.52
Human evidence	$+$ 0.25	1.85	0.67

Appendix

Table 11: Unit size on image–text input (Qwen2.5-Omni-7B).

\rho

: rank correlation between a quote’s score and its token count, per question. Top-1 size: tokens of the top quote over the mean quote size. Image share: share of images among the top-ranked quotes.

Multimodal Model Interpretability

Multimodal model interpretability studies how AI systems combine and use images, text, audio, video, and other signals, and how to explain their decisions.

See all 129 papers

Attending to Multimodal Generation One Token at a Time
Jul 4, 2026Varun Gupta, Vineet Gandhi, Makarand TapaswiMultimodal Large Language Models Multimodal Generation
CVIT, IIIT Hyderabad
Tracing the Evidence: Faithful Token Attribution Through Vision-Language Reasoning
Sep 29, 2026Bowen Yuan, Danny Wang, Ruihong Qiu +2Large Vision-Language Models VLM Interpretability
The University of Queensland
MultAttnAttrib: Training-Free Multimodal Attribution in Long Document Question Answering
Jul 1, 2026Dang Quang Thien Tran, Quang V. Dang, Vinamra Tyagi +7Multimodal QA Document QA
University of Massachusetts, Amherst · 2Adobe Research, San Jose

Jul 4, 2026J/K move · Enter open · S save

Attending to Multimodal Generation One Token at a Time

Varun Gupta, Vineet Gandhi, Makarand Tapaswi

CVIT, IIIT Hyderabad

Multimodal Large Language Models Multimodal Generation Multimodal Foundation Models Multimodal Model Interpretability Multimodal Understanding Modality Reliance in Multimodal Models

Open paper AskPDF arXiv

Multimodal large language models (MLLMs) generate responses autoregressively, integrating visual and linguistic information in an evolving context. Prior work on interpretability has focused on individual layers and circuits (where), leaving the token-level dynamics of multimodal computation during generation (when) underexplored. We address this gap and study attention shifts as per semantic role; tracking model attention to image, text, instruction, and previously generated tokens, One Token at a Time (OTaT). We introduce multimodal tasks that require explicit switching between visual and textual context within a single response. Across two mainstream model families and four open-weight MLLMs of varying sizes, we establish consistent patterns: attention to image peaks at tokens requiring image-derived information, instruction tokens are revisited during task transitions, and attention to previously generated tokens increases as the generation progresses. Causal attention blocking interventions validate the functional role of these trends. We profile model behavior under disrupted attention and observe responses falling back to language priors, or exhibiting cross-modal leakage, denial, or recovery. Finally, informed of the attention dynamics through our novel analysis, we propose a simple test-time intervention to boost attention to the relevant modality at the right time, significantly improving multimodal task performance.

Abstract

Multimodal Model Interpretability

Explore similar work

Attending to Multimodal Generation One Token at a Time

Tracing the Evidence: Faithful Token Attribution Through Vision-Language Reasoning

MultAttnAttrib: Training-Free Multimodal Attribution in Long Document Question Answering