CGCMA: Conditionally-Gated Cross-Modal Attention for Event-Conditioned Asynchronous Fusion

Yunxiang Guo

cs.LGApr 1, 2026

CGCMA: Conditionally-Gated Cross-Modal Attention for Event-Conditioned Asynchronous Fusion

Authors: Yunxiang Guo

Organizations: Independent Researcher, China

Abstract

We study asynchronous alignment, a first-class multimodal learning setting in which a dense primary stream must be fused with sporadic external context whose value depends on when it arrives. Unlike standard multimodal benchmarks that assume structural synchrony, this setting requires models to reason explicitly about freshness and trust. We focus on the event-conditioned case in which continuous market states are paired with delayed web intelligence, and we use high-frequency cryptocurrency markets only as a timestamped, high-noise stress test for this broader problem. We propose CGCMA (Conditionally-Gated Cross-Modal Attention), whose central design principle is to separate text-conditioned grounding from lag-aware trust control. Text first attends over price sequences to identify event-relevant market states, after which a conditional gate uses modality agreement, web features, and lag $τ_{\mathrm{lag}}$ to regulate residual injection and fall back toward unimodal prediction when external context is stale or contradictory. We introduce CMI (Crypto Market Intelligence), an asynchronous evaluation corpus with 27,914 real-news samples pairing high-frequency price sequences with lagged web intelligence. On the current short real-news corpus, CGCMA attains the highest mean downstream Sharpe ratio ( $+0.449 \pm 0.257$ ) among the evaluated baselines under a shared zero-cost threshold-trading evaluation on news-available bars. Additional controls show that the gain is not explained by web scalars alone and is not recovered by simple freshness heuristics. The resulting evidence supports problem validity and a promising asynchronous multimodal gain on this stress-test setting.

Figures & tables

Dataset	Text source	Price granularity	Assets	Samples	Lag $\tau$ recorded	Rolling+SR
PreBit ( Zou and Herremans, 2023 )	Real (Twitter)	Daily	BTC	${\sim}$ 1K	×	×
DAM ( Fu et al., 2024 )	Real (CryptoCompare)	Daily	BTC	${\sim}$ 1K	×	×
CryptoPulse ( Kumar and Ji, 2025 )	Real (Cointelegraph)	Daily	Top-5	${\sim}$ 1K	×	×
Barthélemy et al. ( Gurgul et al., 2025 )	Real (Twitter/Reddit)	Daily	BTC/ETH	${\sim}$ 3yr	×	✓
CryptoMI (synthetic)	LLM-synthesized	15 min	BTC/ETH/SOL	21,374	✓	✓
CryptoMI (real news)	Real (CryptoCompare)	15 min	BTC/ETH/SOL	27,914	✓	✓

Table 1. Comparison of cryptocurrency multimodal datasets. Lag

\tau

indicates whether per-sample modality lag is recorded; Rolling+SR indicates rolling walk-forward evaluation with Sharpe reporting. LLM synthesis trades source authenticity for temporal control and sub-hourly coverage; the associated limitations are discussed in Section 7 .

Figure 1. CGCMA architecture. The proposed fusion is explicitly split into two roles: text-conditioned price attention grounds the aligned text in the full price sequence to form $\mathbf{h}_{c}=\mathrm{LN}(\mathrm{MHA}(\mathbf{h}_{t},\mathbf{H}^{p},\mathbf{H}^{p}))$ , while the conditional gate controls trust in that context using the context shift $\mathbf{\Delta}_{pc}=\mathbf{h}_{p}-\mathbf{h}_{c}$ , web context, and freshness $\tau/60$ . This yields a causally aligned multimodal residual update $\mathbf{h}_{f}=\mathbf{h}_{p}+\mathbf{g}\odot\mathbf{h}_{c}$ that falls back to the price-only path when text is stale or uninformative, while retaining $\mathbf{h}_{w}$ as a direct predictive cue at the task head. The main experiments instantiate the task head as binary direction classification, but the fusion design itself is task-agnostic. Architecture of CGCMA. All inputs are causally aligned at decision time t: the OHLCV window ends at t, the aligned text context satisfies t star less than or equal to t, the web scalars are available at t, and the freshness variable is tau equals t minus t star. A price encoder outputs the full price token sequence H superscript p and a price summary h sub p. A text encoder outputs h sub t, and a web scalar encoder outputs h sub w. Text-conditioned price attention uses the text token as query and the price token sequence as keys and values, followed by layer normalization, to retrieve an event-conditioned market context h sub c. A conditional gate receives h sub p, h sub c, their context-shift term delta sub pc equals h sub p minus h sub c, h sub w, and tau divided by 60, and outputs a per-dimension gate vector g. A gated residual fusion block computes h sub f equals h sub p plus g elementwise-multiplied by h sub c, preserving a price-only fallback path when the gate closes. The fused state is concatenated with a direct predictive web cue and passed to a task head, shown here as binary direction prediction.

Corpus	Asset	Samples	Lag (min)	Pos. rate
Synthetic	BTC	7,335	62.3	$>$ 99%
	ETH	7,039	64.1	$>$ 99%
	SOL	7,000	63.8	$>$ 99%
Real news	BTC	19,241	5.9	45%
	ETH	6,349	6.1	45%
	SOL	2,333	5.7	44%

Table 2. CryptoMI statistics. Synthetic denotes the LLM-synthesized corpus from Jan–Dec 2025; Real news denotes the CryptoCompare corpus from Dec 2025–Mar 2026. We report sample count, mean lag, and positivity rate.

Model	Inputs	Fusion	Mean $\uparrow$	Std $\downarrow$	95% CI	$\Delta$ vs. Price	$\Delta$ vs. BiLSTM	Hit rate $\uparrow$
Unimodal baselines
PriceTx-S	P	Transformer + CLS	$+0.194$	$0.229$	$[0.003,\,0.385]$	—	$-0.015$	0.518
TextOnly	T	SBERT + MLP	$-0.043$	$0.168$	$[-0.183,\,0.098]$	$-0.237$	$-0.252$	0.495
Multimodal baselines
PriceWeb	P+W	Price+web late fusion	$+0.161$	$0.289$	$[-0.299,\,0.621]$	$+0.060$	$-0.099$	0.532
EarlyFusion	P+T+W	Early concat	$+0.198$	$0.150$	$[0.072,\,0.324]$	$+0.004$	$-0.011$	0.525

Table 3. Main real-news comparison on CryptoMI (BTC+ETH+SOL pooled; 27,914 samples; non-overlapping rolling evaluation;

{\approx}35

folds). Entries are mean

\pm

std across 8 runs . Best mean is bold ; second best is underlined .

Model	AUC $\uparrow$	Brier $\downarrow$	Hit rate $\uparrow$
PriceTx-S	0.537	0.261	0.518
PriceWeb	0.536	0.259	0.532
BiLSTM	0.528	0.263	0.521
CGCMA	0.536	0.265	0.535

Table 4. Auxiliary predictive metrics for selected models from Table 3 . Entries are 8-run means on the same pooled real-news protocol, except PriceWeb , which remains a 4-run control.

Model	Key design choice	Mean Sharpe	$\Delta$ vs. PriceTx-S
PriceTx-S	price-only Transformer	$+0.101{\pm}0.197$	—
FixedDecay	fixed exponential decay	$+0.219{\pm}0.382$	$+0.118$
LearnedDecay	learned scalar decay	$+0.014{\pm}0.203$	$-0.087$
HardFilter	hard stale-text filter	$+0.272{\pm}0.408$	$+0.171$
CGCMA-db	conditional late-fusion predecessor	$+0.244{\pm}0.185$	$+0.143$
CGCMA	grounding + vector gate	$+0.375{\pm}0.270$	$+0.274$

Table 5. Dedicated four-seed sweep of task-specific lag-aware controls on the pooled real-news protocol. Entries are 4-run mean

\pm

std Sharpe from a run batch distinct from Table 3 ; compare rows within this table only.

Figure 2. Web-intelligence directional signal Sharpe by modality lag (

\tau_{\mathrm{lag}}

) on the bar-aligned BTC dataset (7,295 samples,

{\approx}440

per bin). The 30–60 min post-publication window provides the strongest positive signal. Very fresh news (

<

30 min) is near-zero, while very stale news (

>

90 min) is strongly negative, indicating a non-monotonic empirical freshness pattern. Bar chart showing annualised signal Sharpe by modality lag bin for bar-aligned BTC data. The 0-to-30-minute bin is slightly negative (-0.30). The 30-to-60-minute bin is the highest (+0.57). The 60-to-90-minute bin is lower but still positive (+0.28). The 90-to-120-minute bin is strongly negative (-1.33). Each bin contains approximately 440 samples.

Model	Mean Sharpe	Std.
PriceTx-S ( $d{=}32$ )	$\mathbf{+0.124}$	$\pm 0.056$
CGCMA-lg (MiniLM-384 + Mixup)	$+0.021$	$\pm 0.052$

Table 6. Full-year scaling study (v5) on BTC (7,335 event-aligned samples; 63 rolling folds;

{\approx}500

training samples per fold). Entries are mean

\pm

std across 4 independent runs . CGCMA-lg is the strongest multimodal variant on the synthetic split.

Multimodal Fusion

Multimodal fusion combines information from different data types to improve prediction, perception, and decision-making.

See all 768 papers

May 27, 2026

GS-FUSE: Granger-Supervised Gated Fusion and Multi-Granularity Alignment for Event-Driven Financial Forecasting

Accurately forecasting the impact of salient financial events on markets is critical for investors and policymakers. However, existing multimodal time-series models typically fuse text and prices symmetrically, without an explicit way to decide when event text is truly predictive, and thus struggle to exploit the directional event-to-price structure and the heterogeneous roles of textual and price signals. In this work, we propose GS-Fuse, a multimodal event-based forecasting framework that employs (i) a Granger-supervised, causal-aware gated fusion module, which learns to open toward event text only when it provides incremental predictive value beyond historical prices, and (ii) a multi-granularity alignment mechanism that jointly aligns high-level event representations and fine-grained textual cues with future market trajectories. Built as a flexible, plug-and-play adapter on top of off-the-shelf large language models and time-series foundation models, GS-Fuse can be instantiated across diverse backbones and market settings. Extensive experiments on real-world financial datasets show that GS-Fuse consistently outperforms state-of-the-art time-series and multimodal baselines across multiple assets and forecasting horizons.

Yang Zhang, En Chun, Ziyun Mao +2

Southwestern University of Finance and Economics Chengdu, China

Financial Forecasting Time Series Forecasting Financial Time Series Multimodal Fusion

Ask PDF

Sep 15, 2026

AsyncCouple-Flow: Asynchronous Cross-Modal Coupling and Flow Matching for Spatio-Temporal Forecasting

Multi-modal spatio-temporal forecasting (MM-STF) supports weather nowcasting, traffic prediction, and earth-system modeling by combining heterogeneous sources such as physical fields, satellite imagery, and in-situ sensors. Three obstacles persist: (i) modalities have different spatio-temporal sampling rates, forcing lossy interpolation onto a unified grid; (ii) modalities are frequently missing at deployment due to sensor outages or revisit gaps, while most methods train with full availability; and (iii) autoregressive decoders accumulate errors over long horizons, amplified by multi-modal conditioning. We propose AsyncCouple-Flow to address these issues jointly. A Modality-Aware Token Sparsification (MATS) module performs scale-aware tokenization and uses a shared importance scorer to select top-k tokens per timestep, producing equal-length sequences. An Asynchronous Cross-Modal Coupling Graph (ACCG) replaces fixed cross-attention with a learnable graph whose edges encode time offsets, semantic similarity, and modality-specific physical priors, enabling fusion under arbitrary asynchrony and missingness. A Flow-Matching Forecasting Head models multi-step prediction as a conditional ODE, trained with stochastic modality dropout and integrated jointly to avoid autoregressive drift. Experiments on ERA5+GOES+ISD weather forecasting and PEMS-BAY traffic prediction with multi-source side information show that AsyncCouple-Flow outperforms state-of-the-art baselines and remains robust with up to two missing modalities. The code will be released upon acceptance.

Zhixiang Wu, Yining Liu, Bo Zhao +4

Institute of Computing Technology, Chinese Academy of Sciences, China · Emory University, USA · University of California, Berkeley, USA +3

Missing-Modality Learning Spatiotemporal Forecasting Multimodal Learning Multimodal Sensor Fusion

Ask PDF

Mar 27, 2026

Context-specific Credibility-aware Multimodal Fusion with Conditional Probabilistic Circuits

Multimodal fusion requires integrating information from multiple sources that may conflict depending on context. Existing fusion approaches typically rely on static assumptions about source reliability, limiting their ability to resolve conflicts when a modality becomes unreliable due to situational factors such as sensor degradation or class-specific corruption. We introduce C

^2

MF, a context-specfic credibility-aware multimodal fusion framework that models per-instance source reliability using a Conditional Probabilistic Circuit (CPC). We formalize instance-level reliability through Context-Specific Information Credibility (CSIC), a KL-divergence-based measure computed exactly from the CPC. CSIC generalizes conventional static credibility estimates as a special case, enabling principled and adaptive reliability assessment. To evaluate robustness under cross-modal conflicts, we propose the Conflict benchmark, in which class-specific corruptions deliberately induce discrepancies between different modalities. Experimental results show that C

^2

MF improves predictive accuracy by up to 29% over static-reliability baselines in high-noise settings, while preserving the interpretability advantages of probabilistic circuit-based fusion.

Pranuthi Tenali, Sahil Sidheekh, Saurabh Mathur +3

†The University of Texas at Dallas · §Department of Computer Science , TU Darmstadt · ‡Air Force Research Lab

Multimodal Robustness Uncertainty-Aware Multimodal Fusion Multimodal Fusion Reliability-Aware Multimodal Fusion

Ask PDF

Abstract

Multimodal Fusion

Explore similar work

GS-FUSE: Granger-Supervised Gated Fusion and Multi-Granularity Alignment for Event-Driven Financial Forecasting

AsyncCouple-Flow: Asynchronous Cross-Modal Coupling and Flow Matching for Spatio-Temporal Forecasting

Context-specific Credibility-aware Multimodal Fusion with Conditional Probabilistic Circuits