---
title: 'Falcon-X: X-ray Benchmark & Time Series Model'
url: https://www.emergentmind.com/topics/falcon-x
type: topic
---

# Falcon-X: X-ray Benchmark & Time Series Model

Falcon-X is not a single universally defined system in the arXiv record. The name is used for at least two distinct research artifacts: a benchmark for compositional threat reasoning in dual-energy X-ray baggage screening, introduced alongside the Falcon model for structured safety reasoning, and a time series foundation model for heterogeneous multivariate forecasting that operates in a shared latent prototype space [2606.25701][2605.27286]. The term therefore denotes domain-specific constructs rather than a unitary architecture, and its meaning depends on the surrounding research context.

## 1. Nomenclature and scope

In current arXiv usage, the name has two principal referents. One is **Falcon-X** as a **benchmark** in safety-critical X-ray vision-language research; the other is **Falcon-X** as a **time series foundation model** in forecasting research. The shared label does not imply a common lineage, objective, or implementation substrate [2606.25701][2605.27286].

| Falcon-X usage | Domain | Role |
|---|---|---|
| Falcon-X | X-ray baggage screening | Benchmark for compositional threat reasoning |
| Falcon-X | Time series forecasting | Foundation model for heterogeneous multivariate modeling |

This bifurcation is central to interpreting the literature. In the X-ray setting, Falcon-X is an evaluation suite and supervision framework paired with a separate model named **Falcon**. In the forecasting setting, Falcon-X is itself the model. A plausible implication is that references to “Falcon-X” without domain qualifiers are intrinsically ambiguous.

## 2. Falcon-X in X-ray compositional threat reasoning

In **“Falcon: Functional Assembly and Language for Compositional Reasoning in X-ray”**, Falcon-X is a **dual-energy X-ray benchmark** for evaluating whether a multimodal system can infer threat from the **functional assembly potential** of spatially dispersed components such as a **battery**, **detonator**, and **main charge**, rather than merely recognize isolated objects [2606.25701]. The underlying problem is formalized as **compositional threat reasoning**, where risk is a relational property over grounded regions.

The benchmark instantiates a component taxonomy
$$
\mathcal{C} = \{\text{battery}, \text{detonator}, \text{main charge}\},
$$
and defines a binary component-presence vector
$$
\mathbf{y} \in \{0,1\}^{|\mathcal{C}|}, \qquad y_c = 1 \iff \text{at least one instance of component } c \text{ is present}.
$$
It also defines a type-level compatibility template
$$
\mathbf{L} \in [0,1]^{|\mathcal{C}| \times |\mathcal{C}|},
$$
where \(L_{uv}\) encodes the functional compatibility strength between component types \(u\) and \(v\). The pair \((\mathbf{y}, \mathbf{L})\) is the benchmark’s basic **structured safety state**.

Falcon-X includes a minimal completeness notion,
$$
\text{complete}(I) = \mathbb{1}\{\mathbf{y} = \mathbf{1}\},
$$
but the paper explicitly distinguishes completeness from risk. Scene-level risk is modeled as a continuous variable
$$
r \in [0,1],
$$
and is **not trivially derived** from the visible presence vector. Even scenes missing one component may receive a high risk score because a part may be occluded, concealed, or ambiguously visible. This makes Falcon-X a benchmark for **uncertainty-aware compositional safety reasoning**, not deterministic set completion.

The benchmark is explicitly positioned against object-centric evaluation. Standard image captioning, object detection, simple VQA, and referring segmentation can all succeed while still failing the targeted safety problem: inferring whether dispersed grounded parts are jointly sufficient and functionally compatible for a plausible assembly. Falcon-X is therefore presented as the **first X-ray benchmark** that jointly supports dense grounding, multimodal understanding, and structured functional threat reasoning over dismantled components.

## 3. Benchmark design, supervision, and evaluation in the X-ray setting

Falcon-X combines **dense instance grounding** with **structured functional labels** and **multimodal task supervision** [2606.25701]. Each image includes bounding boxes and pixel masks, together with component presence, functional completeness, pairwise functional compatibility links, and a scene-level risk score. The associated task suite spans captioning, VQA, referring segmentation, panoptic tasks, functional grounding, missing-component reasoning, and risk prediction.

The dataset statistics are reported with a small unresolved discrepancy. The main paper says Falcon-X contains about **7,000 real dual-energy baggage scans**, the appendix states that **7,000 base X-ray scans** were collected, and the benchmark table reports **6,911** samples. The counterfactual extension expands the corpus to approximately **50,000 images**. Stage 1 uses **16,580 annotated component instances**, while Stages 2 and 3 use **442,287 training instructions**. Data collection used an **ANER K8065 dual-energy X-ray scanner** with settings **100–160 kV, 0.4–1.2 mA**. The documented split is **80/20 train/test**, with all counterfactuals from a base image kept in the same partition; no separate validation split is specified.

A notable feature is the **counterfactual extension**. Components are selectively removed using **mask-guided inpainting**, producing controlled variants spanning single components, partial assemblies, and complete configurations. This supports evaluation of missing-component identification, partial-assembly reasoning, functional completeness, and risk recalibration. Risk labels are generated from component presence, type-level compatibility, and visual uncertainty, then expert-verified using a rubric in which \([0.0,0.3)\) is low risk, \([0.3,0.6)\) is medium or ambiguous risk, and \([0.6,1.0]\) is high functional risk; **72.3%** of generated labels were accepted without change and **27.7%** were corrected by experts.

Falcon-X organizes evaluation into three layers. **Layer I** covers grounded perception under X-ray superposition, including scene captions, panoptic segmentation, referring segmentation, and VQA. **Layer II** covers compositional functional reasoning, including Missing Component Identification, Functional Completeness, and Referring Functional Grounding. **Layer III** covers relationally consistent safety inference, including scene-level risk prediction, functional link estimation, and logically coherent component-set analysis. Metrics vary by task type: **BLEU**, **METEOR**, **ROUGE-L**, and **CIDEr** for language; **cIoU** and **mIoU** for grounding and segmentation; exact-match accuracy, MAE, and binary accuracy for VQA; **Accuracy** and **F1** for classification-style reasoning; and **MAE** or **RMSE** for functional completeness, scene risk, and link-risk regression.

Empirically, the benchmark exposes a clear gap between perceptual adaptation and compositional reasoning. After fine-tuning, many baselines become strong on captioning, VQA, and presence recognition, but remain weak on missing-part reasoning, pairwise compatibility, scene-level risk, and functionally defined grounding. On **Referring Functional Grounding**, Falcon exceeds the next best model by **+30.38 cIoU** and **+40.03 mIoU**. The structured metrics reported for Falcon are **CPC Acc = 98.1**, **FC MAE = 0.017**, **FC RMSE = 0.09**, **SRL MAE = 0.02**, and **CLR MAE = 0.005**. The benchmark’s diagnostic value lies precisely in showing that high object presence accuracy does not imply coherent link estimation or calibrated threat inference.

## 4. Falcon-X as a time series foundation model

In **“Falcon-X: A Time Series Foundation Model for Heterogeneous Multivariate Modeling”**, Falcon-X is an **encoder-only Transformer TSFM** for **heterogeneous multivariate forecasting** [2605.27286]. Its core claim is that prior TSFMs remain either effectively univariate or perform cross-variate interaction too directly in the **raw variate space**, where heterogeneous channels from different datasets are not semantically aligned and where standard non-negative attention cannot explicitly represent antagonistic dependencies.

The formal setup is
$$
\mathcal{E}=\{\mathbf{e}_i\in\mathbb{R}^{m_i\times L}\}_{i=1}^N,
$$
where entity \(\mathbf{e}_i\) is a multivariate time series with \(m_i\) variates and look-back window \(L\), and the total aggregated variate count is
$$
M = \sum_{i=1}^N m_i.
$$
Falcon-X learns a dimension-agnostic mapping
$$
\hat{\mathbf{Y}} = \mathcal{F}_\theta(\mathbf{X}), \quad \text{where } \mathbf{X} \in \mathbb{R}^{M \times L},\; \hat{\mathbf{Y}} \in \mathbb{R}^{M \times T}.
$$

Its defining move is to **decouple physical variates from the interaction space**. Rather than mixing raw channels directly, Falcon-X projects each variate into a small shared set of **latent prototypes**. This is intended to provide a semantic coordinate system in which variates with different physical meanings can nonetheless align to reusable structural patterns. The paper argues that this yields both **semantic alignment** and **zero-shot structural transfer**.

The largest model has about **591M parameters**. It first builds per-variate temporal representations, then applies **Unified Prototype Diff-Attention**, which explicitly computes both positive/synergistic and negative/antagonistic semantic affinities. Cross-variate interaction is then performed in the shared prototype space by **Latent Entity Attention**, after which a **Variate Reassembly Router** reconstructs variate-specific representations for forecasting. This produces a multivariate TSFM intended to be more expressive and more scalable than raw-space multivariate attention, with prototype-space interaction reducing complexity from \(\mathcal{O}(M^2)\) to approximately \(\mathcal{O}(M \cdot C)\) when the prototype count \(C \ll M\).

## 5. Architecture, training, and empirical performance in forecasting

Falcon-X formulates forecasting as a masked reconstruction problem over
$$
\mathbf{X} \in \mathbb{R}^{M \times (L+T)}.
$$
After instance-wise normalization, it applies
$$
\mathbf{\hat{X}} = \arcsin\left(\frac{\mathbf{X} - \mu}{\sigma}\right),
$$
where \(\mu\) and \(\sigma\) are computed only from observed values. The model augments the input with a relative timestamp sequence
$$
\mathcal{T}=\left\{-\frac{L}{L+T}, \dots, 0, \dots, \frac{T-1}{L+T}\right\}
$$
and a binary observation mask \(\mathcal{M}\), then tokenizes the triplet with residual patch embedding:
$$
\mathbf{H} = \text{ResPatchEmbed}(\text{Concat}(\mathbf{\hat{X}}, \mathcal{T}, \mathcal{M}))\in \mathbb{R}^{M \times P \times D}.
$$

Temporal encoding is performed independently along time for each variate:
$$
\mathbf{H}^{(i)} = \text{LayerNorm}\left(\mathbf{H}^{(i-1)} + \text{MHA}\left(\mathbf{H}^{(i-1)}\right)\right),
$$
yielding \(\mathbf{H}_T\). Falcon-X then applies **Unified Prototype Diff-Attention** using two learnable prototype banks \(\mathbf{K}_{\text{pos}}, \mathbf{K}_{\text{neg}} \in \mathbb{R}^{C \times D}\). For an entity \(e_i\), the positive and negative affinity maps are
$$
\mathbf{A}_\text{pos}^i = \text{softmax}\left(\frac{\mathbf{Q}^i\mathbf{K}_\text{pos}^\top}{\sqrt{D}}\right), \quad
\mathbf{A}_\text{neg}^i = \text{softmax}\left(\frac{\mathbf{Q}^i\mathbf{K}_\text{neg}^\top}{\sqrt{D}}\right),
$$
and the prototype representation is
$$
\mathbf{h}_C^i = \left[\mathbf{A}_\text{pos}^i - \lambda \cdot \mathbf{A}_\text{neg}^i\right]^\top \mathbf{V}^i.
$$
After **Latent Entity Attention**, the model reconstructs variate-specific structure through the **Variate Reassembly Router**:
$$
\mathbf{h}_V^i =
\text{softmax}\left(\frac{\mathbf{R}_\text{req}^i (\mathbf{P}_\text{idx}^i)^\top}{\sqrt{D}}\right)\mathbf{S}_\text{ctx}^i.
$$
A gated residual fusion then combines temporal and routed cross-variate information:
$$
\mathbf{\hat{H}} = \mathbf{H}_T + \mathcal{G}(\mathbf{H}_T)\odot \mathbf{H}_V.
$$

The forecasting objective is probabilistic and quantile-based:
$$
\mathcal{L} = \mathcal{L}_\text{pred} + \alpha \mathcal{L}_\text{orth},
$$
where \(\mathcal{L}_\text{pred}\) is the quantile loss and \(\mathcal{L}_\text{orth}\) penalizes overlap between positive and negative prototype banks. Predictions are mapped back to physical scale via
$$
\mathbf{\tilde{Y}^{(q)}} = \sigma \cdot \sin(\mathbf{\hat{Y}^{(q)}}) + \mu.
$$

Pretraining mixes real and synthetic data from **GIFT-Eval**, the **Chronos** training corpus, **QuitoBench**, **TSMixup**, **KernelSynth**, and synthetic multivariate datasets built through grouping and dependency injection. The largest configuration uses **591M parameters**, hidden size \(D=1024\), patch length \(L_p=16\), **16** Time Attention layers, **16** Entity Attention layers, **16 heads per layer**, maximum context length \(L=8192\), and maximum prediction length \(T=480\). Training uses **Megatron-LM**, **1,000,000 iterations**, global batch size **384**, **bf16**, and **AdamW** with \(\beta_1 = 0.9\), \(\beta_2 = 0.95\), weight decay \(0.1\), warmup to \(6\times10^{-5}\), and cosine decay to \(6\times10^{-6}\).

On **GIFT-Eval**, Falcon-X reports **0.666 MASE** and **0.453 CRPS**, improving on **STRIDE** by **1.2%** in MASE, on **Toto-2.0-FT** by **1.9%** in MASE and **2.2%** in CRPS, and on **Timer-S1** by **3.9%** in MASE and **6.6%** in CRPS. Grouped by horizon, it reports **0.65 MASE** in the short term, **0.68** in the medium term, and **0.70** in the long term. On **fev-bench**, Falcon-X reports **0.652 MASE** and **0.490 CRPS**, slightly behind **Chronos-2** at **0.645 MASE** and **0.485 CRPS**. Ablations identify the removal of \(K_\text{neg}\) as the largest drop, support the gated residual, and place the best prototype count around **\(C=6\) or \(C=8\)**. The paper also notes several limitations: sensitivity to the prototype dimension, the need to control cross-variate fusion when dependencies are weak, partial reliance on synthetic multivariate construction, and the fact that the “signed dependence” story remains more architectural than directly interpretable.

## 6. Misidentification within the broader Falcon/FALCON literature

A persistent source of confusion is that many arXiv papers titled **Falcon** or **FALCON** do **not** define Falcon-X. **“Falcon: Fair Active Learning using Multi-armed Bandits”** presents a fair active learning framework and explicitly states that the paper never mentions a variant called “Falcon-X” [2401.12722]. **“Falcon: A Cross-Modal Evaluation Dataset for Comprehensive Safety Perception”** introduces a multimodal safety benchmark named **Falcon** and an evaluator named **FalconEye**, again without a Falcon-X variant [2509.23783]. **“Falcon: A Remote Sensing Vision-Language Foundation Model”** consistently uses **Falcon** for the model and **Falcon\_SFT** for the instruction-tuning dataset, and does not introduce Falcon-X [2503.11070].

This suggests that **Falcon-X** is best treated as a context-dependent label rather than a canonical member of a single named family. In the X-ray literature, Falcon-X is a benchmark paired with a separate model named Falcon. In time-series forecasting, Falcon-X is the model itself. In several other literatures—fair active learning, multimodal safety evaluation, and remote sensing vision-language modeling—the suffix does not appear at all. For rigorous citation and system identification, the surrounding domain and arXiv id are therefore indispensable.

Source: https://www.emergentmind.com/topics/falcon-x