Drax in VideoQA, ASR & Astrophysics

• Drax is a polysemous term referring to a VideoQA method (DRAX), a non-autoregressive ASR framework based on discrete flow matching, and an informal reference to Draco in astrophysics.
• In VideoQA, DRAX employs distraction removal and attended cross-alignment to clean latent cross-modal embeddings, leading to improved benchmark performance.
• In ASR and astrophysics, Drax leverages an audio-conditioned training path for parallel decoding and informs dynamical inference of dark matter, respectively.

Searching arXiv for “Drax” and closely related entries to ground the article in the cited literature. “Drax” is not a single canonical technical term in the arXiv literature. In current usage it denotes at least three unrelated objects: DRAX, a video question answering method whose name expands to Distraction Removal and Attended Cross-Alignment; Drax, a non-autoregressive automatic speech recognition framework based on discrete flow matching; and, in some astrophysical discussion, an informal shorthand for Draco, the Milky Way dwarf spheroidal galaxy analyzed in dark-matter dynamical studies. The term is therefore polysemous across vision-language learning, speech modeling, and galactic dynamics, and any rigorous use requires domain-specific disambiguation (Dogra et al., 2023, Navon et al., 5 Oct 2025, Yang et al., 3 Jul 2025).

1. Nomenclature and scope

Usage	Domain	Defining characterization
DRAX	VideoQA / VLU	Distraction Removal and Attended Cross-Alignment
Drax	ASR	Discrete flow matching framework for non-autoregressive speech recognition
“Drax” = Draco	Astrophysics	Informal reference to the Draco dwarf spheroidal galaxy

In the vision-language setting, the term is uppercase and explicitly acronymic: DRAX addresses distractors in latent cross-modal representations during VideoQA. In speech recognition, Drax is a proper model name rather than an acronym in the supplied material, and denotes a probability-path design plus training and inference framework for non-autoregressive decoding. In astrophysics, “Drax” is not a formal object name; it is treated explicitly as an informal reference to Draco in the discussion of dwarf-spheroidal dark-matter inference (Dogra et al., 2023, Navon et al., 5 Oct 2025, Yang et al., 3 Jul 2025).

A common source of confusion is that these usages are methodologically and scientifically unrelated. The overlap is lexical, not conceptual.

2. DRAX in robust video question answering

DRAX was introduced for video question answering on the SUTD-TrafficQA benchmark, with the stated aim of removing latent distractors rather than sparsifying the raw input or relying on external knowledge. The method targets the observation that even low-weight attention on irrelevant content can corrupt attended representations during training and inference. Its two named components are Distraction Removal (DR) and Attended Cross-Alignment (AX) (Dogra et al., 2023).

The core DR mechanism operates on a cross-attention matrix $A$. For a query row $i$, DRAX defines the representative relevance as

$\rho_i = \max_j A_{i,j},$

then sets a threshold

$\tau_i = d_f \cdot \rho_i,$

where $d_f$ is the distraction factor. Any attention weight below $\tau_i$ is treated as a distractor and zeroed out. In the multi-head setting this is done head-wise, so masking suppresses specific subspaces rather than entire tokens. Across layers, the threshold is progressively tightened by incrementing $d_f$; for DRAX-base with $K=3$, the schedule is $0.3 \rightarrow 0.6 \rightarrow 0.9$ (Dogra et al., 2023).

Architecturally, DRAX is instantiated as a hierarchical VideoQA model built on offline features: appearance from ResNet-18, motion from ResNeXt-101, and text from GloVe embeddings. The hierarchy proceeds through an Appearance–Motion DRAX block, a Question DRAX block, and an Answer DRAX block. Each stage uses self-attention encoders, a cross-encoder with DR-modified multi-head cross-attention, and cross-aligned fusion. The AX component treats attention as a vector-space transformation from a tailing space into an anchor space: $X_t^{align} = A_{a \to t}^{masked} \cdot X_t,$ followed by learned fusion

$i$0

This avoids plain concatenation or exclusive reliance on a CLS token (Dogra et al., 2023).

Training uses a 4-way multiple-choice hinge loss, following HCRN and TGIF-QA, with no separate explicit loss term for distraction removal or alignment: $i$1 The total objective is therefore effectively $i$2 (Dogra et al., 2023).

Empirically, on SUTD-TrafficQA the reported full-dataset accuracies are 39.63 ± 0.24 for DRAX-base and 40.40 ± 0.76 for DRAX-large, compared with 37.05 for ECLIPSE and 36.49 for HCRN. Ablations show that removing cross-aligned fusion lowers full accuracy from 39.63 to 38.75, removing distraction masking lowers it to 38.93, and removing both lowers it to 38.06. The gains are especially pronounced on Attribution, Reverse Reasoning, Counterfactual Inference, and Introspection, where distractor suppression and hierarchical cross-modal conditioning appear most consequential (Dogra et al., 2023).

Within VideoQA, DRAX is best understood not as an input-pruning method but as a latent-space attention-control mechanism. Its novelty lies in hard-zeroing relative low-relevance attention entries and reusing the same filtered attention as a semantic alignment operator during fusion.

3. Drax in non-autoregressive speech recognition

In automatic speech recognition, Drax denotes a non-autoregressive (NAR) speech recognition framework based on discrete flow matching (DFM). Its motivating problem is the latency of autoregressive systems such as Whisper and Qwen2-Audio, whose decoding time scales with output length, whereas classical NAR approaches such as CTC are faster but generally less expressive for long-range dependencies. Drax is presented as both a training-and-inference framework and a concrete model instantiation: a Whisper encoder plus a DiT-style transformer decoder trained with DFM (Navon et al., 5 Oct 2025).

The formal substrate is a continuous-time Markov chain on token sequences $i$3, with a time-indexed probability path $i$4 interpolating from a source distribution $i$5 to a target transcription distribution $i$6. Standard DFM for discrete sequences often uses a two-way mixture between noise and target, but Drax argues that this produces a train-inference mismatch for ASR, because real inference trajectories pass through acoustically plausible but incorrect transcripts rather than pure random noise (Navon et al., 5 Oct 2025).

Its distinctive design is the audio-conditioned tri-mixture path

$i$7

where $i$8 is an audio-conditioned middle distribution approximating a good but imperfect ASR hypothesis. Early times are dominated by uniform noise, middle times by the audio-conditioned hypothesis, and late times by the ground truth. Training uses a conditional DFM cross-entropy objective together with an auxiliary middle-distribution loss,

$i$9

The paper’s theoretical analysis links the generalization gap to divergence between training occupancies and inference occupancies, controlled by cumulative velocity errors along the generated path, thereby motivating the path design (Navon et al., 5 Oct 2025).

The implementation uses a frozen Whisper large-v3 encoder and a DiT-style transformer decoder with cross-attention to audio features at every layer. The main Drax model has 16 transformer blocks, 20 heads, hidden size 1280, corresponding to approximately 580M decoder parameters; Drax-flash uses 4 blocks with the same width, at approximately 250M parameters. The separate middle-distribution network $\rho_i = \max_j A_{i,j},$0 is a small transformer with 1 block + projection and around 28M parameters. Total size is reported as approximately 1.2B parameters including encoder and middle network (Navon et al., 5 Oct 2025).

Inference is fully parallel over positions. Sampling proceeds for a fixed number of function evaluations (NFEs), such as 4, 8, or 16, rather than token-by-token generation. At test time Drax does not use the middle distribution in the sampler; the paper reports that including $\rho_i = \max_j A_{i,j},$1 at inference worsens WER on all datasets. Because the flow is stochastic, multiple candidate transcriptions can be sampled and then selected by mode voting, minimum Bayes risk (MBR), Whisper rescoring, or an internal ELBO-style score, with MBR and Whisper rescoring reported as particularly effective (Navon et al., 5 Oct 2025).

On English Open-ASR benchmarks, the reported average WERs and runtimes are: Whisper large-v3 at 7.6 with RTFx ~18, Qwen2-Audio at 7.8 with RTFx 6.4, Voxtral at 7.4 with RTFx 10.9, Drax (single sample, 16 NFEs) at 8.4 with RTFx 32.2, Drax + MBR (8 NFEs, 16 candidates) at 8.0 with RTFx 20.8, and Drax + Whisper rescoring (8 NFEs, 16 candidates) at 7.4 with RTFx 17.8. On LibriSpeech, Drax is also reported to outperform prior diffusion-based ASR models such as TransFusion, FFDM, and Whisfusion in WER (Navon et al., 5 Oct 2025).

Within ASR, Drax is therefore a discrete-flow alternative to autoregressive decoding whose central contribution is not merely parallel decoding, but a training path explicitly constructed to resemble plausible inference errors.

4. “Drax” as Draco in dark-matter dynamics

In the astrophysical material, “Drax” is treated as an informal reference to Draco, one of the classical Milky Way dwarf spheroidal galaxies. Draco is analyzed as a dark-matter-dominated system relevant to the core–cusp problem and to indirect detection through its astrophysical $\rho_i = \max_j A_{i,j},$2- and $\rho_i = \max_j A_{i,j},$3-factors. Two supplied papers address Draco with different dynamical formalisms: an axisymmetric JAM analysis of DESI plus literature spectroscopy, and a later distribution-function-based axisymmetric treatment (Yang et al., 3 Jul 2025, Pascale et al., 27 Apr 2026).

The DESI-based study combines DESI Milky Way Survey Year-3 spectroscopy with Walker et al. (2023) data, yielding a Draco member catalog with $\rho_i = \max_j A_{i,j},$4 stars after selection. It models the stellar tracers with either a single-population Jeans analysis or a two-population chemodynamical model separating metal-rich and metal-poor stars. In this framework the dark halo is spherical with generalized NFW density

$\rho_i = \max_j A_{i,j},$5

with $\rho_i = \max_j A_{i,j},$6, and the key inferential target is the inner logarithmic slope $\rho_i = \max_j A_{i,j},$7. The preferred chemodynamical result for Draco is

$\rho_i = \max_j A_{i,j},$8

while the single-population model gives

$\rho_i = \max_j A_{i,j},$9

The same study reports for Draco at $\tau_i = d_f \cdot \rho_i,$0: $\tau_i = d_f \cdot \rho_i,$1 It also finds two chemically and dynamically distinct stellar populations: a centrally concentrated, dynamically colder metal-rich component and a more extended, hotter metal-poor component (Yang et al., 3 Jul 2025).

The later distribution-function analysis introduces fully DF-based, multi-component axisymmetric models in which two chemo-dynamically distinct flattened stellar populations orbit in a spherical potential generated by a dominant dark-matter halo plus a possible intermediate-mass black hole (IMBH). For Draco, the fiducial flattened two-component model gives

$\tau_i = d_f \cdot \rho_i,$2

with the G26 sample and

$\tau_i = d_f \cdot \rho_i,$3

with the W23 sample. It reports

$\tau_i = d_f \cdot \rho_i,$4

and at $\tau_i = d_f \cdot \rho_i,$5,

$\tau_i = d_f \cdot \rho_i,$6

The same paper reports no evidence for an IMBH and places a $\tau_i = d_f \cdot \rho_i,$7 upper limit

$\tau_i = d_f \cdot \rho_i,$8

It also emphasizes that Draco’s dark-matter inference is unusually stable across data sets and modeling variants, while spherical stellar models can bias the inferred inner slope toward cuspier values in more flattened systems (Pascale et al., 27 Apr 2026).

Taken together, these studies support a consistent qualitative picture of Draco as a dense, dark-matter-dominated, cusp-like dwarf spheroidal, while also showing that the precise value of $\tau_i = d_f \cdot \rho_i,$9 and the line-of-sight integrals relevant to indirect detection remain sensitive to modeling assumptions, data selection, and outer-halo treatment.

5. Methodological contrasts across the three usages

The three meanings of “Drax” occupy unrelated scientific domains, but each is organized around an internal representation problem.

In DRAX for VideoQA, the operative object is a cross-attention matrix whose low-relevance entries are hard-zeroed relative to a query-wise maximum. The problem is contamination of latent cross-modal embeddings by weakly relevant information. The method is therefore an architectural intervention inside multi-head attention and cross-modal fusion (Dogra et al., 2023).

In Drax for ASR, the operative object is a time-indexed probability path over token sequences and its associated velocity field under a CTMC. The problem is mismatch between the states visited during training and those encountered during generation. The method is therefore a training-path and decoding framework that shapes the occupancy of intermediate token states (Navon et al., 5 Oct 2025).

In Draco dynamics, the operative objects are the stellar phase-space distribution and the dark-matter density profile inferred from discrete stellar observations. The problem is degeneracy among halo slope, anisotropy, flattening, and sample selection. The methods are dynamical inference procedures—first Jeans-based, then DF-based—that estimate $d_f$0, $d_f$1, $d_f$2, and, in the later work, an IMBH mass limit (Yang et al., 3 Jul 2025, Pascale et al., 27 Apr 2026).

A plausible implication is that the lexical convergence of these names is almost accidental: the papers do not share a common technical lineage. What they do share is a concern with suppressing or controlling hidden confounding structure—attention distractors in VideoQA, unrealistic intermediate states in NAR ASR, and modeling degeneracies in dynamical astronomy.

6. Terminological precision and recurrent misconceptions

The most basic misconception is to treat “Drax” as a single established concept. In the supplied literature it is not. DRAX and Drax are unrelated methods, and the astrophysical “Drax” is explicitly only an informal reference to Draco (Dogra et al., 2023, Yang et al., 3 Jul 2025).

Within the VideoQA literature, DRAX should not be confused with frame-sampling or external-knowledge approaches. The paper explicitly positions it against methods that sparsely sample frames or tokens, fine-grain the input, or add external knowledge, arguing instead for explicit cleaning of latent embeddings (Dogra et al., 2023).

Within ASR, Drax should not be interpreted as a standard diffusion-from-noise model. The paper distinguishes discrete flow matching from multinomial diffusion and argues that its audio-conditioned middle distribution is a training-only device. A direct misunderstanding would be to assume that $d_f$3 is also beneficial at inference; the reported ablation states the opposite, namely that using it during decoding worsens WER (Navon et al., 5 Oct 2025).

Within Draco dark-matter studies, the later DF-based analysis directly warns against assuming that spherical modeling is innocuous for flattened systems. It reports that spherical stellar models can bias the inner slope toward cuspier values, even though Draco itself remains cusp-like across the tested models. The earlier DESI-based analysis likewise stresses that dynamical inferences remain subject to uncertainties in both methodology and observed data (Yang et al., 3 Jul 2025, Pascale et al., 27 Apr 2026).

For technical writing, the term therefore requires immediate contextual qualification: DRAX (VideoQA), Drax (DFM-ASR), or Draco (“Drax”). Without that qualification, the label is underdetermined.