AMix-1: Multi-Domain Generative Models

• AMix-1 is a label assigned to three distinct constructs across protein modeling, music mixing, and quantum error correction, requiring clear domain disambiguation.
• In protein engineering, it employs Bayesian Flow Networks with in-context learning and evolutionary test-time scaling to achieve significant performance improvements.
• In music and quantum applications, MEGAMI uses diffusion-based generative mixing while the mixed-alphabet quantum code surpasses traditional noiseless channels in error correction.

AMix-1 is a label attached in the arXiv literature to three distinct technical constructs rather than to a single unified method. In protein machine learning, it denotes a family of protein foundation models built on Bayesian Flow Networks and extended with in-context learning and test-time scaling (Lv et al., 11 Jul 2025). In automatic music production, it denotes the MEGAMI framework, a generative approach to multitrack music mixing that models the conditional distribution of professional mixes given unprocessed tracks (Moliner et al., 11 Nov 2025). In quantum error correction, it denotes a mixed-alphabet code construction for partial-noisy channels within the “$\tfrac12+\tfrac12>1$” setting (Wang et al., 2012). The term therefore requires domain-specific disambiguation.

1. Uses of the term

In the supplied literature, “AMix-1” appears in protein modeling, automatic music mixing, and mixed-alphabet quantum coding. The usages are technically unrelated, but each is associated with a concrete model or construction.

Usage	Domain	Primary source
AMix-1	Protein foundation model	(Lv et al., 11 Jul 2025)
MEGAMI (“AMix-1”)	Automatic music mixing	(Moliner et al., 11 Nov 2025)
AMix-1 code construction	Quantum error correction	(Wang et al., 2012)

This multiplicity is significant because the same label refers to different mathematical objects: a generative sequence model, a generative audio-mixing system, and a graph-state code. A plausible implication is that citations to “AMix-1” require explicit domain qualification to avoid ambiguity.

2. Protein AMix-1: Bayesian Flow Network foundation model

AMix-1, in the protein-design sense, is a family of protein foundation models built on the Bayesian Flow Network (BFN) framework with an encoder-only Transformer backbone (Lv et al., 11 Jul 2025). Rather than directly modeling discrete amino acids, BFNs learn continuous distributions over one-hot–encoded sequences via iterative Bayesian updates. At timestep $i$, the sender distribution is

$$p_s(\mathbf{y}_i\mid \mathbf{x};\alpha_i) = \mathcal{N}\bigl(\alpha_i(K\,\mathbf{e}(\mathbf{x})-\mathbf{1}),\,\alpha_i\,K\,\mathbf{I}\bigr),$$

with $\alpha_i=\beta(t_i)=\beta_1\,t_i^2$, while the receiver forms

$$p_r(\mathbf{y}_i\mid p_\text{out},\alpha_i) = \mathbb{E}_{\mathbf{x}'\sim p_\text{out}}\bigl[p_s(\mathbf{y}_i\mid \mathbf{x}';\alpha_i)\bigr].$$

Training minimizes

$$\mathcal{L}_\text{BFN} = \mathbb{E}_{\mathbf{x}\sim p_{\rm data}} \sum_{i=1}^n D_{\rm KL}\bigl[p_s(\mathbf{y}_i\mid \mathbf{x},\alpha_i)\,\Vert\,p_r(\mathbf{y}_i\mid p_\text{out},\alpha_i)\bigr].$$

The Transformer backbone is RoPE-augmented and scales from $8$ M to $1.7$ B parameters.

Model size	Architecture	Heads
8 M	6 layers, hidden dim 320	20
35 M	12 layers, hidden dim 480	20
150 M	30 layers, hidden dim 640	20
350 M	33 layers, hidden dim 960	20
650 M	33 layers, hidden dim 1280	20
1.7 B	48 layers, hidden dim 1680	40

The pretraining setup uses UniRef50, with approximately $41$ M training sequences and $83$ k validation sequences, an $i$0 noise scheduler with $i$1, AdamW, peak learning rate $i$2, $i$3 M steps, and mixed-precision BF16. Predictive scaling laws are fit against total training FLOPs $i$4, where $i$5 is the number of parameters and $i$6 the number of tokens, using

$i$7

The reported interpretation is that robust scalability depends on a predictive scaling law and on an emergence analysis conducted from the loss perspective. Under intermediate noise, specifically $i$8, sequence consistency and structural metrics exhibit sudden nonlinear gains, whereas at extreme noise levels $i$9 or $$p_s(\mathbf{y}_i\mid \mathbf{x};\alpha_i) = \mathcal{N}\bigl(\alpha_i(K\,\mathbf{e}(\mathbf{x})-\mathbf{1}),\,\alpha_i\,K\,\mathbf{I}\bigr),$$0 such transitions vanish. This suggests that the model’s structural understanding is coupled to calibrated noise schedules rather than to parameter count alone.

3. Protein AMix-1: in-context learning, test-time scaling, and results

AMix-1 uses a Multiple Sequence Alignment (MSA)-based in-context learning mechanism that converts MSAs into position-wise frequency profiles

$$p_s(\mathbf{y}_i\mid \mathbf{x};\alpha_i) = \mathcal{N}\bigl(\alpha_i(K\,\mathbf{e}(\mathbf{x})-\mathbf{1}),\,\alpha_i\,K\,\mathbf{I}\bigr),$$1

where $$p_s(\mathbf{y}_i\mid \mathbf{x};\alpha_i) = \mathcal{N}\bigl(\alpha_i(K\,\mathbf{e}(\mathbf{x})-\mathbf{1}),\,\alpha_i\,K\,\mathbf{I}\bigr),$$2 is the number of homologs, $$p_s(\mathbf{y}_i\mid \mathbf{x};\alpha_i) = \mathcal{N}\bigl(\alpha_i(K\,\mathbf{e}(\mathbf{x})-\mathbf{1}),\,\alpha_i\,K\,\mathbf{I}\bigr),$$3 the sequence length, and $$p_s(\mathbf{y}_i\mid \mathbf{x};\alpha_i) = \mathcal{N}\bigl(\alpha_i(K\,\mathbf{e}(\mathbf{x})-\mathbf{1}),\,\alpha_i\,K\,\mathbf{I}\bigr),$$4 the residue types (Lv et al., 11 Jul 2025). Conditional generation is written as

$$p_s(\mathbf{y}_i\mid \mathbf{x};\alpha_i) = \mathcal{N}\bigl(\alpha_i(K\,\mathbf{e}(\mathbf{x})-\mathbf{1}),\,\alpha_i\,K\,\mathbf{I}\bigr),$$5

The inference procedure requires no finetuning: the profile is fed forward at test time and each position is greedily decoded. The stated purpose is to preserve both evolving structural motifs and functional specificity within a unified design framework.

The test-time scaling component, EvoAMix-1, is an evolutionary propose-verify-update loop. At round $$p_s(\mathbf{y}_i\mid \mathbf{x};\alpha_i) = \mathcal{N}\bigl(\alpha_i(K\,\mathbf{e}(\mathbf{x})-\mathbf{1}),\,\alpha_i\,K\,\mathbf{I}\bigr),$$6, the proposal distribution is

$$p_s(\mathbf{y}_i\mid \mathbf{x};\alpha_i) = \mathcal{N}\bigl(\alpha_i(K\,\mathbf{e}(\mathbf{x})-\mathbf{1}),\,\alpha_i\,K\,\mathbf{I}\bigr),$$7

and the prompt is updated from top-scoring sequences rather than by parameter finetuning. Performance is summarized through a verifier-budget relation $$p_s(\mathbf{y}_i\mid \mathbf{x};\alpha_i) = \mathcal{N}\bigl(\alpha_i(K\,\mathbf{e}(\mathbf{x})-\mathbf{1}),\,\alpha_i\,K\,\mathbf{I}\bigr),$$8 with $$p_s(\mathbf{y}_i\mid \mathbf{x};\alpha_i) = \mathcal{N}\bigl(\alpha_i(K\,\mathbf{e}(\mathbf{x})-\mathbf{1}),\,\alpha_i\,K\,\mathbf{I}\bigr),$$9, where $\alpha_i=\beta(t_i)=\beta_1\,t_i^2$0, indicating monotonic gains as verification budget increases.

The reported experimental results include a wet-lab AmeR design task in which AMix-1-650M generated $\alpha_i=\beta(t_i)=\beta_1\,t_i^2$1 variants with at most $\alpha_i=\beta(t_i)=\beta_1\,t_i^2$2 mutations each; the best variant achieved up to $\alpha_i=\beta(t_i)=\beta_1\,t_i^2$3 activity increase over wild type, approximately $\alpha_i=\beta(t_i)=\beta_1\,t_i^2$4 gain versus an EvoAI baseline. In silico directed-evolution benchmarks cover six tasks: orphan protein foldability, general protein family design, optimal temperature, optimal pH, EC number reprogramming, and specific reaction activity. Across all six, EvoAMix-1 is reported to outperform or match ALDE, EVOLVEpro, and MLDE in final performance and improvement rate as a function of verifier calls. The broader implication stated in the source is a pathway toward lab-in-the-loop protein engineering, with open challenges including multimodal prompting, calibrated noise schedules, and theoretical analysis of proposal distribution dynamics.

4. MEGAMI (“AMix-1”): generative automatic music mixing

In automatic music mixing, AMix-1 refers to a detailed formulation of MEGAMI, a generative framework for multitrack mixing that treats professional mixing as a one-to-many problem rather than as deterministic regression (Moliner et al., 11 Nov 2025). The unprocessed tracks are denoted $\alpha_i=\beta(t_i)=\beta_1\,t_i^2$5, the professional mixed tracks by $\alpha_i=\beta(t_i)=\beta_1\,t_i^2$6, and the per-track processed outputs by $\alpha_i=\beta(t_i)=\beta_1\,t_i^2$7. Assuming no master bus effects, the final mix is

$\alpha_i=\beta(t_i)=\beta_1\,t_i^2$8

so that

$\alpha_i=\beta(t_i)=\beta_1\,t_i^2$9

MEGAMI factorizes $$p_r(\mathbf{y}_i\mid p_\text{out},\alpha_i) = \mathbb{E}_{\mathbf{x}'\sim p_\text{out}}\bigl[p_s(\mathbf{y}_i\mid \mathbf{x}';\alpha_i)\bigr].$$0 by introducing latent effect embeddings $$p_r(\mathbf{y}_i\mid p_\text{out},\alpha_i) = \mathbb{E}_{\mathbf{x}'\sim p_\text{out}}\bigl[p_s(\mathbf{y}_i\mid \mathbf{x}';\alpha_i)\bigr].$$1:

$$p_r(\mathbf{y}_i\mid p_\text{out},\alpha_i) = \mathbb{E}_{\mathbf{x}'\sim p_\text{out}}\bigl[p_s(\mathbf{y}_i\mid \mathbf{x}';\alpha_i)\bigr].$$2

where $$p_r(\mathbf{y}_i\mid p_\text{out},\alpha_i) = \mathbb{E}_{\mathbf{x}'\sim p_\text{out}}\bigl[p_s(\mathbf{y}_i\mid \mathbf{x}';\alpha_i)\bigr].$$3 are CLAP content embeddings and $$p_r(\mathbf{y}_i\mid p_\text{out},\alpha_i) = \mathbb{E}_{\mathbf{x}'\sim p_\text{out}}\bigl[p_s(\mathbf{y}_i\mid \mathbf{x}';\alpha_i)\bigr].$$4 is a deterministic effect processor. The generative task is therefore to model $$p_r(\mathbf{y}_i\mid p_\text{out},\alpha_i) = \mathbb{E}_{\mathbf{x}'\sim p_\text{out}}\bigl[p_s(\mathbf{y}_i\mid \mathbf{x}';\alpha_i)\bigr].$$5.

Ground-truth effect embeddings are extracted from wet stems $$p_r(\mathbf{y}_i\mid p_\text{out},\alpha_i) = \mathbb{E}_{\mathbf{x}'\sim p_\text{out}}\bigl[p_s(\mathbf{y}_i\mid \mathbf{x}';\alpha_i)\bigr].$$6 through an injective “FxEncoder++” $$p_r(\mathbf{y}_i\mid p_\text{out},\alpha_i) = \mathbb{E}_{\mathbf{x}'\sim p_\text{out}}\bigl[p_s(\mathbf{y}_i\mid \mathbf{x}';\alpha_i)\bigr].$$7,

$$p_r(\mathbf{y}_i\mid p_\text{out},\alpha_i) = \mathbb{E}_{\mathbf{x}'\sim p_\text{out}}\bigl[p_s(\mathbf{y}_i\mid \mathbf{x}';\alpha_i)\bigr].$$8

and then augmented with a $$p_r(\mathbf{y}_i\mid p_\text{out},\alpha_i) = \mathbb{E}_{\mathbf{x}'\sim p_\text{out}}\bigl[p_s(\mathbf{y}_i\mid \mathbf{x}';\alpha_i)\bigr].$$9-dimensional “dynamic+stereo” Fourier-feature vector to obtain $$\mathcal{L}_\text{BFN} = \mathbb{E}_{\mathbf{x}\sim p_{\rm data}} \sum_{i=1}^n D_{\rm KL}\bigl[p_s(\mathbf{y}_i\mid \mathbf{x},\alpha_i)\,\Vert\,p_r(\mathbf{y}_i\mid p_\text{out},\alpha_i)\bigr].$$0. At inference, embeddings are sampled with a conditional diffusion model using the Elucidating Diffusion Models schema. Forward diffusion adds noise independently to each track,

$$\mathcal{L}_\text{BFN} = \mathbb{E}_{\mathbf{x}\sim p_{\rm data}} \sum_{i=1}^n D_{\rm KL}\bigl[p_s(\mathbf{y}_i\mid \mathbf{x},\alpha_i)\,\Vert\,p_r(\mathbf{y}_i\mid p_\text{out},\alpha_i)\bigr].$$1

and the reverse process uses the probability-flow ODE

$$\mathcal{L}_\text{BFN} = \mathbb{E}_{\mathbf{x}\sim p_{\rm data}} \sum_{i=1}^n D_{\rm KL}\bigl[p_s(\mathbf{y}_i\mid \mathbf{x},\alpha_i)\,\Vert\,p_r(\mathbf{y}_i\mid p_\text{out},\alpha_i)\bigr].$$2

initialized as $$\mathcal{L}_\text{BFN} = \mathbb{E}_{\mathbf{x}\sim p_{\rm data}} \sum_{i=1}^n D_{\rm KL}\bigl[p_s(\mathbf{y}_i\mid \mathbf{x},\alpha_i)\,\Vert\,p_r(\mathbf{y}_i\mid p_\text{out},\alpha_i)\bigr].$$3. A score network $$\mathcal{L}_\text{BFN} = \mathbb{E}_{\mathbf{x}\sim p_{\rm data}} \sum_{i=1}^n D_{\rm KL}\bigl[p_s(\mathbf{y}_i\mid \mathbf{x},\alpha_i)\,\Vert\,p_r(\mathbf{y}_i\mid p_\text{out},\alpha_i)\bigr].$$4 is trained by denoising score matching.

The score network is implemented as a set Transformer. Self-attention over $$\mathcal{L}_\text{BFN} = \mathbb{E}_{\mathbf{x}\sim p_{\rm data}} \sum_{i=1}^n D_{\rm KL}\bigl[p_s(\mathbf{y}_i\mid \mathbf{x},\alpha_i)\,\Vert\,p_r(\mathbf{y}_i\mid p_\text{out},\alpha_i)\bigr].$$5 captures inter-track style correlations, while cross-attention from $$\mathcal{L}_\text{BFN} = \mathbb{E}_{\mathbf{x}\sim p_{\rm data}} \sum_{i=1}^n D_{\rm KL}\bigl[p_s(\mathbf{y}_i\mid \mathbf{x},\alpha_i)\,\Vert\,p_r(\mathbf{y}_i\mid p_\text{out},\alpha_i)\bigr].$$6 to $$\mathcal{L}_\text{BFN} = \mathbb{E}_{\mathbf{x}\sim p_{\rm data}} \sum_{i=1}^n D_{\rm KL}\bigl[p_s(\mathbf{y}_i\mid \mathbf{x},\alpha_i)\,\Vert\,p_r(\mathbf{y}_i\mid p_\text{out},\alpha_i)\bigr].$$7 injects content information. Permutation equivariance under track re-ordering is enforced by random permutation during training and by appending a one-hot “track index” vector to each $$\mathcal{L}_\text{BFN} = \mathbb{E}_{\mathbf{x}\sim p_{\rm data}} \sum_{i=1}^n D_{\rm KL}\bigl[p_s(\mathbf{y}_i\mid \mathbf{x},\alpha_i)\,\Vert\,p_r(\mathbf{y}_i\mid p_\text{out},\alpha_i)\bigr].$$8 and $$\mathcal{L}_\text{BFN} = \mathbb{E}_{\mathbf{x}\sim p_{\rm data}} \sum_{i=1}^n D_{\rm KL}\bigl[p_s(\mathbf{y}_i\mid \mathbf{x},\alpha_i)\,\Vert\,p_r(\mathbf{y}_i\mid p_\text{out},\alpha_i)\bigr].$$9. Variable track counts are handled by padding up to $8$0 with masking. The effect processor $8$1 is a track-agnostic temporal convolutional network taking mono $8$2 plus concatenated $8$3, injected by FiLM. Stereo panning and width are encoded in $8$4; the input is collapsed to mono and re-stereoized by $8$5. This architecture is designed to support arbitrary unlabeled tracks while maintaining shared processing structure.

5. MEGAMI (“AMix-1”): domain adaptation, losses, and evaluation

MEGAMI includes a domain-adaptation mechanism intended to use large wet-only corpora without paired dry stems (Moliner et al., 11 Nov 2025). Let $8$6 denote the dry-stem distribution and $8$7 the wet-stem distribution. A small MLP adaptor $8$8 is trained to remove wet-effect artifacts from CLAP content embeddings so that, for paired dry/wet singles, $8$9, $1.7$0, and $1.7$1. The adaptor loss is

$1.7$2

During diffusion-model training, each wet stem contributes a conditioning embedding

$1.7$3

where the added Gaussian is said to simulate the smoothing kernel in the convolutional-distribution matching formulation.

The full objective combines diffusion, reconstruction, feature, and adaptor terms. The score-network loss is

$1.7$4

For the effect processor, the per-track multi-scale spectral reconstruction loss is

$1.7$5

and the deep feature consistency loss is

$1.7$6

The total loss is

$1.7$7

with $1.7$8 chosen by validation.

Evaluation uses Kernel Audio Distance (KAD), computed as MMD-based distances between system mixes and human mixes under AFxRep, FxEncoder, FxEncoder++, and CLAP embeddings. The reported table states that MEGAMI (I-L) obtains the lowest KAD across almost all embeddings.

Embedding	MEGAMI (I-L)	Comparator
AFxRep	5.21	FxNorm-AutoMix L = 11.77
FxEncoder	1.72	DMC = 75.74
FxEncoder++	3.90	E2E-Flow = 14.98
CLAP	0.84	FxNorm-AutoMix L = 1.31

The subjective listening test is a multi-stimulus test on $1.7$9 songs $41$0 $41$1 trials with $41$2 listeners, including $41$3 professional engineers. Each page contains the human reference, Equal Loudness, E2E-Flow, FxNorm-AutoMix L, and MEGAMI (I-L). MEGAMI (I-L) is reported to outperform all baselines in median and mean ratings and, in several cases, to be rated above the human reference. Deterministic baselines include Equal Loudness, FxNorm-AutoMix (S/L), DMC, MixWaveUNet, and E2E-Flow. The stated conclusion is that generative modeling of the one-to-many mix space yields objectively closer style distributions and subjectively preferred mixes relative to single-best regression.

6. Quantum AMix-1: mixed-alphabet code for half-noisy channels

In quantum error correction, AMix-1 denotes a specific mixed-alphabet code construction introduced in the context of partial-noisy, or half-noisy, channels (Wang et al., 2012). A half-noisy channel starts from a $41$4-level qudit viewed as two $41$5-level subsystems, with generalized Pauli operators $41$6 for $41$7. It acts as a fully general error channel on subsystem $41$8 but approximately as the identity on subsystem $41$9. The notation “$83$0” denotes a half-noisy channel and “$83$1” a fully noiseless channel.

The AMix-1 construction is the mixed-alphabet code

$83$2

which encodes $83$3 logical states into eight noisy qubits and one half-noisy ququart. The $83$4-level qudit is regarded as two qubits, $83$5 and $83$6, with qubit $83$7 transmitted noiselessly. The construction uses a $83$8-weighted graph $83$9 on $i$00, with adjacency as in Fig. 1(B) of the paper, and a joint graph state $i$01. The composite coding clique $i$02 is generated by five independent vectors,

$i$03

producing $i$04 mutually orthogonal basis states. The logical basis is

$i$05

All single-“qubit” errors on the nine noisy subsystems are correctable, namely

$i$06

together with their products up to weight $i$07. For any two correctable errors $i$08 with total weight $i$09, the Knill–Laflamme condition

$i$10

holds on the logical basis.

The construction is used to illustrate the “$i$11” phenomenon. The comparison given in the source is between the standard entanglement-assisted code

$i$12

and the half-noisy mixed-alphabet code

$i$13

for which $i$14. The paper therefore states that one half-noisy ququart outperforms one perfectly noiseless qubit by a factor $i$15 in code dimension. A second numerical example compares $i$16 with $i$17, showing $i$18.

The code also appears in the discussion of the unified quantum Singleton bound. For any $i$19 code,

$i$20

The paper states that $i$21 saturates this bound. The practical lesson drawn there is that increasing subsystem dimension and allowing partial noise can surpass rates achievable by lower-dimensional noiseless channels.