{"id":"767e0a88-f081-4543-b448-2619a5397a34","arxiv_id":"2505.08175","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"ARC post-training speeds up text-to-audio generation to near-real-time speeds on GPUs and a few seconds on phones, without distillation or classifier-free guidance.","lead":"This paper introduces a training method that makes text-to-audio AI generate sound much faster by teaching an existing model with an adversarial 'fake versus real' game. It reports generating about 12 seconds of stereo audio in roughly 75 milliseconds on a high-end GPU and in about 7 seconds on a phone.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sign error in Eq. (5)/(8) reverses the min-max objective: the discriminator is trained to reward generated samples and incorrect prompt pairings, making the central claim irreproducible as written.","rationale":"The reader's weakest assumption focused on whether the contrastive loss LC, which trains only the discriminator, can indirectly improve generator prompt adherence through the relativistic loss LR. That is a reasonable mechanistic concern, and the paper's ablation partially addresses it. However, a more fundamental and previously unnoticed defect appears in the very definition of the objective. With f(x) = −log(1+e^{−x}), which is increasing, the min-max game specified in Eq. (4) trains the discriminator to maximize logits of generated samples relative to real samples and to maximize logits of incorrect prompt pairings relative to correct ones. This is the precise opposite of both the textual description in §2.3–2.4 and the behavior needed for the claimed results. The empirical results in Table 1 are therefore either unattainable with the stated objective, or the implementation diverges from the equations—either way, the central claim is not reproducible from the manuscript. This is a load-bearing correctness issue, so the appropriate verdict is REJECT (or at minimum UNVERDICTED pending code inspection), rather than the reader's CONDITIONAL. The reader's noted issues about overclaimed priority, missing error bars, and unvalidated CCDS are secondary to this sign inconsistency.","tokens_in":10816,"tokens_out":16535,"duration_ms":160442,"concrete_test":"Compute the gradient of LR in Eq. (5) with respect to the discriminator output on a generated sample, treating f(x) = −log(1+e^{−x}) and ∆gen = Dψ(q(x̂_s|·)). Since f'(x) = e^{−x}/(1+e^{−x}) > 0, the gradient with respect to ∆gen is positive, so gradient ascent on ψ increases generated-sample logits. Then inspect the released code (or re-train a small model on a single prompt) to verify whether the implemented loss matches Eq. (5) with f(x) = −log(1+e^{−x}) or uses the sign-flipped form f(x) = −log(1+e^{x}). If the code matches Eq. (5), training should produce a discriminator that scores generated samples above real samples, and a generator that degrades prompt adherence; if the code uses the flipped form, the paper's equations must be corrected to restore reproducibility.","verdict_should_be":"REJECT","load_bearing_attack":"The central objective in Eq. (4) is specified as min_φ max_ψ LARC = LR + λ·LC, with LR in Eq. (5) using f(x) = −log(1+e^{−x}). This f is strictly increasing in x, so LR = f(∆gen − ∆real) is maximized by making ∆gen > ∆real. Thus the discriminator Dψ (which maximizes LR) is explicitly trained to assign higher logits to generated samples than to paired real samples, the opposite of its intended role. The generator Gφ (which minimizes LR) is thereby pushed to make its samples less real than real ones. This directly contradicts the text in §2.3, which states that Gφ wants each generated sample to be 'more real than its paired real sample' and Dψ wants each real sample to be 'more real than its paired generated sample.' The same sign error appears in the contrastive loss LC (Eq. 8): maximizing f(∆real(·,P[c]) − ∆real(·,c)) with an increasing f trains Dψ to assign higher logits to incorrect audio–prompt pairs than to correct ones, the opposite of the intended contrastive signal described in §2.4. If the implementation follows the written equations, the method would reward poor prompt adherence and unrealism; the strong results in Table 1 could not arise. If the implementation flips a sign (e.g., uses f(x) = −log(1+e^{x}) or minimizes/maximizes the opposite direction), then the paper's equations mis-specify the actual method, making the central objective non-reproducible. This is a fundamental correctness issue that precedes the reader's concern about whether LC propagates semantically through LR.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces ARC (Adversarial Relativistic-Contrastive) post-training, an acceleration method for pre-trained rectified-flow text-to-audio models. ARC replaces the velocity-matching loss with a relativistic adversarial loss computed on paired real and generated audio (Eq. 5) and adds a contrastive discriminator loss over correct versus shuffled prompt-audio pairs (Eq. 8). The generator is trained without distillation or classifier-free guidance and is sampled with a ping-pong scheme. On a variant of Stable Audio Open, the authors report 8-step generation of about 12 seconds of 44.1 kHz stereo audio in about 75 ms on an H100, about 187 ms on a 3090, and about 6.6 seconds on a mobile phone after quantization. Experiments on AudioCaps include objective metrics, a proposed conditional diversity metric (CCDS), ablations against Presto and standard least-squares adversarial losses, and a small listening test.","tokens_in":11192,"tokens_out":7896,"duration_ms":79914,"significance":"If the paper's equations and results are taken at face value, ARC is a potentially useful contribution: it is a non-distillation, CFG-free adversarial post-training recipe for audio that appears to preserve diversity better than the distillation baseline while matching its speed, and the on-device deployment experiments are a valuable practical data point. The paper also commits to public code and a demo, and the empirical evaluation uses an external benchmark and listening tests rather than fitted constants. The main caveat is that the two defining losses are written with reversed signs, so the method as specified is not reproducible; this must be corrected before the contribution can be assessed. The concern that the contrastive loss only trains the discriminator and may not propagate to the generator is partially addressed empirically by the LR vs. LR+LC ablation, which shows improved CLAP score and prompt adherence when LC is included, although no theoretical analysis is provided.","major_comments":[{"comment":"The relativistic loss is sign-reversed. Since f(x) = -log(1+e^{-x}) is strictly increasing, maximizing LR with respect to D_psi pushes Delta_gen above Delta_real, and minimizing LR with respect to G_phi pushes Delta_gen below Delta_real. This is the opposite of the training goal stated in Sec. 2.3, where G should make generated samples 'more real' than their paired real sample and D should make real samples 'more real' than their paired generated sample. As written, the discriminator is rewarded for assigning higher logits to generated samples and the generator is rewarded for lowering them, which would not produce the reported quality. Please correct the sign (e.g., use f(x) = -log(1+e^{x}) or swap the arguments) and confirm that the text, equations, and released code agree; the reward-model interpretation in Sec. 2.6 is also reversed under the current sign.","section":"Sec. 2.3, Eqs. (4)-(5)"},{"comment":"The contrastive loss is also reversed. With f increasing, maximizing LC trains D_psi to assign higher logits to shuffled incorrect prompt-audio pairs than to correct pairs, which is the opposite of a contrastive objective intended to make the discriminator reward correct audio-text alignment. The equation should use f(Delta_real(x0,s,c) - Delta_real(x0,s,P[c])) or a decreasing f; otherwise the method as written actively rewards wrong prompts and would be expected to degrade prompt adherence rather than improve it.","section":"Sec. 2.4, Eq. (8)"},{"comment":"The speed claims are not supported by the paper's own table. In Table 1, +Presto and +ARC (ours) have identical RTF (156.42) at 8 steps, so 'the fastest text-to-audio model to our knowledge' is contradicted by the presented baseline, and no comparison is made to other cited fast text-to-audio models (e.g., the consistency-distillation models in refs. [26]-[29]). Additionally, the abstract and introduction claim the model is '100x faster' than SAO, but the Table 1 RTF values imply about a 44x latency improvement over 100-step SAO (3.56 vs. 156.42 in RTF, i.e., 12/3.56 s vs. 12/156.42 s). Please restrict the speed claims to the specific setting that supports them or add the missing comparisons.","section":"Abstract, Sec. 1, and Table 1"}],"minor_comments":[{"comment":"There is a typo in the section title: 'Objective Evalaution' should be 'Objective Evaluation'.","section":"Sec. 3.3 heading"},{"comment":"The objective metrics are reported without confidence intervals, error bars, or significance tests; adding variance across evaluation seeds or bootstrap intervals would strengthen the comparison, especially because several values are close (e.g., CLAP 0.27 for ARC and Presto).","section":"Table 1"},{"comment":"The listening test uses only 14 participants; please report the number of stimuli per condition and the statistical test used for the claims that differences are or are not significant, since the 95% confidence intervals alone do not establish significance across conditions.","section":"Sec. 3.4"},{"comment":"The sentence 'This decreases inference time from 15.3s (original F32) to 6.6s, and reducing peak runtime RAM usage...' has a grammatical issue ('and reducing' should be 'and reduces'); also clarify whether the reported edge times are for the same 12-second audio length and state the corresponding RTF on the device.","section":"Sec. 3.7"},{"comment":"The RTF column header would benefit from a footnote repeating the definition 'audio duration divided by latency' directly in the table caption, since the term is often defined in the opposite direction in the literature.","section":"Sec. 3.5"}],"recommendation":"major_revision","confidential_remarks":"The sign errors in Eqs. (5) and (8) are central and must be reconciled with the released code before publication; if the implementation indeed uses the opposite sign, the equations and the associated textual description need a correction, not just a clarification. The overclaim about being the fastest text-to-audio model is straightforward to fix by rewording and adding comparisons. I do not see grounds for rejection if the experiments are reproducible after the sign correction, but the current manuscript is not reproducible as written."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the method is genuinely interesting: it is the first adversarial post-training recipe for text-to-audio that avoids both distillation and classifier-free guidance, and the reported speedups are impressive. Second, the central equations are inconsistent with the text—the sign of the relativistic loss is flipped, so the discriminator is trained to reward generated samples over real ones, and the contrastive loss rewards incorrect prompt pairings. That cannot be what the authors implemented; the tables would not look like that. This is almost certainly a typo, but it is a load-bearing one: the method as written is not reproducible, and anyone trying to build on it from the equations alone would train the opposite objective.\n\nThe real contributions are worth crediting. Combining a relativistic adversarial loss with a contrastive discriminator objective is a plausible way to get prompt adherence without CFG, and it seems to work in practice if the signs are corrected. The CCDS metric for conditional diversity is a nice idea and appears to align with listening tests. The ablations are sensible, and the edge-device experiments with quantization and ARM optimizations are a useful data point for anyone deploying text-to-audio on phones.\n\nThe soft spots beyond the sign error: the “fastest text-to-audio model” claim is not supported by Table 1, where Presto has the same RTF, and there is no comparison with other cited fast audio models. Objective metrics have no error bars, so the differences between ARC and Presto are uncertain. The code link is missing from the manuscript, which matters because the equation problem makes the code the only reliable spec. The CCDS metric is only lightly validated, though the evidence presented is encouraging.\n\nThe paper deserves a serious referee, not a desk reject. The problem is fixable, and the underlying approach is worth engaging with. I would ask the authors to correct the objective, rerun the ablations to confirm the results, add error bars, and tone down the speed claim.\n\nThis is a paper for people working on fast generation or audio synthesis. Once the sign is fixed, I would cite it. For now, treat the equations as provisional.","headline":"Promising method and strong experiments, but the core objective in Eq. (5)/(8) has an inverted sign that contradicts the prose; the paper is not reproducible as written.","tokens_in":11747,"tokens_out":5325,"would_cite":false,"duration_ms":48059,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A relativistic-contrastive post-training objective converts a pre-trained rectified flow text-to-audio model into an eight-step generator, matching distilled quality while preserving diversity and dropping classifier-free guidance.","keywords":["text-to-audio generation","adversarial post-training","rectified flow","relativistic adversarial loss","contrastive discriminator","classifier-free guidance","ping-pong sampling","conditional diversity"],"falsifier":"Run ARC with and without the contrastive discriminator loss and compare eight-step prompt-adherence scores on the same prompts; if removing the contrastive loss changes nothing, the claimed semantic transfer from discriminator to generator is not happening. A second check is to test the trained discriminator alone on shuffled-prompt detection: if it cannot rank correct above incorrect prompts, the contrastive training did not make it semantically aware.","tokens_in":10621,"feed_emoji":"🎧","tokens_out":9236,"duration_ms":83688,"temperature":0.7,"pith_summary":"The paper introduces Adversarial Relativistic-Contrastive (ARC) post-training, an acceleration recipe for text-to-audio rectified flow models that needs neither distillation nor classifier-free guidance. It replaces the flow model's velocity-matching loss with a relativistic adversarial loss, paired with a contrastive discriminator objective, and claims this turns a pre-trained model into an eight-step generator competitive with distillation-based accelerators. The speed payoff is large: roughly twelve seconds of 44.1 kHz stereo audio in about 75 ms on an H100 GPU and about seven seconds on a phone CPU. The authors argue this makes interactive and on-device text-to-audio realistic without the training cost, memory overhead, or diversity loss of distillation.","feed_headline":"Text-to-audio now runs 100x faster without distillation","feed_subtitle":"ARC post-training turns a pre-trained flow model into an 8-step generator in about 75 ms on an H100.","key_machinery":"The load-bearing object is the ARC objective: a relativistic adversarial game over same-prompt real/generated audio pairs, defined in the paper as $\\min_\\phi \\max_\\psi \\left[L_{\\mathrm{R}}(\\phi,\\psi) + \\lambda L_{\\mathrm{C}}(\\psi)\\right]$. $L_{\\mathrm{R}}$ is a relativistic logistic loss that rewards the generator for producing samples judged more real than their paired real sample, and $L_{\\mathrm{C}}$ is a contrastive loss that trains only the discriminator to separate correct from shuffled audio-prompt pairs. The claimed interaction between the two, with the discriminator also initialized from the pre-trained flow model, is what substitutes for distillation and classifier-free guidance. Ping-pong sampling carries the final eight-step trajectory by alternating denoising and re-noising at decreasing noise levels.","core_discovery":"ARC post-training is a min-max objective $L_{\\mathrm{R}} + \\lambda L_{\\mathrm{C}}$ in which a text-conditioned discriminator compares pairs of noised real audio and noised generated audio that share the same prompt. The relativistic loss $L_{\\mathrm{R}}$ pushes each generated sample to score as more real than its paired real sample, while the contrastive loss $L_{\\mathrm{C}}$ additionally trains the same discriminator to score correct audio-prompt pairs above shuffled ones. The paper argues that this semantically aware discriminator makes the relativistic signal carry prompt adherence, so classifier-free guidance can be dropped; with ping-pong sampling the post-trained generator runs in eight steps. Across objective metrics and listening tests, the result is competitive with an eight-step distillation baseline while retaining substantially more generative diversity, and it is roughly one hundred times faster than the original hundred-step model.","pith_inferences":["Nothing in the ARC objective is audio-specific, so the same recipe should transfer to other conditional flow and diffusion models, such as image or video generators; that transfer is an extension the paper does not claim.","The paper does not train the generator with the contrastive loss, so a natural untested modification is adding a small contrastive term to the generator's update and checking whether prompt adherence improves further without eroding diversity.","The paper reports audio-to-audio style transfer by seeding ping-pong sampling with a recording; a testable extension would be systematic evaluation of this as a beat-alignment or voice-control mechanism.","The same-prompt diversity metric introduced here could serve as a standard evaluation companion for any conditional generative model that reports both fidelity and diversity."],"forward_implications":["A pre-trained rectified flow text-to-audio model can be accelerated to eight sampling steps without distillation, avoiding the need to generate and store teacher trajectory-output pairs.","Because classifier-free guidance is removed, inference needs only a single text-conditioned forward pass per step, which cuts peak memory and is what makes CPU-based edge-device deployment practical.","Diversity is preserved better than with distillation-based acceleration, as measured by the paper's new same-prompt conditional diversity score and by listening tests.","The reported timings, about 75 ms for roughly twelve seconds of 44.1 kHz stereo audio on an H100, put text-to-audio latency in a range suitable for interactive creative tools."],"supporting_citations":[{"why":"Supplies the relativistic adversarial loss formulation that ARC extends to conditional flow post-training.","marker":"[30]"},{"why":"Introduces the relativistic discriminator idea that grounds the paired real-generated objective.","marker":"[31]"},{"why":"Provides the distillation-based acceleration baseline, the disentangled noise schedules, and the discriminator-initialization scheme that ARC builds on.","marker":"[18]"},{"why":"Provides the pre-trained latent audio model and data recipe used as the starting point for ARC post-training.","marker":"[4]"},{"why":"Defines the rectified flow pretraining objective and the shifted logit-normal noise distribution used for the discriminator.","marker":"[6]"},{"why":"Supplies the audio-text contrastive learning formulation behind the contrastive discriminator loss and one of the evaluation metrics.","marker":"[33]"},{"why":"Documents the diversity and over-saturation costs of classifier-free guidance that ARC removes.","marker":"[23]"},{"why":"Earlier adversarial post-training baseline that reported limited gains, motivating the stronger ARC combination.","marker":"[24]"},{"why":"Earlier adversarial post-training work that required distillation initialization, a requirement ARC removes.","marker":"[25]"},{"why":"Introduces the alternating denoise-renoise refinement strategy that ARC adopts as ping-pong sampling.","marker":"[14]"}],"fun_headline_variants":["ARC post-training makes text-to-audio 100x faster","No-distillation adversarial post-training for fast audio","Adversarial post-training slashes audio generation latency","100x speedup in text-to-audio with ARC post-training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that making the discriminator good at telling correct from shuffled audio-prompt pairs also makes the generator follow prompts better, even though the generator is never trained on that pairing signal directly.","fun_headline_variants_meta":{"raw":{"variants":["ARC post-training makes text-to-audio 100x faster","No-distillation adversarial post-training for fast audio","Adversarial post-training slashes audio generation latency","100x speedup in text-to-audio with ARC post-training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000574,"raw_usage":{"total_tokens":2678,"prompt_tokens":881,"completion_tokens":1797,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":497,"completion_tokens_details":{"reasoning_tokens":1730}},"tokens_in":497,"tokens_out":1797,"duration_ms":12750,"temperature":1.0,"reasoning_tokens":1730,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:02:26.817584+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ARC with and without the contrastive discriminator loss and compare eight-step prompt-adherence scores on the same prompts; if removing the contrastive loss changes nothing, the claimed semantic transfer from discriminator to generator is not happening. A second check is to test the trained discriminator alone on shuffled-prompt detection: if it cannot rank correct above incorrect prompts, the contrastive training did not make it semantically aware.","supporting_citations":[{"cited_title":"The gan is dead; long live the gan! a modern baseline gan,","cited_arxiv_id":null,"evidence_quote":"Supplies the relativistic adversarial loss formulation that ARC extends to conditional flow post-training."},{"cited_title":"Presto! distilling steps and layers for accelerating music generation","cited_arxiv_id":null,"evidence_quote":"Provides the distillation-based acceleration baseline, the disentangled noise schedules, and the discriminator-initialization scheme that ARC builds on."},{"cited_title":"Scaling rectified flow transformers for high-resolution image synthesis,","cited_arxiv_id":null,"evidence_quote":"Defines the rectified flow pretraining objective and the shifted logit-normal noise distribution used for the discriminator."},{"cited_title":"Ufogen: You forward once large scale text-to-image generation via diffusion gans,","cited_arxiv_id":null,"evidence_quote":"Earlier adversarial post-training baseline that reported limited gains, motivating the stronger ARC combination."},{"cited_title":"Consistency models,","cited_arxiv_id":null,"evidence_quote":"Introduces the alternating denoise-renoise refinement strategy that ARC adopts as ping-pong sampling."}],"review_version":1}