acceptodds
Desk rejected at ICLR 2027

AOI: All-in-One Audio Deepfake Detection via Evidence-Grounded Rationales and On-Policy Distillation

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Generative audio models can now synthesize speech, music, and environmental sounds that are difficult to distinguish from genuine recordings, yet existing audio deepfake detectors are typically trained for a single audio type and a narrow family of forgery algorithms, forcing practitioners to deploy a collection of type-specific models. We propose AOI, an all-in-one audio deepfake detection framework that covers multiple audio types and multiple deepfake types with a single AudioLLM. AOI first trains three audio-type-specific experts by supervised fine-tuning (SFT) on evidence-grounded chain-of-thought (CoT) targets constructed by a separate multimodal reasoning model. This annotator receives measured acoustic descriptors when writing the targets, but neither the experts nor the final student is given those measurements as input, so the detector must perceive the evidence it recites. The experts are then distilled into one student by on-policy distillation with reward extrapolation: the student generates its own CoT trajectories and the domain experts supply dense token-level supervision on them. We evaluate on speech, music, and environmental-audio benchmarks for fine-grained forgery classification and all-type generalization ( held-out clips). AOI classifies five-way over three audio types with a single checkpoint at weighted accuracy, within of three per-type label-only experts that require oracle routing, and recovers a substantial share of the unification cost by on-policy distillation; among all-in-one rationale methods, it achieves the highest five-way accuracy.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
audio deepfake detection, on-policy distillation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_PL4vOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AOI, an all-in-one AudioLLM framework for deepfake detection across speech, music, and environmental audio. The method first trains three type-specific experts using evidence-grounded rationales constructed with privileged acoustic measurements, while the detectors themselves only receive raw audio. These experts are then distilled into a single model using on-policy distillation with reward extrapolation (ExOPD). AOI performs five-way classification over genuine, synthetic, edited, codec-reconstructed, and vocoder-reconstructed audio. The final model achieves 97.95% weighted five-way accuracy, approaching the 99.33% performance of three separately routed label-only experts while producing explanatory rationales with a single checkpoint.

Critical strengths
  • Unifying multiple audio types into a single detector without requiring an explicit type router is a useful direction, particularly as audio generation increasingly extends beyond speech.
  • The evidence-grounded rationale construction is well motivated
  • The paper separately investigates the cost of rationale training versus generative decoding and attempt to quantitatively evaluate rationale grounding. This makes the empirical analysis more informative than simply comparing final classification numbers
Critical weaknesses
  • The benefit of the proposed Stage-2 distillation appears relatively modest. ExOPD improves weighted five-way accuracy from 97.71% to 97.95%. While statistically significant on a very large test set, the practical magnitude is small, making it unclear whether the additional complexity of training three experts followed by on-policy distillation is sufficiently justified.
  • The paper does not convincingly isolate the contribution of ExOPD itself. The main comparison is essentially mixed SFT versus ExOPD. Important controls such as standard OPD (λ=1), alternative distillation strategies, and a sweep over the extrapolation coefficient are missing.
  • The rationale evaluation remains somewhat weak. The main evidence relies on LLM-as-a-judge evaluation, and the grounding analysis is performed on only 500 examples and four acoustic descriptors. Moreover, correlation between values mentioned in generated explanations and measured descriptors does not necessarily establish that those explanations faithfully reflect the evidence actually used for classification.
  • The experiments use path-disjoint held-out splits, but the paper’s motivation emphasizes robustness across audio types and forgery algorithms. Evaluation on genuinely unseen generators, codecs, datasets, or recording conditions would provide stronger evidence that AOI learns transferable forensic cues rather than dataset-specific characteristics.
Questions for authors
  1. How much of the Stage-2 improvement is actually attributable to reward extrapolation?
  2. Does AOI generalize to completely unseen forgery methods or datasets? Given that cross-generator generalization is one of the central challenges in audio deepfake detection, an out-of-distribution evaluation is necessary.
  3. How faithful are the generated rationales to the model’s actual decision process? The descriptor correlations show that the student can estimate certain acoustic properties, but do not establish that those properties causally influence its prediction.
  4. Given the relatively small improvement from ExOPD (97.71% → 97.95%), what is the computational cost of Stage 2, including training three experts and performing on-policy distillation?
2 comments

Thank you for the careful distinction between the utility of unification, the Stage-2 gain, and the evidence needed for stronger generalization and faithfulness claims.

Due to limited GPU resources, we are unable to conduct additional experiments during the rebuttal period. We focus on clarifying the method (§3) and existing results (§4) and providing implementation details in Appendices F and G. We welcome further discussion of AOI's design and experimental setting.

1. Magnitude and attribution of the Stage-2 gain. Weighted five-way accuracy improves from 97.71% to 97.95%, or 0.24 percentage points (§4.3, Table 1). The reported 26% recovery uses the 0.92-point gap between mixed SFT (97.71%) and the rationale-producing per-type experts (98.63%): 0.24 / 0.92 ≈ 26%.

Section 4.3 reports a 2.12-point gain on speech codec reconstruction, and Appendix E describes 127 speech codec clips corrected from genuine to codec-reconstruction. These observations locate the gains in codec-reconstruction detection. The comparison evaluates the Stage-2 procedure as a whole; a λ = 1 control is not included. The λ = 1.25 choice follows the original work's recommended setting, and the reported checkpoint uses 14% of the planned schedule, at which point training was stopped to stay within the available compute budget (Appendix G.2).

2. Training cost and whether Stage 2 is justified. AOI (§§3.3–3.4) requires training three type-specific rationale experts and a mixed SFT model, followed by on-policy student distillation. Each ExOPD example entails student generation and scoring by the relevant frozen expert and fixed reference, in addition to the student update. These operations account for the additional Stage-2 training cost. The practical motivation is consolidating specialized knowledge into one rationale-producing checkpoint.

3. Unseen forgery methods and datasets. Our experiments evaluate unification across speech, music, and environmental audio (§4.1). The reported path-disjoint splits and shortcut controls define this evaluation setting (§§3.3, 4.2; Appendix B). A dedicated unseen-generator or external-dataset evaluation is not included.

4. Faithfulness of the rationales. The 500-clip descriptor study (Appendix D) tests agreement with signal measurements across four acoustic features. The judge study (§4.4) provides complementary preference evidence. Together, these results support partial acoustic grounding and the ability to inspect a model's stated rationale. Causal attribution of the model's decision was not tested.

Thank you for the response and clarifications. However, my main concerns remain unresolved. The contribution of reward extrapolation cannot be isolated without a standard OPD baseline, particularly given the modest Stage-2 improvement and unquantified training cost. Moreover, the evaluation demonstrates unification across audio types but does not establish generalization to unseen generators or external datasets. The faithfulness of the generated rationales to the model's actual decisions also remains unverified. These issues concern the paper's central methodological claims and empirical support. I therefore maintain my original rating.

Reviewer_GPHwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper targets the current challenge of domain complexity (speech, environment, music, etc.) in audio deepfake detection and proposes an “all-in-one” audio deepfake detection framework through on-policy distillation of expert knowledge into a single student model. By combining evidence-grounded supervised fine-tuning and on-policy knowledge distillation, the framework seeks to show that knowledge from several specialized detectors can be consolidated into a single, explainable AudioLLM without sacrificing too much detection performance. Findings from experiment validations demonstrate the feasibility of generating explanations while retaining high classification accuracy.

Critical strengths

1.The research problem (addressing specialized deepfake detector dependence on audio-specific architectures or routing strategies) is well motivated and relevant to audio deepfake detection research community, and the move beyond binary classification toward identifying manipulation categories makes the problem more interesting 2.The overall architecture is conceptually sound. While the individual components of this framework do not involve fundamentally new techniques, their integration into an all-type audio deepfake detection pipeline is interesting. 3.The authors are explicitly aware of and honest about the shortcut issue, which helps contextualize the contributions made in this paper.

Critical weaknesses

1.The per-type performance comparison against HoliAntiSpoof, as presented in the paper, is misleading. For HoliAntiSpoof, the authors used the released checkpoint that was trained for speech antispoofing, and domain drift therefore accounts for much of the gap between AOI and HoliAntiSpoof figures. While the authors acknowledge this, the inclusion of this comparison seems disingenuous. 2.In absence of the AOI vs HoliAntiSpoof comparison, however, the claim that “AOI is the strongest all-in-one rationale method” is somewhat questionable in the sense that the performance increases are minor and would require confidence intervals to contextualize. 3.The question of generalization ability of the framework is not adequately addressed. The paper’s motivation emphasizes the need for detectors capable of handling varied and evolving generative audio techniques. However, the experimental design primarily evaluates held-out samples originating from the same benchmark and generation ecosystems used to construct training data. 4.The ablation studies seem somewhat coarse (audio-type prompting and decoding ablations), but without finer control variables, such as ExOPD lambda values, inclusion vs absence of rationales, etc.

Questions for authors

1.How does explanation quality change when the rationale-generation model receives only the audio and ground-truth label, without the measured acoustic descriptors? 2.Is it possible to provide stronger evidence that the acoustic artifacts mentioned in a rationale actually influence the model's predicted class, rather than merely producing a plausible explanation? 3.How does ExOPD compare with standard OPD and continued SFT under roughly equal computational budgets? 4.Have the authors considered comparing AOI against a router + smaller specialized detector setup? I suspect that the latter setup could lower the computation cost further.

1 comment

Thank you for recognizing the motivation and integration of the framework, and for highlighting comparison fairness and the scope of the explanation claims.

Due to limited GPU resources, we are unable to conduct additional experiments during the rebuttal period. We focus on clarifying the method (§3) and existing results (§4) and providing implementation details in Appendices F and G. We welcome further discussion of AOI's design and experimental setting.

1. HoliAntiSpoof comparison and the strength of the claim. The released speech checkpoint's music and environmental results measure cross-domain transfer; Table 1's mixed-data rows (§4.3) provide the corresponding all-in-one comparison. The relevant all-in-one comparison is therefore 97.95% versus 97.91%, a difference of 0.04 percentage points. We correct §4.3 to use this mixed-data comparison and remove the significance claim based on the released checkpoint. AOI attains the highest numerical five-way accuracy among the evaluated all-in-one rationale systems in §4.3, Table 1.

2. Rationales without acoustic measurements. Section 4.4 reports an annotation-side bandwidth correlation of −0.076 without supplied measurements and +1.000 with them. This quantifies the numerical consistency of the annotation targets. Student-side grounding is assessed separately on audio-only inputs in Appendix D.

3. Do cited artifacts influence the prediction? Measurement-free detector inputs (§3.2) prevent direct copying from a supplied measurement block, and the Appendix D correlations support partial signal consistency. Causal attribution of the category prediction was not tested. The DailyTalkEdit example in Appendix E also illustrates why a correct observation can support an incorrect interpretation. These results evaluate partial grounding and diagnostic usefulness.

4. ExOPD versus OPD and continued SFT at equal compute. Section 4.3 compares mixed SFT with the implemented ExOPD stage. It contains neither standard OPD nor compute-matched continued SFT. The 97.71% → 97.95% result evaluates the implemented ExOPD stage as a whole. Our use of ExOPD follows prior work; our contribution is its use in consolidating the type-specific teachers into one audio-only detector (§3.4).

5. Router plus smaller specialized detectors. We agree that this is a valid alternative and may reduce training cost. For deployment across the three audio types, it maintains three specialized detectors plus a router, whereas AOI serves one checkpoint (§3.1). Only the selected detector needs to execute after routing. Actual inference latency and memory depend on detector sizes, routing overhead, and model-loading strategy. AOI reduces the number of deployed model components and removes explicit inference-time routing. AOI uses type-based teacher selection during training (§3.4).

6. Generalization. The reported evaluation covers three audio types and five forgery categories on path-disjoint held-out splits (§4.2). Our source matching and duration controls (§§3.2–3.3; Appendix B) address specific shortcuts, while inverse-frequency sampling addresses training exposure. Generator-disjoint and external-dataset evaluations were not conducted. Our results support unification across the three evaluated audio types and five categories under the reported splits.

Reviewer_71p3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces AOI, an all-in-one audio deepfake detection framework built on an AudioLLM designed to detect forgeries across speech, music and environmental sound. It classifies audio as genuine, synthetic, edited, codec-reconstructed, or vocoder-reconstructed within a single unified checkpoint, eliminating the need for modality routers.

Critical strengths

Unlike all-type ADD approaches that rely on modality-specific routers and separate domain-specific models, AOI uses a single unified checkpoint across speech, music, and environmental sound.

Existing ALLM4ADD methods either hallucinate post-hoc explanations or require hand-crafted acoustic measurement pipelines during inference. The paper innovatively utilizes acoustic descriptors solely at target generation time while enforcing an audio-only prompt during both expert SFT and student inference, successfully compelling the underlying audio encoder to internalize verifiable physical features.

The authors demonstrate commendable rigor in data construction by countering trivial shortcuts (e.g., duration cues and corpus-specific acoustic fingerprints) through paired source generation and uniform length clipping. Furthermore, the paper honestly identifies the rationale cost, attributing 62% to autoregressive sequence risk and 38% to representation shift.

Critical weaknesses

While the AOI framework is attractive, the actual target generation and cleaning pipeline remains largely opaque.

  1. The paper mentions 5 descriptor categories in Sec. 3.2, claims "eleven checkable quantities" in Appendix E, yet only benchmarks 4 descriptors in Appendix D. A complete mathematical specification and listing of all acoustic descriptors is missing.

  2. The exact prompt template provided to Qwen3-Omni-Thinker is omitted.

  3. The "text-only cleaning pass" is vaguely described without disclosing whether this relies on rule-based regular expressions or an auxiliary LLM rewriter.

  4. In Sec. 4.3, the authors state that ExOPD closes 26% of the gap "after 14% of the training schedule," but fail to report the final performance upon schedule completion. It is unclear whether the model suffered from policy collapse, degenerate CoT generation, or reward hacking beyond this checkpoint.

  5. The extrapolation hyperparameter is arbitrarily fixed to λ= 1.25 without any sensitivity analysis or ablation.

  6. The reproduction of FT-GRPO in Appendix C completely collapses by step 50, predicting 100% fake, due to GPU resource downscaling, rendering the reported all-in-one FT-GRPO baseline questionable.

  7. Relying exclusively on Weighted and Macro Accuracy masks severe recall deficiencies on minority classes. Standard evaluation metrics, such as Macro-F1, are omitted.

  8. In the blind LLM evaluation, the judge is likely susceptible to jargon and length bias. Gemini 3.8 Flash lacks forensic-level acoustic spectral resolution to verify whether energy actually dropped by 3.3 dB at 2.03 kHz; it may simply favor AOI's authoritative, lab-style scientific prose over competitors. Crucially, the authors do not provide side-by-side textual outputs of FT-GRPO or HoliAntiSpoof in Appendix E for qualitative inspection.

Questions for authors
  1. Why did the authors benchmark the model at 14% of the ExOPD training schedule rather than at convergence? Did the student model experience language degeneration, repetitive loops, or format degradation during later training steps? Please provide the training loss curve and performance trajectory across the full schedule.

  2. Could the authors provide i) The full system and instruction prompt given to Qwen3-Omni-Thinker; ii) The exact definition and list of all 11 extracted acoustic descriptors; iii) The concrete implementation details of the "text-only pass".

  3. What is the sensitivity of student performance to the reward extrapolation factor λ? How does the model perform under standard on-policy distillation versus aggressive extrapolation (λ in {1.1, 1.5, 2.0})?

  4. Please report the Macro-F1 score on the held-out test sets to confirm whether partial edits and codec artifacts are reliably disentangled.

  5. Could the authors provide representative, side-by-side output examples comparing AOI, FT-GRPO, and HoliAntiSpoof on the exact same audio clips to substantiate that the LLM judge's 52% preference stems from forensic reasoning fidelity rather than superficial stylistic verbosity?

2 comments
AuthorsOpenReview ↗

Response (1/2)

Thank you for the detailed questions about reproducibility, training, and explanation evaluation.

Due to limited GPU resources, we are unable to conduct additional experiments during the rebuttal period. We focus on clarifying the method (§3) and existing results (§4) and providing implementation details in Appendices F and G. We welcome further discussion of AOI's design and experimental setting.

1. Annotation prompts, descriptors, and text-only cleaning. We have added Appendix F, “Annotation and Target-Construction Details,” to the revised manuscript. Appendix F.1, “Annotation Model, Inputs, and Exact Prompts,” provides the full system and user prompts, all template substitutions, the measurement rendering template, and decoding settings. The annotator is Qwen3-Omni-30B-A3B-Thinking; the measured-evidence configuration uses audio and raw measurements without a spectrogram image. Neither the experts nor the student receives the measurements or ground-truth label as detector input.

Appendix F.2, “Acoustic Descriptor Definitions,” specifies all 11 returned fields: ten scalar fields, including duration, and one variable-length boundary list, with preprocessing, mathematical definitions, boundary-selection rules, and stored precision. This distinguishes the field count from the five broad acoustic aspects in §3.2, the four features audited in Appendix D, and the numerical values cited by the example in Appendix E. We also clarify that the annotation prompt receives raw measurements; the optional calibration is not applied to the measurement block (Appendix F.1).

Appendix F.3, “Text-Only Rewriting and SFT Target Construction,” describes the text-only cleaning pass: it uses the same checkpoint as a text-only rewriter, with deterministic preprocessing and filtering. It includes both exact rewriting prompts, decoding settings, and the final target-construction steps. The rewriter receives only the original notes, including any numbers or label references already present in them. This process removes explicit references to a supplied answer while retaining the clip-specific analysis requested by the rewriting prompt.

2. Why only 14% of the ExOPD schedule? ExOPD uses the same training set as SFT, but its on-policy generation and teacher/reference scoring make it substantially more expensive. Training was stopped at approximately 14% of the planned schedule to stay within the available compute budget (Appendix G.2). At this checkpoint, ExOPD recovers 26% of the gap between mixed SFT and the per-type rationale experts (§4.3). No later checkpoints or full-schedule trajectory are available.

3. Choice and sensitivity of λ. We used λ = 1.25 following the settings recommended in the original ExOPD work (Appendix G.2). In §3.4, Equation (4) makes the relationship explicit: λ = 1 recovers standard OPD, while λ > 1 adds (λ − 1)(ℓ_teacher − ℓ_reference) to the token-level advantage. A λ sweep and a λ = 1 control were not conducted.

AuthorsOpenReview ↗

Response (2/2)

4. FT-GRPO reproduction and result provenance. Our reproduction used the authors' official FT-GRPO code, following the released training configuration with the resource-related differences documented in Appendix C. Specifically, we used four rather than eight GPUs, adjusted gradient accumulation to preserve the effective batch size, and ran 300 of the original 1,698 GRPO steps. By step 50, all sampled answers were fake. We report these details to make the official-code reproduction transparent.

5. Macro-F1 and minority classes. Thank you for suggesting a metric that gives equal weight to the classes. In the main experiment (§4.3; Appendix G.1), AOI Stage 2 (ExOPD) achieves a five-way Macro-F1 of 0.97310048, or 97.31%, alongside its weighted five-way accuracy of 97.95%. Macro-F1 averages the F1 scores of the five classes equally and thus complements the sample-weighted accuracy in this imbalanced setting.

6. Judge bias, grounding, and side-by-side examples. The 52% figure is AOI's win rate in a blind, three-way explanation-preference evaluation. The evaluation hides method identities, randomizes ordering, and supplies the audio and ground-truth category. It reports 0.88 self-agreement under order permutation and the same model ranking with a second judge, with win rates within ±0.04 (§4.4, Figure 2). These checks assess order sensitivity and judge dependence; length and style are not separately controlled. Numerical grounding is evaluated separately against signal measurements.

The independent measurement comparison in Appendix D provides independent, signal-based evidence: on 500 stratified clips, student-quoted values correlate with four measured features, with pooled r = 0.58–0.72. This supports partial numerical grounding; causal decision faithfulness is a separate evaluation question. Appendix E illustrates the diagnostic value of the rationale by separating an observed discontinuity from its attribution to editing.

We have added the three verbatim responses for the same synthetic speech clip in Appendix H. The example uses the first existing FT-GRPO speech inference record, with responses matched by exact audio path: AOI Stage 2, our official-code FT-GRPO reproduction checkpoint-300, and the released HoliAntiSpoof checkpoint. AOI gives the correct five-way verdict and quotes checkable spectral and temporal quantities; FT-GRPO provides qualitative frequency- and time-domain descriptions, while the stored HoliAntiSpoof response provides structured verdict, region, and method fields. This example exposes the specificity and checkability of AOI's explanation; the aggregate preference evaluation is reported separately in §4.4.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.