Omni-modal large language models deployed in real-world environments encounter external noise that can interfere with their perception and understanding of multimodal inputs. We study their robustness in audio-visual understanding, focusing on question answering under environmental noise and competing speech. The challenge is to resist acoustic interference while preserving useful audio evidence. On-policy distillation provides dense teacher feedback on student-generated responses, but uniform token weighting does not explicitly prioritize positions affected by acoustic interference. We introduce OP-CAD (On-Policy Clean-Audio Distillation), a curriculum-based privileged self-distillation framework for robust audio-visual understanding. Training progresses from mild to severe environmental noise and competing speech, with selective token-level supervision at each stage. The student generates responses from corrupted audio-visual input, while a frozen teacher uses clean audio and the verified answer to supervise the same response prefixes. To allocate this supervision, OP-CAD compares teacher predictions under clean, corrupted, and visual-only contexts without revealing the answer. These matched comparisons measure sensitivity to audio removal and corruption; a bounded weighting rule emphasizes positions identified by either signal while retaining supervision throughout the response. OP-CAD outperforms the compared methods across all evaluated noise conditions. Paired analyses further show improved preservation of clean-correct answers under strong interference, with no observed aggregate clean-accuracy penalty. These results demonstrate the value of directing clean-teacher supervision toward acoustically sensitive predictions for robust audio-visual reasoning.
Figure 2: Overview of the OP-CAD training framework.
Figure 3: Token-level supervision allocation in OPSD and OP-CAD.
Daily-Omni
OmniVideoBench
Method
Clean
Env.
Speech
Clean
Env.
Speech
Base
67.75
61.38
57.84
37.40
36.70
34.80
SFT warm-up
68.76
61.74
59.82
38.10
36.80
35.93
+ SFT
68.42 ↓ 0.34
61.88 ↑ 0.14
59.12 ↓ 0.70
39.00 ↑ 0.90
36.23 ↓ 0.57
36.20 ↑ 0.27
+ GRPO
68.34 ↓ 0.42
62.91 ↑ 1.17
59.93 ↑ 0.11
41.50 ↑ 3.40
37.87 ↑ 1.07
36.87 ↑ 0.94
+ OPSD
69.34 ↑ 0.58
63.49 ↑ 1.75
60.62 ↑ 0.80
41.40 ↑ 3.30
37.40 ↑ 0.60
35.13 ↓ 0.80
Table 1: Full test accuracy (%). Env./Speech: three-SNR averages. Bold: best result. Arrows: changes from SFT warm-up in percentage points, using displayed scores.
Daily-Omni
Cond.
Method
Question Type
Video Duration
Avg.
AV Align
Comp.
Ctx. Und.
Evt. Seq.
Infer.
Reas.
30s
60s
Clean
Base
58.82
77.86
60.62
60.13
81.82
81.14
68.01
67.45
67.75
+ SFT
61.34 ↑ 2.52
77.10 ↓ 0.76
64.77 ↑ 4.15
57.84 ↓ 2.29
83.77 ↑ 1.95
80.57 ↓ 0.57
68.62 ↑ 0.61
68.18 ↑ 0.73
68.42 ↑ 0.67
+ GRPO
59.24 ↑ 0.42
79.39 ↑ 1.53
63.73 ↑ 3.11
59.80 ↓ 0.33
82.47 ↑ 0.65
80.00 ↓ 1.14
70.48 ↑ 2.47
65.82 ↓ 1.63
68.34 ↑ 0.59
+ OPSD
56.72 ↓ 2.10
77.86 ↔ 0.00
68.91 ↑ 8.29
60.78 ↑ 0.65
84.42 ↑ 2.60
82.29 ↑ 1.15
71.41 ↑ 3.40
66.91 ↓ 0.54
69.34 ↑ 1.59
Table 2: Daily-Omni accuracy (%) by question type and duration. Arrows: changes from Base within each condition in percentage points, using displayed scores. Bold: best per column.
OmniVideoBench
Cond.
Method
Audio Type
Video Duration
Avg.
Music
Sound
Speech
(0,1] min
(1,5] min
(5,10] min
(10,30] min
Clean
Base
30.77
32.65
39.11
49.12
35.29
36.40
33.33
37.40
+ SFT
32.97 ↑ 2.20
37.41 ↑ 4.76
40.03 ↑ 0.92
47.95 ↓ 1.17
39.41 ↑ 4.12
35.96 ↓ 0.44
35.25 ↑ 1.92
39.00 ↑ 1.60
+ GRPO
36.26 ↑ 5.49
36.73 ↑ 4.08
43.04 ↑ 3.93
49.12 ↔ 0.00
40.59 ↑ 5.30
39.47 ↑ 3.07
39.46 ↑ 6.13
41.50 ↑ 4.10
+ OPSD
28.57 ↓ 2.20
40.82 ↑ 8.17
43.04 ↑ 3.93
48.54 ↓ 0.58
42.06 ↑ 6.77
40.79 ↑ 4.39
36.40 ↑ 3.07
41.40 ↑ 4.00
Table 3: OmniVideoBench accuracy (%) by audio type and duration. Arrows: changes from Base within each condition in percentage points, using displayed scores. Bold: best per column.
Daily-Omni
OmniVideoBench
Variant
Clean
Env.
Speech
Clean
Env.
Speech
OPSD
69.34
63.49
60.62
41.40
37.40
35.13
Audio-rem. proxy
69.34
64.47
61.71
40.70
39.63
38.50
Corr.-sens. proxy
68.76
65.41
62.13
43.80
39.73
39.00
OP-CAD
70.09
66.97
63.55
41.60
40.23
38.93
Table 4: Accuracy (%) with single-signal and combined allocation. Bold marks the best result.
Training schedule
Daily-Omni
OmniVideoBench
Clean
Env.
Speech
Clean
Env.
Speech
Randomly mixed
69.93
65.16
62.67
42.30
38.27
37.13
Curriculum learning
70.09
66.97
63.55
41.60
40.23
38.93
Table 5: Curriculum order ablation: accuracy (%) under matched training budgets.
Figure 4: Prediction stability at 0 dB: harm rate (left) and clean-correct retention (right).
Figure 5: Token-level supervision allocation on an OP-CAD training rollout.
While large audio-language models have achieved remarkable progress in auditory perception, they still lag behind text-based large language models in deep logical reasoning, primarily due to the scarcity of high-quality audio reasoning data. To bridge this gap, we propose X3-OPD, a cross-modal on-policy distillation framework that transfers reasoning capabilities from a powerful text teacher to an audio-language student. During training, the student generates reasoning trajectories conditioned on its own acoustic perception, while the teacher provides token-level guidance using matched textual inputs and verified answers. We further construct a three-tier symmetric corpus covering textual reasoning rendered into speech, audio-event reasoning grounded in complex acoustic scenes, and spoken-dialogue reasoning involving paralinguistic cues. This design extends cross-modal distillation beyond textually recoverable content to reasoning grounded in non-linguistic events, prosody, and conversational context. Experiments on MMSU, MMAU, BIG Bench Audio, and MMAR demonstrate that X3-OPD substantially improves audio-grounded reasoning and chain-of-thought quality while largely preserving the model's existing capabilities under domain shift.
Omni-modal models provide a unified interface for text, images, and audio. However, improving these abilities together remains difficult, as post-training on pooled multimodal data often fails to preserve the strengths of modality teachers. On-policy distillation (OPD) has recently become popular in model post-training. It samples responses from the current student and compares the teacher's and student's next-token distributions along those responses, yielding dense supervision while reducing the mismatch between training and inference. Despite these advantages, standard OPD does not readily extend to several modality teachers. Their guidance may favor conflicting changes to the shared model, while matching each teacher's next-token distribution can prevent the student from moving beyond that teacher. To address these challenges, we propose On-Policy Omni Distillation (OPOD), which consolidates text, image, and audio teachers into one omni model. OPOD routes each response to the corresponding teacher, controls the teachers independently, and applies guidance only when the teacher assigns a higher probability to the generated token. The selected teacher also evaluates answer confidence and whether the reasoning increases support for the answer. Extensive experiments on twelve benchmarks show that OPOD achieves the best average at three model scales, reaching 70.8, 51.7, and 46.2 and outperforming the strongest comparator by 2.1, 1.8, and 1.7 points. At 30B, it surpasses the base model and pooled RL training on all twelve benchmarks, and ranks first or second on eleven even when the teachers are included. Only the student is retained for deployment.
Large Audio Language Models (LALMs) remain vulnerable to acoustic noise, which can obscure task-relevant evidence and produce unreliable responses. We propose EchoDistill, a noisy-to-clean self-distillation framework that uses clean audio as privileged information during post-training. A noisy-input student samples candidate responses reflecting its inference-time behavior, while a frozen copy of the same backbone processes the corresponding clean audio. EchoDistill combines masked response-token distillation, task-gated consistency shaping, and teacher-referenced group-relative optimization to align noisy-input generation with clean-conditioned semantics. Only the student is retained at inference time, introducing no additional inference cost. Across three LALM backbones and three audio domains at -10dB, EchoDistill improves average noisy-input accuracy by 1.63 percentage points over the strongest baseline. On Qwen2.5-Omni, it raises noisy-input accuracy from 59.33% to 62.94%, while clean-audio accuracy increases from 76.56% to 77.56%. Replacing matched audio with random, shuffled, or silent inputs reduces accuracy by 3.08-6.42 points, confirming that matched acoustic evidence contributes to its predictions. Additional evaluations show improvements on held-out additive noises and external benchmarks, while revealing that these gains do not reliably extend to non-additive distortions. These results demonstrate robust post-training improvements under severe additive noise without sacrificing clean-audio capability across diverse tasks.