- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
SONATINE separates monaural semantic understanding from multichannel spatial measurement. The Audio LLM receives the FOA W channel and question through its existing interface, then selects a spatial tool and time window. Tools access the complete multichannel recording and return direction, trajectory, distance, or spatially cropped audio. The central design is to select windows in which the queried sound is clear and interference from other sources is limited.
Critical strengths
-
In the duration-matched ablation, target-informative supervision exceeds high-energy-window supervision by 5.47 percentage points under heavy overlap.
-
The interface supports closed-source tool use and six-channel transfer through backend replacement while retaining much of the model’s general-audio capability.
Critical weaknesses
-
Window-supervision ablations are provided, but a single-round or fixed-pipeline control with the same tools and target temporal grounding is missing, leaving the additional value of iterative tool reasoning unclear.
-
Human agreement rates for the answer extractor are not reported, leaving the stability of the GRPO and overall gains uncertain.
Questions for authors
-
Could you include a pipeline baseline using the same spatial tools and target temporal grounding without multiple interaction rounds?
-
What is the agreement rate between the free-form answer extractor and human annotations?
-
What are the complete adaptation costs, including distance-tool training, trajectory synthesis, and GRPO, and the end-to-end inference latency?
-
What spatial error does the six-channel-to-FOA conversion introduce, and how does a simplified policy perform using the same six-channel backend?
5 comments
Response to Reviewer cTy5 (2/3)
W2/Q2. Agreement between the extractor and human annotation
We would first like to describe the role of the extractor in more detail. The extractor in SONATINE does not decide whether an answer is correct. Without seeing the gold answer, the extractor only extracts what the free-form answer claims, such as a source name, left or right, a time interval, or a transcript. The comparison with the gold answer is performed by deterministic code, and numeric tasks such as angle, distance, and onset are scored by rules without an LLM. The extractor therefore only reads the answer and does not score it. Because it does not know the correct answer, it has little room for bias toward particular answers.
Nevertheless, as the reviewer points out, the paper did not report how accurate the extraction is, so we measured it by direct comparison with human-made labels. On the evaluation side, for the 11 tasks scored with the extractor, we compared 200 SO-Bench answers from three systems with different answer styles (Spatial-Omni-7B, SONATINE 7B, and Gemini with prompting only). On the GRPO reward side, we compared 200 completions from the 7 tasks for which the extractor gave the reward during the GRPO training. One annotator wrote down what each answer claims without seeing the gold answer, the extractor output, the system that produced the answer, or whether the item came from evaluation or training. The human labels were converted into scores by the same scoring code as the extractor outputs.
Result 2: extractor accuracy against human labels.
| SO-Bench answers (evaluation, Qwen3.8-27B) | GRPO completions (reward, Qwen3-4B) | |
|---|---|---|
| Items | 200 | 200 |
| Agreement with human label | 97.0% | 98.0% |
| Pass/fail decision accuracy | 99.0% | 99.5% |
| Cohen's kappa (pass/fail) | 0.980 | 0.990 |
| Mean score (extractor / human) | 0.463 / 0.460 | 0.430 / 0.435 |
The extractor output agreed with the human label on 97.0% and 98.0% of the items, and the pass/fail decision differed for only 3 of the 400 items. In two of them the extractor misread the answer, once giving credit to a wrong answer and once giving zero to a correct one. In the remaining one, the human and the extractor read almost identical time intervals, which fell on either side of the pass threshold. The difference in mean score was 0.003 to 0.005, which is too small to change the results.
We also checked, using the metrics logged during training, whether the signal from the extractor was stable during GRPO training. If the policy had drifted toward answers that the extractor cannot read or answers that fool the extractor, the fraction of answers whose content the extractor could not read (unclear) would have risen, or the reward would have approached 1.
Result 3: the extractor's reward signal during GRPO (extractor-scored tasks).
| Steps | Reward | Extractor unclear |
|---|---|---|
| 1-50 | 0.487 | 1.87% |
| 51-100 | 0.495 | 1.42% |
| 101-150 | 0.511 | 1.22% |
| 151-200 | 0.513 | 0.66% |
As training progressed, the unclear rate decreased from 1.9% to 0.7%, the reward rose slowly without saturating, and the answer length remained constant. Together with the GRPO column of Result 2, this supports that the gain from GRPO does not come from extractor errors or exploits. We will add the role of the extractor, the blind annotation procedure, and the two results above to App. E.3.
Q3. Adaptation cost and end-to-end latency
We summarize, stage by stage, the cost of building SONATINE 7B from Qwen2.5-Omni-7B, including the training of the distance tool.
Result 4: adaptation cost.
| Stage | Data | Compute |
|---|---|---|
| Trajectory construction | 14,990 trajectories | CPU, a few minutes per task |
| Rationale synthesis (Qwen3.6-27B) | 14,990 | ~4 GPU-h, Pro 6000 |
| Distance regressor | 199K windows, 38.7M params | ~18-20 GPU-h, Pro 6000 |
| SFT 7B | 14,990, 2 epochs | 5 h 52 min, Pro 6000 ×2 |
| GRPO 7B | 200 steps | ~24 h, H100 ×4 |
We also measured the end-to-end inference latency. On 315 questions randomly sampled from SO-Bench, we measured the time and the number of generated tokens per question, running sequential inference with batch size 1 on a single RTX Pro 6000.
Result 5: per-question inference cost.
| Per question | SONATINE 7B |
|---|---|
| Seconds (mean / median / p90) | 4.75 / 4.18 / 8.54 |
| Tool execution (mean) | 0.33 |
| Tokens (mean / median / p90) | 297 / 253 / 548 |
| Tool calls / model replies | 1.88 / 2.88 |
Most of the time is spent on generating tokens over multiple tool calls, and tool execution takes only a small part of it. We will add both tables to App. D.
Response to Reviewer cTy5 (3/3)
Q4. Error of the six-channel-to-FOA conversion and simplified policies
In short, the six-channel-to-FOA conversion loses both directional information and the spectral characteristics of the signal, and even with the same six-channel backend, an untrained policy did not reach the performance of SONATINE. First, to examine what the conversion preserves independently of any model, we passed a single plane wave without noise or reverberation through the six-channel array from every direction of the measured array transfer functions, converted it to FOA, and analyzed it per frequency band. We examined two conversions. One is the full-band complex least-squares Ambisonics signal matching (ASM) that we used in the main text to produce the inputs of Spatial-Omni-7B and JAEGER. The other is a variant that is identical to ASM below 1.5 kHz and replaces the objective with magnitude least squares (MagLS) only above it. Since this is the case of a single source without noise, these numbers correspond to an upper bound on the information the conversion can preserve.
Result 6: per-band analysis of the six-channel-to-FOA conversion.
| Band | Azimuth error, median (ASM / MagLS variant) | Accuracy within 20° (ASM / MagLS variant) | W level (ASM / MagLS variant) | Dipole-to-W ratio (MagLS variant, ideal 0 dB) |
|---|---|---|---|---|
| 200-500 Hz | 4.2° / 4.2° | 100% / 100% | -0.4 / -0.4 dB | -0.6 dB |
| 0.5-1 kHz | 12.1° / 12.1° | 75% / 75% | -1.7 / -1.7 dB | 0.3 dB |
| 1-1.5 kHz | 21.2° / 21.2° | 47% / 47% | -3.5 / -3.5 dB | 0.7 dB |
| 1.5-2 kHz | 24.2° / 24.1° | 44% / 43% | -6.3 / -2.2 dB | -3.4 dB |
| 2-3 kHz | 38.7° / 19.6° | 30% / 51% | -10.9 / -1.4 dB | -4.3 dB |
| 3-4 kHz | 55.1° / 26.3° | 20% / 39% | -13.6 / -1.6 dB | -5.3 dB |
| 4-6 kHz | 48.3° / 27.7° | 25% / 38% | -13.7 / -1.2 dB | -4.7 dB |
| 6-8 kHz | 45.6° / 38.0° | 27% / 27% | -14.9 / -2.2 dB | -4.7 dB |
Direction is preserved only below 1 kHz. Above it, even for a single noise-free source, a large azimuth error of around 20° or more occurs.
The conversion also changes the spectral distribution of the signal. ASM attenuates the W channel by 11 to 15 dB above 2 kHz, producing a signal with its high frequencies cut, whereas the MagLS variant keeps the level but discards the high-frequency phase, so that the summed energy of the dipoles (X, Y, Z channels) becomes 3 to 5 dB smaller than W. In real FOA, this ratio is exactly 0 dB for a single plane wave. That is, even for a single clean source, the converted FOA systematically has a dipole deficit that does not exist in real FOA. An FOA-trained encoder therefore receives inputs outside its training distribution whichever conversion is used, and the failures of the baselines can be regarded as stemming from this.
We believe this is a limitation of the internalized design. To be used on a new array, an FOA-trained model must either be retrained or receive input converted to FOA, and as shown above, converting the signals of a real six-channel array to FOA incurs a loss. SONATINE, on the other hand, keeps the policy as is and replaces only the tool backend with one for six channels, so no conversion is involved.
Next, we compared simplified policies on the same six-channel backend, with the same 350 questions and the same scoring. Base 7B is Qwen2.5-Omni 7B without any additional training, and Base 7B + tools is the same model given the same tools as SONATINE by prompting. Lower is better for MAE and WER.
Result 7: simplified policies on the six-channel backend.
| Policy | Single-speaker azimuth MAE (°) | Single-speaker azimuth acc@20 | Left-right acc. | Order-conditioned azimuth acc@20 | Location-conditioned WER | Distance MAE (m) |
|---|---|---|---|---|---|---|
| Base 7B, no tools | 21.7 | 0.614 | 0.500 | 0.371 | 0.637 | 1.177 |
| Base 7B + tools | 27.7 | 0.471 | 0.486 | 0.229 | 0.672 | 1.034 |
| SONATINE 7B | 10.1 | 0.857 | 0.600 | 0.471 | 0.616 | 0.329 |
When the untrained base model is given the same tools by prompting, it falls below even the base model without tools on every task except distance. The performance on EasyCom can therefore be attributed to the learned policy rather than to the six-channel backend. We will add the per-band conversion analysis and the policy comparison on the same backend to the paper.
We thank the reviewer once again for the helpful questions.
Response to Reviewer cTy5 (1/3)
We thank the reviewer for the specific questions on the contribution of iterative tool reasoning, the reliability of the evaluation, the adaptation cost, and format transfer. We report below the additional experiments and analyses we conducted to answer each question.
W1/Q1. The additional value of iterative tool reasoning
To isolate the effect of iterative tool reasoning, we ran a single-round experiment. With the same SONATINE 7B policy and the same tools, we kept the first tool call and its window unchanged and forced the model to answer immediately after receiving its result, and evaluated on the full SO-Bench.
Result 1: the paper model with the second tool call blocked.
On the direct-estimation tasks, a single round already achieved performance similar to iterative calling. In contrast, on the compare tasks and the multi_hop task, which deal with relations between sources, the single-round performance was much lower.
This difference shows that two kinds of abilities are separated. The measurement itself, such as direction or distance, is handled by adaptive window selection. Because the first call of the model already uses a window in which the target is clearly heard, a single measurement is as accurate as iterative calling. That is, the window-selection ability resides in the learned policy, and using it does not require multiple rounds.
Where iterative tool reasoning contributes, in contrast, is spatial understanding that requires combining measurements to produce an answer. The single-round setting lowers the accuracy of compare_distance, compare_elevation, relative_left_right, and multi_hop by 7 to 21 points. The first three require measuring two sources separately and comparing them, and multi_hop requires gathering the directions of several sources and selecting the one that satisfies a condition. These tasks do not end with measuring a single source. The spatial information of several sources must be gathered and the relations among them inferred. Such reasoning cannot be replaced by a single measurement and is possible only with iterative tool reasoning that interleaves measurement and reasoning. Moreover, SONATINE outperforms Spatial-Omni-7B on three of these tasks (relative_left_right, compare_elevation, and compare_distance), and this advantage disappears or shrinks considerably in the single-round setting. Because a measurement in SONATINE is a tool call, the model can explicitly choose a target-informative window for each source the question refers to and obtain a separate measurement for each. The results show that this provides a real benefit and that iterative tool reasoning is what makes it possible.
In summary, in the SONATINE framework, adaptive window selection is responsible for the accuracy of the measurements, and iterative tool reasoning is responsible for the reasoning that combines the measurements. We will add this experiment and its interpretation to the paper.