acceptodds
Under review as a conference paper at ICLR 2027

AURAL: Adaptive Latent Reasoning with Joint Chunk for Speech Language Models

Abstract

Model intelligence and fast response jointly shape the quality of interaction with speech language models, yet remain difficult to achieve together. Explicit chain-of-thought (CoT) improves reasoning and audio understanding, but generating intermediate reasoning tokens delays responses. Describing fine-grained acoustic cues further lengthens CoT and increases latency. Latent reasoning can reduce this overhead, yet existing methods often trail CoT and remain limited by single-path supervision and reasoning budgets that do not adapt to problem difficulty. We introduce **AURAL**, which models a distribution over multiple plausible reasoning continuations in latent space and jointly predicts chunks of future states to reduce sequential forward passes and reasoning latency. To provide initial supervision for latent reasoning, we construct AuralReason-685K: 685K bilingual speech utterances (about 1,000 hours) with concise CoT for emotion recognition, empathetic dialogue, and general reasoning. AURAL-RL then explores beyond these traces, rewarding concise reasoning that yields high-quality answers and adapting reasoning effort to each problem. Across two backbones, AURAL-RL achieves performance comparable to CoT-RL, with larger gains over the respective supervised checkpoints on most metrics. Analysis further shows that harder questions elicit more latent reasoning steps. On Qwen2.5-Omni, it reduces time to the first answer token by 11.8×, from 1.22 to 0.10 s, versus 0.05 s for direct answering.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Latent reasoning, Speech language model, Adaptive reasoning, Reinforcement learning, Low-latency inference

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_cZBoOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The main motivation of this paper is to use a latent reasoning chain that is composed of continuous vectors sampled from GMMs whose parameters are estimated by the LLM backbone. This enables reducing the number of time points that CoT utilizes. The experimental results show comparable performance with regular text CoT while showing advantages for latency.

Critical strengths
  • The paper proposes a feasible solution to convert the regular CoT approach to latent reasoning. It achieves this via chunking the the reasoning time steps, and learns continuous vectors instead of text decoded CoTs. This chunking results in reduction in latency which is useful for realtime conversation systems.

  • The experimental performance remains comparable with text-CoT based approaches largely.

Critical weaknesses
  • The gains over Latent-MSE for instance remains low. And we do not have a comparison in terms of latency with other latent reasoning models.
Questions for authors
  • Did you compare the latency gains over the latent reasoning baselines such as Coconut, CODI, CoLaR, LatentMSE or LatentGMM?

  • Did you compare with larger base LALMs like QWEN 3.0 Omni? Having this would be helpful to see if the idea is helpful with larger models as well.

3 comments
AuthorsOpenReview ↗

Response to Reviewer cZBo

Dear Reviewer cZBo,

Thank you for your review. All revisions are marked in blue in the updated PDF.

(1) Gains over Latent-MSE (Response to Weakness, first part)

“Weakness: The gains over Latent-MSE for instance remains low.”

Please refer to Line 382–383 on Page 8 of the revised manuscript.

AURAL outperforms Latent_MSE at both stages and, after RL, also reaches the first answer token about 3× faster (see (2) below). The accuracy gap is modest after SFT but more than doubles after RL (Table 1):

Qwen2.5-OmniMean Δ (10 metrics)Largest gains
AURAL-SFT vs. Latent_MSE+1.13GPQA +2.38, MCQ +2.26
AURAL-RL vs. Latent_MSE-RL+2.54OBQA +5.05, GPQA +4.76

SFT imitates one recorded CoT, so the benefit of modeling alternative continuations appears mainly once RL explores: Latent_MSE gains +0.95 from RL, while AURAL gains +2.36.

(2) Latency of the latent baselines (Response to Weakness, second part & Question 1)

“Weakness: And we do not have a comparison in terms of latency with other latent reasoning models.”

“Question 1: Did you compare the latency gains over the latent reasoning baselines such as Coconut, CODI, CoLaR, LatentMSE or LatentGMM?”

Please refer to Line 417–418 on Page 8 and the new Appendix D.5 (Line 1376–1398, Page 26).

As Sec. 3.1 notes, a latent trace of length L takes L sequential forward passes without chunking and ⌈L/m⌉ with it. We now measure this with the protocol of App. D:

SystemLatent statesSerial passesTTFT (s)SpeedupAvg. (10)
CoT-RL64.25 tokens64.251.2181.0×63.90
Latent_Coconut6 (fixed)60.1727.1×58.01
Latent_CODI6 (fixed)60.1737.0×58.75
Latent_CoLaR-RL13.6213.620.3064.0×62.43
Latent_MSE-RL14.1814.180.3153.9×61.09
Latent_GMM-RL13.0713.070.2984.1×63.03
AURAL-RL12.361.790.10311.8×63.63
  • The adaptive baselines reason over about as many states as AURAL-RL but need one pass per state, which makes them about 3× slower.
  • Coconut and CODI are faster only because their budget is short and fixed, and they are the least accurate latent methods.
  • AURAL-RL has both the lowest TTFT and the highest average accuracy among latent methods.

(3) Larger backbone (Response to Question 2)

“Question 2: Did you compare with larger base LALMs like QWEN 3.0 Omni? Having this would be helpful to see if the idea is helpful with larger models as well.”

Please refer to Line 389–390 on Page 8 and the new Appendix H (Line 2195–2213, Page 41).

We repeated the main comparison on Qwen3-Omni-30B-A3B-Instruct, a mixture-of-experts model with 30B total and 3B active parameters. Because of its training cost, CoT-RL and AURAL-RL are trained on a stratified 200K subset of AuralReason-683K with the same recipe.

Qwen3-Omni-30B-A3BMCQOpenQIEMOCAPMELDMMSUMMAU-ProGPQAOBQAVB-MMSUBBHIFEvalTTFT (s)
Base67.853.55268.3355.1068.0062.9767.0389.0167.6078.7079.350.074
CoT-RL73.923.72472.2860.5469.4266.7570.3391.2170.1479.2079.151.978
AURAL-RL73.813.73472.0460.6569.3866.1371.2591.8770.3078.5078.710.160

AURAL-RL improves over Base on 9 of 11 metrics, including 3.71–5.96 points on EchoMind MCQ, IEMOCAP and MELD, and is slightly below the strong base on BBH and IFEval. It stays within 0.8 points of CoT-RL on 10 of 11 metrics. As on Qwen2.5-Omni, CoT-RL leads on BBH, here by 0.70. TTFT drops from 1.978 s to 0.160 s (12.4×), so both the accuracy pattern and the speedup carry over to the larger model.

Thank you again for your review. We hope these results answer your questions, and we would be grateful if you would take them into account in your final assessment.

Thank you for your detailed rebuttal. I will keep my positive score.

Thank you so much for your kind words and positive feedback! We truly appreciate your constructive review, which helps us improve our work. Thank you!

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

AURAL replaces explicit chain-of-thought generation with continuous latent chunks in speech language models. It pools CoT token embeddings into supervised targets, predicts a Gaussian mixture over several future states, and uses semantic alignment and scheduled sampling to support decoding from sampled latents. An RL stage updates latent generation, stopping and answer generation using outcome quality and a conciseness bonus. The paper also constructs AuralReason-683K and evaluates two backbones. Results indicate broadly comparable performance to CoT-RL with substantially lower time to the first visible answer token.

Critical strengths
  1. The efficiency claim is supported by measured latency rather than forward-pass counts alone. On 256 paired examples, the default Qwen configuration reduces mean time to the first visible answer token from 1.218 s to 0.103 s, reported as 11.8x, and full text-response time from 1.510 s to 0.322 s (Appendix D, Table 9). The protocol includes model prefill, synchronizes GPU computation and reports a repeat on another node. These are model-computation timings after CPU preprocessing, rather than complete spoken-turn latency, but they substantiate a useful improvement.

  2. The implementation and ablations address several substantive challenges in latent reasoning. The paper provides exact marginal scoring for partial chunks, a causal feedback and stopping procedure, and controls for semantic alignment, scheduled sampling, mixture size and covariance rank (Appendix A and Table 2). The component interventions also test whether different mixture components produce complementary correct answers, with within-component and collapsed-mixture controls (Table 3). Evaluation on two backbones supports the method beyond a single architecture.

Critical weaknesses
  1. Small accuracy differences are interpreted without estimates of training or decoding variation. Every setting uses one training run with seed 42, while latent sampling and EOL decisions remain stochastic at evaluation (Appendix B). For example, Qwen EchoMind MCQ is 68.63% versus 68.42% for CoT-RL, while BBH is lower at 66.00% versus 68.10% (Table 1). Broadly comparable performance is plausible, but the finer claims about consistent gains and component contributions need uncertainty estimates. The second backbone is useful corroboration, but does not quantify variation within a setting.

  2. The RL analysis does not isolate the contribution of the latent-action loss. Appendix A.3 explains that the dimension-normalized multiplier is a surrogate and that clipping is inactive under the one-step update. It also notes that direct mixture-weight gradients are small after dimension normalization. Meanwhile, latent, stopping and answer losses all update the model. The final-checkpoint component interventions show useful alternatives, but do not establish how much improvement comes from optimizing the latent distribution rather than the answer and stopping objectives. This limits the explanation of why AURAL benefits more from RL.

Questions for authors
  1. Can you provide paired uncertainty estimates on the main comparison and repeat stochastic latent/EOL decoding on the same checkpoints? Any available independent training runs for the principal configuration would further clarify which small gains are robust, without requiring the entire hyperparameter grid to be retrained.

  2. What happens if the latent-action loss weight is set to zero during RL while the stopping and answer losses, prompts and schedule remain unchanged? Reporting quality, latent depth and component-diversity diagnostics for this control would clarify the contribution of direct latent-policy optimization.

2 comments
AuthorsOpenReview ↗

Response to Reviewer vkRF (Part 1/2)

Dear Reviewer vkRF,

Thank you for the careful review. All revisions are marked in blue in the updated PDF.

(1) Uncertainty estimates (Response to Weakness 1 & Question 1)

“Weakness 1: Small accuracy differences are interpreted without estimates of training or decoding variation. … the finer claims about consistent gains and component contributions need uncertainty estimates.”

“Question 1: Can you provide paired uncertainty estimates on the main comparison and repeat stochastic latent/EOL decoding on the same checkpoints? Any available independent training runs for the principal configuration would further clarify which small gains are robust …”

Please refer to Line 363–365 on Page 7 and the new Appendix G (Line 2160–2193, Page 41).

We did three things:

  • re-evaluated the reported AURAL-RL checkpoint with five decoding seeds, which resample the latent chunks and EOL decisions;
  • compared it with CoT-RL (greedy decoding) using a paired bootstrap over examples;
  • trained two additional RL runs of AURAL-RL (seeds 43 and 44) from the same SFT checkpoint.
MetricCoT-RLAURAL-RL (5 decoding seeds)Δ [95% CI]RL seed 43RL seed 44
EchoMind MCQ68.4268.66 ± 0.09+0.24 [−0.35, 0.83]68.4268.75
EchoMind OpenQ3.6583.666 ± 0.004+0.008 [−0.006, 0.022]3.6593.672
IEMOCAP71.8071.38 ± 0.31−0.42 [−2.35, 1.51]71.7270.83
MELD60.0459.63 ± 0.24−0.41 [−1.74, 0.92]59.2359.96
MMSU64.3464.72 ± 0.12+0.38 [−0.58, 1.34]64.5865.04
MMAU-Pro62.7861.41 ± 0.15−1.37 [−2.48, −0.26]61.0561.52
GPQA37.5538.76 ± 0.81+1.21 [−3.38, 5.80]38.1039.56
VB-OpenBookQA87.0386.42 ± 0.71−0.61 [−3.51, 2.29]86.3786.15
VB-MMSU63.3763.58 ± 0.19+0.21 [−1.16, 1.58]63.3463.96
VB-BBH68.1065.20 ± 0.66−2.90 [−5.03, −0.77]65.4064.80
VB-IFEval55.5254.36 ± 0.41−1.16 [−5.25, 2.93]54.2054.49
  • Decoding variation is small. The std across seeds is at most 0.15 on the three large percentage benchmarks and at most 0.81 elsewhere, and every AURAL-RL value in Table 1 lies within 0.83 points of its decoding mean.
  • RL-seed variation. With the SFT checkpoint fixed, the two additional RL runs (evaluated with the decoding seed of Table 1) reach ten-metric averages of 63.24 and 63.51, against 63.63 for the reported run. Single metrics move by up to 1.28 points, most on GPQA. All three runs improve at least 10 of 11 metrics over AURAL-SFT, and their ten-metric average gains (1.97–2.36 points) exceed CoT's 0.83-point gain.
  • The main conclusions hold. AURAL-RL shows no significant difference from CoT-RL on 9 of 11 metrics, and CoT-RL is better on MMAU-Pro and BBH. The main ablation effects in Table 2 (2.53–3.89 points on EchoMind MCQ) are far larger than the variation above. We dropped one claim: AURAL-RL's small leads on EchoMind and MMSU are within noise, so Sec. 5.2 now counts them as ties.
AuthorsOpenReview ↗

Response to Reviewer vkRF (Part 2/2)

(2) Contribution of the latent-action update (Response to Weakness 2 & Question 2)

“Weakness 2: The RL analysis does not isolate the contribution of the latent-action loss. … This limits the explanation of why AURAL benefits more from RL.”

“Question 2: What happens if the latent-action loss weight is set to zero during RL …? Reporting quality, latent depth and component-diversity diagnostics … would clarify the contribution of direct latent-policy optimization.”

Please refer to Table 2 and Line 451–458 on Page 9, and the new Appendix E.4 (Line 1563–1592, Pages 29–30).

We ran exactly this control: ω_lat = 0 during RL, starting from the same AURAL-SFT checkpoint, with the stopping and answer losses, prompts, schedule and reward unchanged. In both runs, sampled latent chunks are stored rollout actions and enter the stopping and answer losses as fixed inputs, so only ω_lat differs. With ω_lat = 0, the latent head receives no gradient and the latent distribution changes only through the backbone.

VariantMCQMMSUGPQAOBQABBHAvg. (10)Latent statesH(π)JDCR (%)
AURAL-SFT (no RL)68.2063.5433.8882.2065.0061.2716.420.710.24129.64
RL w/o latent-action update68.3964.1236.4584.8465.1062.3613.080.700.25232.53
AURAL-RL68.6364.7639.3887.2566.0063.6312.360.610.28338.88

Table 14 reports all 11 metrics and the EOL hit rate, which is 100% for both RL variants. J is the language-projection distance between mode-intervention branches, and DCR is the "two valid paths" rate of Table 3.

  • Quality. Without the update, RL still improves all 11 metrics, but the average gain over AURAL-SFT falls from 2.36 to 1.09 points, close to CoT's 0.83. The update adds the remaining 1.27 points and improves every metric, most on GPQA and OpenBookQA.
  • Latent depth. Most of the depth reduction remains without the update, since the stopping decisions still receive the conciseness bonus.
  • Component diversity. With the update, branches from different components diverge more (higher J) and more often yield two distinct correct paths (DCR +9.24 vs. +2.89 points). Mixture entropy also decreases much more.

Direct latent-policy optimization therefore accounts for most of AURAL's larger RL gain over CoT, and it makes the mixture components lead to more distinct paths.

Thank you again for the precise suggestions. We hope these additions resolve your concerns, and we would be grateful if you would take them into account in your final assessment.

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

The paper proposes, AURAL, a new mechanism for latent reasoning replacing explicit CoT for speech LLMs. Similar to CoLaR, CoT token embeddings are pooled into latent targets and a head predicts a Gaussian distribution. The AURAL's novelty is it uses mixture of Gaussian instead of CoLaR's single Gaussian. Additionally, the head predicts a Gaussian mixture (low rank + diagonal covariance) over a chunk of 10 future latent states simultaneously in a single pass. The model is trained during SFT with augmented loss including chunkwise semantic alignment (CSA) and latent scheduled sampling (LSS).

The SFT is followed by a RL stage with quality-gated conciseness reward which is comparable to CoT-RL with significant improvements with token latency.

The authors compile AuralReason speech dataset with CoTs transcribed by Gemini model.

Experiments are conducted on Qwen2.5 Omni and Kimi-Audio demonstrating significant reduction in latency metrics.

Critical strengths
  • Latent reasoning in audio LLMs that can have low latency profile is of sufficient interest in the community.
  • Well designed experiments and ablations especially latency profiling providing detailed insights on the latency aspects.
  • Elaborate analysis of Gaussian Mixture components including across and within resampling and collapsed mixture.
  • AuralReason dataset on its own is valuable with well documented process.
  • Results are encouraging: especially latency reduction. Proposed mixture prediction beats a single prediction consistently.
Critical weaknesses
  • While the Gaussian mixture output heads is novel in the context of latent reasoning, there are related ideas outside - the authors should cite them.
  • The paper incorporates latent reasoning baselines from text methods but no speech-based baselines like FLAiR, CoAT, LatentOmni (they are cited but not used as baselines)
  • Some crucial baselines missing to separate contributions of the AuralReason dataset: (1) Direct SFT, (2) Direct RL on top of Direct SFT, i.e., no CoT but finetuned on AuralReason. Accuracy gains over untuned base are substantial but cannot be fully attributed to latent reasoning.
  • Latent reasoning is not applied explicitly to speech generation, it is only applied to producing text. The latency measurements are also tied to text tokens - speech synthesis is excluded from all timings.
Questions for authors
  • Can you add tuned baselines for SFT and RL (without CoT) on AuralReason?
  • Can you compare with any speech-LM latent method? or explain why not?
  • What latency do you get including speech synthesis, end-to-end?
2 comments
AuthorsOpenReview ↗

Response to Reviewer NPuA (Part 1/2)

Dear Reviewer NPuA,

Thank you for the thoughtful review. All revisions are marked in blue in the updated PDF.

(1) Related ideas outside latent reasoning (Response to Weakness 1)

“Weakness 1: While the Gaussian mixture output heads is novel in the context of latent reasoning, there are related ideas outside - the authors should cite them.”

Please refer to Line 131–138 on Page 3 of the revised manuscript.

We added a Related Work paragraph covering three lines of work:

  • Mixture density outputs (Bishop, 1994), used for handwriting and sketches (Graves, 2013; Ha & Eck, 2018), speech synthesis (Zen & Senior, 2014; Meng et al., 2025), robot policies (Mandlekar et al., 2021) and image tokens (GIVT, Tschannen et al., 2024).
  • Low-rank plus diagonal output covariances (Monteiro et al., 2020).
  • Predicting several future steps at once: action chunking (ACT, Zhao et al., 2023), multi-token prediction (Gloeckle et al., 2024) and continuous autoregressive LMs (CALM, Shao et al., 2025).

These works model observed outputs. AURAL instead places the mixture and the chunk on unobserved reasoning states, which are fed back into the backbone and trained through the final answer.

(2) Speech latent baselines (Response to Weakness 2 & Question 2)

“Weakness 2: The paper incorporates latent reasoning baselines from text methods but no speech-based baselines like FLAiR, CoAT, LatentOmni (they are cited but not used as baselines)”

“Question 2: Can you compare with any speech-LM latent method? or explain why not?”

Please refer to Table 1 on Page 7 and Appendix B.5 (Line 1159–1180, Page 22).

We chose Coconut, CODI and CoLaR because they are modality-agnostic. We can train them with exactly our backbone, data and budget, so the comparison isolates the latent reasoning mechanism. For the speech methods:

  • FLAIR generates latent thoughts in the listening slots of full-duplex interaction, so its depth is set by how long the user speaks. Because our benchmarks provide the complete waveform before the model responds, FLAIR would need additional adaptation for our setting.
  • CoAT distills five frozen audio experts into a thinking block of fixed length and uses no reasoning traces. Building its multi-expert distillation pipeline was outside the shared training setup of Table 1.
  • LatentOmni takes its alignment targets from the backbone's own audio encoder, needs no external expert model, and fits our setting. We re-implemented it in audio-only form (LatentOmni†):
    • 8 latent states (the audio share of its 40-state budget) replace the verbalized acoustic evidence, or precede the reasoning in CoTs that have none, and the rest of the CoT stays in text;
    • its text loss and latent alignment loss are kept, with anchors pooled from the backbone's audio features as in the original;
    • the visual stream and the audio-visual synchronization terms are dropped;
    • like the original, it is trained with SFT only, on the same backbone, data and two-epoch budget.
Qwen2.5-OmniMCQIEMOCAPMELDGPQAOBQABBHAvg. (10)TTFT (s)
AURAL-SFT68.2068.8256.9733.8882.2065.0061.270.107
LatentOmni†67.8170.2758.1634.8085.0567.7062.611.231
AURAL-RL68.6371.3159.6939.3887.2566.0063.630.103

LatentOmni† keeps most of its reasoning in text, so its average accuracy stays close to CoT (62.61 vs. 63.07; all 11 metrics are in Table 1). For the same reason, it still generates that reasoning token by token and needs 1.231 s to the first answer token, versus 0.107 s for AURAL-SFT.

AuthorsOpenReview ↗

Response to Reviewer NPuA (Part 2/2)

(3) Tuned baselines without CoT (Response to Weakness 3 & Question 1)

“Weakness 3: Some crucial baselines missing to separate contributions of the AuralReason dataset: (1) Direct SFT, (2) Direct RL on top of Direct SFT, i.e., no CoT but finetuned on AuralReason. Accuracy gains over untuned base are substantial but cannot be fully attributed to latent reasoning.”

“Question 1: Can you add tuned baselines for SFT and RL (without CoT) on AuralReason?”

Please refer to Table 1 and Line 366–372 on Page 7 of the revised manuscript.

We totally agree that a tuned baseline without reasoning was missing. We now add two:

  • Direct-SFT: two epochs on the same examples with the CoT removed.
  • Direct-RL: the same RL prompts and schedule on top of Direct-SFT. With no reasoning, our reward reduces to R = (1+λ)q − β·I_end (L = 1, no EOL term).
Qwen2.5-OmniMCQOpenQIEMOCAPMELDMMSUMMAU-ProGPQAOBQAVB-MMSUBBHIFEval
Base58.093.49266.2452.8761.6055.6623.4479.7851.5366.7054.64
Direct-SFT63.413.58368.0155.4862.4657.2926.5681.1055.8664.6054.18
AURAL-SFT68.203.63768.8256.9763.5459.9133.8882.2060.3865.0053.81
Direct-RL64.273.61169.1456.8662.9858.0128.3982.8657.4265.1054.37
AURAL-RL68.633.66871.3159.6964.7661.3739.3887.2563.6366.0054.29
  • The data helps. Fine-tuning without CoT already improves Base on 9 of 11 metrics, so part of the gain over the untuned model does come from AuralReason-683K.
  • On the same data, CoT-supervised latent reasoning adds a clear further gain.
    • AURAL-SFT beats Direct-SFT on 10 of 11 metrics, by 2.38 points on average.
    • AURAL-RL beats Direct-RL on 10 of 11 metrics, by 3.69 on average, including +10.99 on GPQA and +6.21 on VoiceBench MMSU.
    • The cost is about 0.05 s in time to first token over direct answering with the untuned Base (0.050 s vs. 0.103 s, Table 9).

(4) Speech synthesis (Response to Weakness 4 & Question 3)

“Weakness 4: Latent reasoning is not applied explicitly to speech generation, it is only applied to producing text. The latency measurements are also tied to text tokens - speech synthesis is excluded from all timings.”

“Question 3: What latency do you get including speech synthesis, end-to-end?”

Please refer to the new Appendix D.6 (Line 1400–1405, Pages 26–27).

Our benchmarks score the text response, so Table 9 times the SLM up to its first answer token. Following your advice, we additionally measured time to first audio on the same 256 examples. AURAL-RL returns the first audio packet in 0.265 s, compared with 1.381 s for CoT-RL, reducing the mean delay by 1.116 s and achieving a 5.2× speedup under this setup.

AURAL shortens the reasoning before the answer and leaves speech generation unchanged. Applying latent reasoning to speech generation itself is a separate problem, which we leave to future work.

Thank you again for the constructive review. We hope these additions address your concerns, and we would be grateful if you would take them into account in your final assessment.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

Dear Reviewers and ACs,

We thank all three reviewers for their constructive and detailed reviews. We are glad that the reviewers found low-latency latent reasoning for SLMs useful (NPuA, cZBo), the efficiency claim supported by measured latency (vkRF), the ablations and GMM component analyses thorough (NPuA, vkRF), and AuralReason-683K valuable (NPuA).

We have uploaded a revised PDF with all changes in blue. The main additions are:

  1. Direct-SFT and Direct-RL (NPuA). These baselines are trained on AuralReason-683K without CoT. The data alone helps, but AURAL-RL still exceeds Direct-RL on 10 of 11 metrics, by 3.69 points on average (Table 1).
  2. A speech latent baseline (NPuA). An audio-only implementation of LatentOmni keeps most of its reasoning in text, so it stays close to CoT in accuracy but needs 1.23 s to the first answer token, versus 0.11 s for AURAL-SFT (Table 1, App. D.5).
  3. Time to first audio (NPuA). With streaming speech synthesis, AURAL-RL starts speaking after 0.27 s, versus 1.38 s for CoT-RL (App. D.6).
  4. Latency of the latent baselines (cZBo). Without chunking, the fully latent baselines need one forward pass per state and take 0.17–0.32 s to the first answer token (App. D.5).
  5. Uncertainty of the main comparison (vkRF). Five decoding seeds, paired bootstrap intervals, and two additional RL runs (App. G).
  6. RL without the latent-action update (vkRF). Without this update, AURAL's RL gain drops from 2.36 to 1.09 points, close to CoT's 0.83. The update adds 1.27 points and more distinct correct continuations (Table 2, App. E.4).
  7. A larger backbone (cZBo). On Qwen3-Omni-30B-A3B, AURAL-RL stays within 0.8 points of CoT-RL on 10 of 11 metrics and reduces TTFT by 12.4× (App. H).
  8. Related work (NPuA). Sec. 2 now discusses related ideas outside latent reasoning, such as mixture density outputs and predicting several future steps at once, and how AURAL differs from them.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.