- Rating
- 1(−1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper introduces ReMux, which superposes multiple queries into a shared representation, performs collaborative latent reasoning, and decodes separate answers using the original queries and a shared KV cache. LoRA and stepwise trajectory supervision train this behavior. Experiments across eight benchmarks, multiple model families, and two agent topologies report approximately preserved accuracy with substantially reduced token counts and runtime.
Critical strengths
--- A relevant efficiency problem. Sharing computation across independent queries is practically meaningful, and extending multiplexing to iterative latent collaboration is a potentially useful contribution, especially for multi LLM system.
--- Broad empirical evidence. Multiple model families, eight benchmarks, two agent topologies, and component ablations provide useful evidence that the proposed system can retain task performance.
Critical weaknesses
--- The incremental contribution and decoupling mechanism are insufficiently established. Random binding and superposition closely follow DataMUX/MUX-PLMs, while latent recurrence and KV communication build on LatentMAS. RevMUX is also directly relevant, and the additional contribution of CSO needs clearer isolation. More seriously, Theorem A.3’s proof assumes that separation survives learned value projections, which a rank-one (W_V) can invalidate. Its guarantee additionally concerns a projected output, although the implemented decoder contains no such projector. The claimed equivalence to downstream decoding is not established.
--- The efficiency gains cannot yet be assessed under consistent accounting and controls. For ARC, Eq. 53 and the stated settings in Appendix D.5 imply at most 587 reasoning steps, but Table 1 reports 3,251 for LatentMAS? Different counting conventions or budgets may explain this, but are not explained. ReMux also receives additional training without clearly matched baseline adaptation. Eq. 54 measures batch time rather than per-request latency.
Questions for authors
-- What specifically advances beyond RevMUX-style multiplexing and LatentMAS? Can the contribution of CSO be isolated while keeping the alignment module and training budget fixed?
-- Can Theorem A.3 be corrected to address rank-collapsing value projections? Which operation in the actual decoder realizes the query-specific projector (P_j)?
-- Can the authors align the token formula, generation budgets, and reported counts, and provide comparisons with training matched, ordinarily batched baselines under identical hardware, decoding budgets, and timing conventions?
-- Can the authors provide code, exact checkpoint IDs and reasoning modes, and details of gradient propagation through the recurrent latent steps and shared KV caches?
2 comments
4. Align the reported counts, and generation budgets, and training matched, and batched baselines.
We address this issue in the following:
- We align the generation budgets report in Appendix D.5. One major difference between 3251 and 587 is that the latent reasoning step is set to 800 for the first (m-1) agents. As shown in Appendix F.4 of the revised manuscript, performance improves with more latent reasoning steps after fine-tuning. For LatentMAS and REMUX, we set the latent reasoning step to 800 for ARC-E, 1000 for ARC-C and GSM8K, 2000 for MedQA, 1400 for MBPP+ and HumanEval+, 7000 for GPQ and AIME25.
- For fair comparison, all baselines are fine-tuned with LoRA under identical configurations on the same training dataset as REMUX, with supervision from the cross-entropy (CE) loss.
- All baselines are evaluated in batch mode, with a batch size equal to n, where n is the number of queries superimposed by REMUX with the same backbone model. Thus, both the baselines and REMUX process n original queries per forward pass. REMUX superimposes these queries into a single query, resulting in an input batch size of 1.
- We have corrected the reported Time in Appendix D.4, which computes the amortized per-query processing time. We confirm that we have aligned the generation budgets, reported counts, training-aligned comparisons, and batch sizes. These details are explicitly provided in Appendices D.4, D.5, and E of the revised manuscript.
We confirm that we have aligned the generation budgets, reported counts, training-aligned comparisons, and batch sizes. These details are explicitly provided in Appendices D.4, D.5, and E of the revised manuscript.
5. Code & checkpoint & details of gradient propagation?
We are currently preparing the code and model checkpoints for public release and expect to make them available soon. In the meantime, we have provided additional implementation details in the revised manuscript to facilitate reproducibility.
We sincerely appreciate the reviewer’s insightful comments, which will help us further improve the quality and clarity of our manuscript. Please find our detailed responses below.
1. Advances beyond RevMUX-style multiplexing and LatentMAS?
RevMUX-style multiplexing primarily focuses on accelerating inference for single-step discriminative tasks using relatively small models, such as Transformers (DataMUX[1], RevMUX[3]), BERT (DataMUX[1] MUX-PLMs[2]), and LLaMA adapted for discriminative tasks (RevMUX[3]). In contrast, REMUX targets autoregressive generative tasks in multi-agent systems (MASs) composed of LLM-based agents, aiming to reduce token consumption (a major bottleneck in MASs) and accelerate inference without modifying the backbone LLM architecture. Directly applying existing multiplexing techniques from prior work to this setting leads to unacceptably poor reasoning results and fails to deliver the token savings and inference acceleration achieved by REMUX. As demonstrated in Table2.pdf of the link, we directly adapt the multiplexing techniques from prior work to Qwen3-4B, Qwen3-8B, and Qwen3-14B, and report the average reasoning accuracy (ACC), token savings (TOKEN), and inference time to provide a quantitative comparison.
REMUX introduces key innovations from both theoretical and methodological perspectives. First, theoretical advances: Theorems A.1 and A.2 establish the feasibility of random binding-based superposition in LLMs for autoregressive generative tasks, despite their substantially larger parameter scales and more complex nonlinear architectures. Theorem A.3 further provides a theoretical foundation for conditional demultiplexing under superposition in autoregressive generation. Second, methodological advances: Guided by these theoretical insights, we integrate LoRA fine-tuning and continuous supervised alignment to enable effective superposition reasoning. We further introduce a dedicated decoupling agent that recovers query-specific responses by leveraging both the current superimposed representation and historical superimposed trajectories preserved in the KV cache.
Notably, while LatentMAS establishes a one-to-one latent reasoning paradigm for MASs, REMUX extends this paradigm to many-to-many superposition reasoning, enabling substantial reductions in token consumption and inference latency.
Overall, REMUX extends multiplexing from single-step discriminative inference to autoregressive LLM-based multi-agent reasoning, achieving both inference acceleration and token savings, while advancing LatentMAS from one-to-one to many-to-many latent reasoning. We have incorporated these clarifications into Section 6 (Discussion) of the revised manuscript.
2. Can CSO be isolated while keeping the alignment and training budget fixed?
We conduct controlled ablations fixing the alignment module and training budget to isolate CSO’s contribution on Qwen3-4B, Qwen3-8B, Qwen3-14B under the sequential MAS setting. The results reported in the Figure1.pdf of the link. Our ablation studies reveal that incorporating the CSO loss brings the maximum performance uplifts of 3.7%, 2.7% and 1.9% absolute points for Qwen3-4B, Qwen3-8B and Qwen3-14B backbones. We have added the ablation study of CSO to Tables 2, 3, and 4 in the revised manuscript.
3. Can Theorem A.3 be corrected to address rank-1 projections. Which operation realizes the projector (P_j).
Since rank-one value-projection matrices mix reasoning trajectories from multiple queries, potentially introducing substantial decoupling errors, we have revised Theorem A.3 by removing the projection P_j and instead developing a conditional readout mechanism: when the decoupling error remains within the decoding margin, REMUX can recover outputs identical to those produced by native reasoning. Importantly, Theorem A.3 indicates that our LoRA fine-tuning mitigates the decoupling error. Further experimental results presented in Tables 2, 3, and 4 of the manuscript demonstrate the effectiveness of fine-tuning in mitigating decoupling errors. The updated Theorem A.3 is included in the revised manuscript.
[1] Murahari et al. DataMUX: Data Multiplexing for Neural Networks. NeurIPS 2022.
[2] Murahari et al. MUX-PLMs: Data Multiplexing for High-throughput Language Models. EMNLP 2023.
[3] Xu et al. RevMUX: Data Multiplexing with Reversible Adapters for Efficient LLM Batch Inference. EMNLP 2024.