acceptodds
Withdrawn from ICLR 2027

Many Minds, One Pass: Multiplexing Collaborative Reasoning via Superposition

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multi-agent systems (MASs) enable emergent system-level intelligence, driving a paradigm shift in problem solving toward collaborative reasoning. However, standard MAS reasoning procedures process each query independently, incurring substantial redundant computation and token expenditure as query volume grows and creating a bottleneck to large-scale deployment. To address this challenge, we propose REMUX, a novel token-efficient collaborative reasoning framework that uses reasoning superposition to process multiple queries concurrently in a single pass through a MAS. Given a batch of queries, REMUX encodes their reasoning signals in approximately mutually orthogonal subspaces and superposes them in a shared latent space. By preserving near-orthogonality throughout reasoning, REMUX enables a demultiplexing agent to de-multiplex the signals with minimal cross-query interference and recover the corresponding answers. Extensive experiments demonstrate that, by multiplexing up to 20 queries concurrently, REMUX reduces token consumption by a factor of – and achieves a – speedup over state-of-the-art baselines while maintaining comparable task accuracy. The code for REMUX will be released publicly.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multi‑Agent Systems, Latent Collaborative Reasoning, Representation Superposition, Data Multiplexing, Token‑Efficient Inference

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_SLStOpenReview ↗
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

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.

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.

Reviewer_brD3OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes REMUX to share intermediate computation across several queries in a multi-agent reasoning system. Query embeddings are transformed by fixed random vectors and superposed into one sequence. Intermediate agents reason in this shared latent space and pass their KV caches. A learned alignment matrix and supervision from individual-query trajectories reduce representation drift. The final agent receives each original query and the shared cache to generate separate answers. The paper reports similar accuracy to LatentMAS across eight benchmarks and several model families, with large reductions in average reasoning steps and amortized inference time.

Critical strengths
  1. Sharing intermediate latent computation across queries is an interesting contribution for reducing the cost of multi-agent reasoning. Sections 4.1 to 4.4 combine multiplexing, trajectory alignment, and cache communication into a trainable system. This goes beyond simply shortening messages between agents.
  2. The empirical coverage is substantial. Table 1 and Appendix F.7 cover eight benchmarks, several model families, and sequential and hierarchical systems. Table 2 and Appendix F.2 test the random transformation, cache communication, alignment objective, and fine-tuning. The query-count study also acknowledges the trade-off between efficiency and interference.
Critical weaknesses
  1. I do not see how the theoretical recovery guarantee matches the implemented readout. Section 4.3 constructs the decoding query from the original input, while Equation 31 in Appendix A.2 constructs it from the randomly transformed input. I understand Theorem A.3 to bound recovery after applying a query-specific value-space projector in Equations 32 and 38. I find the next step unproven, because the paper treats downstream computation as equivalent to this projector without establishing that equivalence. Attention over a summed value cache still contains contributions from other queries before this projection. I regard this as a substantive gap in the theoretical justification, although it does not by itself invalidate the empirical method. I would like the authors to explain the actual readout and either justify this missing step or limit the theorem to the projected setting it proves.
  2. I find the reported speedup difficult to attribute from the stated comparison protocol. Appendix D.4 divides batch runtime by the number of superposed queries, so it measures amortized processing time. Appendix F.7 lists a superposition count of one for baselines, but leaves ordinary query batching unspecified. Appendix E describes fine-tuning REMUX, while the adaptation status of LatentMAS and the single-agent baseline is not explicit. I appreciate that the existing ablations show training matters, but they do not resolve this comparison. I see these as fixable reporting and comparison concerns. I would like the authors to state baseline batch sizes and training status and describe the speedup claim at the level supported by the existing runs.
Questions for authors
  1. I would like to know whether the actual decoder uses the original query or the same randomly transformed query as Equation 31. I would also like the authors to explain which operation implements the projector in Theorem A.3, or which argument makes it unnecessary for the actual decoder.
  2. I would like the authors to specify the ordinary query batch sizes for LatentMAS and Single-Agent in the reported timing runs and identify which baselines received fine-tuning on the same training data. I would also like them to clarify the existing configurations and distinguish amortized throughput from individual-query latency.
1 comment

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. Decoder uses original query or same randomly transformed query as Eq.31.

We explain this issue along the following two lines.

  • The first m-1 agents employ randomly transformed query to reason in superposition. This is because random transformation approximately maps each query into an orthogonal region and mitigates interference during superposition reasoning. We establish the corresponding theoretical guarantees in Theorems A.1 and A.2.
  • The final agent leverages the original query to decouple the shared trajectory and generate a textual answer per query. The reason is that the original query attends to its matching reasoning trajectory in the attention layer. Theorem A.3 is provided to validate this separation procedure.

2. Which operation implements the projector in Theorem A.3.

As stated at the end of Theorem A.3, the projection P was originally introduced as a theoretical assumption to characterize the property that query-specific reasoning trajectories occupy distinct regions. Importantly, no explicit projection module is required or implemented in practice.

To address the reviewer’s concern, we have removed the projection P and instead developed a conditional readout mechanism to theoretically characterize the recoverability of query-specific outputs. Specifically, when the decoupling error remains within the decoding margin, REMUX can recover outputs identical to those produced by native reasoning. Crucially, our fine-tuning procedure is employed to suppress the decoupling error, thereby facilitating this recovery. 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 provided in the newly uploaded manuscript.


3. Batch sizes and fine-tuning for baseslines. Distinguish amortized throughput from individual-query latency.

We supplement the configuration details as follows.

  • Batching configurations. All baselines in the time comparison are also evaluated in batch mode, with a batch size equal to n, where n is the number of queries superimposed by REMUX. Thus, both the baselines and REMUX process n original queries per run. REMUX superimposes these queries into a single query, resulting in an input batch size of 1.
  • The differentiation between amortized throughput and per-query latency. As detailed above regarding our batching configurations, the reported Time in Appendix D.4 computes the amortized per-query processing time.
  • Baseline adaptation. 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.

These points are now explicitly noted in Appendices D.4, D.5, and E of the revised manuscript.

Reviewer_ZPd1OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The authors argue that standard MAS reasoning processes each query separately, which wastes a lot of tokens. To address this, the authors propose ReMux, which superposes a batch of queries into one latent input with random near-orthogonal masks. The first m−1 agents reason over this shared state once, and the final agent separates the answer for each query from the shared KV cache. In detail, the backbone is fine-tuned with LoRA, plus an alignment loss (CSO) to reduce drift in the superposed states. On eight benchmarks with several backbones, ReMux keeps accuracy comparable to LatentMAS while reporting 10 to 16× fewer tokens and 2 to 10× speedup.

Critical strengths
  1. Processing multiple queries in one pass through a MAS is a novel direction, bringing data multiplexing to generative LLM agents, and it is evaluated on many backbones, including an MoE model.

  2. The results are promising, as accuracy stays close to LatentMAS even on code, GPQA and AIME25, which are not in the training data.

Critical weaknesses
  1. The reported token savings do not match the paper’s own analysis. By Eq. 43, the saving over one-to-one reasoning is at most m (about 3.3× for m=4 and n=15), but Table 1 reports 10 to 24×. Since the final agent still decodes each query separately, the gain likely comes from shorter final outputs after fine-tuning, not from superposition.

  2. The time comparison is not fair. Baselines run with n=1, while ReMux time is the batch time divided by n, so the speedup mostly reflects batching, which the baselines could also use.

Questions for authors

Can the authors report the results of other baseline method with the same LoRA fine-tuning, as this can isolate the advantage of the proposed method clearer?

1 comment

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. Can the authors report the results of other baseline method with the same LoRA fine-tuning, as this can isolate the advantage of the proposed method clearer?

Eq. (43) assumes that all agents generate outputs of equal length, denoted by c. In contrast, Table 1 reports end-to-end token consumption, where neither the intermediate agents' output lengths nor the final response length are constrained to c. Therefore, Eq. (43) is not intended to provide an upper bound on end-to-end token consumption. Rather, its purpose is to establish a unified analytical framework for comparing the token consumption of REMUX and the baselines under consistent assumptions. We have clarified this distinction in the revised manuscript.

To address the reviewer’s concern, we report the final-response generation lengths after fine-tuning for TextMAS, LatentMAS, and REMUX on Qwen3-4B, Qwen3-8B, and Qwen3-14B backbones under the sequential MAS setting. The results are provided in Table1.pdf of the link. Our experimental results show that the token savings from fine-tuning constitute merely 0.05% to 2.40% of LatentMAS’s total token budget. We have added these results and discussions in Appendix F.3 of the revised manuscript.


2. Batch size for baselines and REMUX.

We clarify that all baselines in the time comparison are also 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 run. REMUX superimposes these queries into a single query, resulting in an input batch size of 1. We have added the batch size details in Appendix D.5 of the revised manuscript.

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

Related papers

Discussion (0)

Sign in to comment.