acceptodds
Under review as a conference paper at ICLR 2027

SLC: Constructing Semantic-Aware Latent Communication for LLM-based Multi-Agent Collaboration

Abstract

Latent communication has shown promising potential for efficient inter-agent communication by exchanging continuous model-internal representations. Yet bypassing natural language alone does not suffice to enable effective multi-agent collaboration. Existing approaches either share prefill-derived representations or propagate each latent response along with accumulated context. Our analysis shows that decoupling latent information flow from the underlying coordination topology alters how intermediate reasoning is propagated and composed across agents, limiting effective collaboration along the multi-agent topology. To bridge this gap, we propose Semantic-aware Latent Communication (SLC), a latent communication protocol that supports effective collaboration along the multi-agent topology. Specifically, SLC employs a reusable semantic-aware hidden input-output alignment for latent response generation and transfers only Context-Augmented KV caches with Boundary-Mediated Latent Integration. Extensive experiments across seven benchmarks demonstrate that SLC achieves competitive accuracy with TextMAS, while enabling up to 4.1 faster inference and reducing input and output token usage by up to 82.6% and 79.6%, respectively.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multi-Agent Systems, Latent Communication, Large Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_tgZ9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies latent communication between LLM agents in a multi-agent system. It argues that current latent methods break the link between information flow and the agent topology. To address this issue, the authors propose a framework named SLC. SLC has three parts: the first is a closed-form alignment matrix fitted with top-k token probabilities, the second sends only the KV states of new latent response from each agent and one sender position, and the third inserts these blocks between boundary tokens at the receiver. The authors report accuracy close to TextMAS on seven benchmarks, and up to 4.1x faster inference with large token savings.

Critical strengths
  1. The problem framing is clear. Figure 1 and 2 are helpful to show how each paradigm changes the information flow, and why adding agents helps text communication but not existing latent methods.

  2. The proposed alignment step is simple and efficient.

  3. The efficiency study is detailed, showing results with time, memory, TFLOPs, and input/output tokens.

Critical weaknesses
  1. The results don't show any variance estimate, though agents sample at temperature = 0.7. In Table 5, the top-k sweep shows accuracy moves between 83.46 and 87.8 with no apparent trend. And the noise is larger than the average gain over TextMAS (74.46 vs. 73.6) in Table 1. Thus, the the claim that SLC "achieves the best accuracy on most tasks" is not yet supported.

  2. It is not clear how kk, mm, and ρ\rho are chosen per task. The best result of AQuA in Table 5 equals the main result in Table 1, which suggests tuning on the test set. The source of the calibration inputs is also not clear.

  3. The claim of 4.1x speedup in the abstract is about inference time only. The end-to-end (E2E) speedup is actually 2.5x (line 413). Also, could you explain the gap between E2E time and inference time? According to Table 6, the gap seems large for the proposed method (over 2,000 s), though the gap is just 28s for TextMAS.

  4. KV states computed at the sender's positions are inserted at new positions in the receiver. It is unclear how positional encoding stays consistent.

  5. All agents share one Qwen3 model. If time allows in the rebuttal, could you add a second model family, or at least a limitation section discussing this? Would SLC work across different models?

Questions for authors
  1. Could you rerun Table 1 with at least three seeds? If time is limited in the rebuttal, reporting partial benchmarks are fine. Please report mean and standard deviation, or a paired bootstrap interval against TextMAS.

  2. Could you clarify how kk, mm, ρ\rho were chosen for each task and model? Was any validation split used? Could you clarify where the N=50 calibration inputs come from? A clear description of setup with no test tuning would address my main concern.

  3. Could you explain the gap between E2E time and inference time? Does it include fitting the alignment matrix or KV handling?

  4. Could you clarify my questions regarding positional encoding and SLC across models in W4 and W5.

2 comments
AuthorsOpenReview ↗

Response to Reviewer tgZ9

We thank the reviewer for the constructive suggestions. We provide additional experiments and implementation clarifications below.

W1, W5 & Q1: Statistical Robustness and Generalization Across Model Families

We evaluated TextMAS and SLC with seeds 0, 1, and 2 on AQuA under the four-agent Chain topology, using Qwen3-8B and Llama-3.1-8B-Instruct. We report accuracy and textual output tokens per sample, with mean ± sample standard deviation:

ModelMethodSeedAcc. (%)Output Tokens / Sample
Llama-3.1-8B-InstructTextMAS060.634,394
160.244,003
256.694,656
Mean ± Std59.19 ± 2.174,351 ± 329
SLC057.871,038
159.061,103
258.271,002
Mean ± Std58.40 ± 0.611,048 ± 51
Qwen3-8BTextMAS085.832,064
185.432,094
285.042,088
Mean ± Std85.43 ± 0.402,082 ± 16
SLC083.07480
184.65480
286.61519
Mean ± Std84.78 ± 1.77493 ± 23

SLC's mean accuracy is 0.79 and 0.65 percentage points below TextMAS on Llama and Qwen3, respectively, while reducing textual output tokens by 75.9% and 76.3%. These results support a favorable accuracy-efficiency trade-off on both evaluated model families. All agents within each system share the same backbone; heterogeneous KV communication remains future work.

W2 & Q2: Hyperparameter Selection and Calibration Data

We selected kk, mm, and ρ\rho based on AQuA evaluation performance and reused the selected configurations on other benchmarks, rather than tuning each benchmark separately. We considered m∈{40,50,60}m\in\{40,50,60\} latent steps and k∈{16,32,64}k\in\{16,32,64\} token alternatives, with ρ=10−4\rho=10^{-4} for regularization. The latent-step setup follows LatentMAS, while retaining multiple token alternatives is motivated by CIPHER's distribution-based semantic representations. Parameter settings and sensitivity analyses appear in Sections 5.1 and 5.4 and Tables 4–5. We fixed N=50N=50 as a practical calibration budget; each TextMAS trajectory supplies multiple response positions for fitting agent-specific matrices.

Calibration inputs were selected from benchmark evaluation questions without using reference answers or correctness feedback to select samples or fit matrices. For AQuA, these were the first 50 inputs in dataset order, also included in the full 254-example evaluation. We report Pass@1 (%) separately on calibration and remaining examples for two code benchmarks (Qwen3-8B, four-agent Chain):

DatasetMethodCalibration (50)RemainingFull
MBPP+TextMAS74.0066.4667.46
SLC70.0068.9069.05
HumanEval+TextMAS86.0078.0780.49
SLC86.0074.5678.05

W3 & Q3: Online Inference and End-to-End Time

The 4.1× speedup refers to online inference; Section 5.2 (Lines 411–413) also reports 2.5× end-to-end (E2E) acceleration. The reported E2E time includes model loading, calibration-response generation, hidden-state extraction, and alignment fitting. Online inference excludes these setup costs but includes KV handling. More intermediate agents require more alignment matrices, accounting for the growing setup overhead. The matrices remain fixed during inference (Appendix C.2, Lines 874–879). With four agents, SLC's reported E2E time is higher than KVComm's and LatentMAS's, while its online inference time is lower than both. We will clarify the timing breakdown.

W4: Position IDs and RoPE During KV Transfer

We do not re-rotate sender keys. Each transferred KV block retains its RoPE encoding and bounds [s,e)[s,e), where ss is the minimum position ID and ee is one past the maximum. A receiver prefix of LL valid tokens, including the opening boundary, receives IDs s−L,…,s−1s-L,\ldots,s-1, without clamping negative values. Sender KV is inserted unchanged; the closing boundary and suffix begin at ee. With multiple senders, KV blocks and validity masks are concatenated in workflow order, using s=min⁡jsjs=\min_j s_j and e=max⁡jeje=\max_j e_j for receiver positioning. Overlapping or noncontiguous sender position ranges are preserved. In our implementation, causal visibility follows cache sequence indices and validity masks, separately from numerical RoPE IDs. We will clarify this procedure in the manuscript.

Thank you for your response. However, my main concerns are still not resolved. The accuracy claims are not yet supported by variance estimates across the main results, the hyperparameters appear to chosen using a validation set that overlaps the test set, the 4.1× speedup applies only to online inference and the extra setup time is not explained, and the positional encoding procedure is described but not verified. Given the remaining concerns, I'll keep my original score.

Reviewer_YJ2mOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the limitations of existing latent space communication methods in LLM-based MAS—such as insufficient reasoning information transfer, redundant historical context, and high communication overhead—by proposing a Semantic-aware Latent Communication framework (SLC). The method features two core designs: (1) Semantic-aware Latent Response Generation, which constructs a semantic-aware input-output alignment matrix using Top-k token probabilities to enable efficient autoregressive generation in the latent space; and (2) Latent Context Composition, which transmits the KV Cache of newly generated responses and integrates sender-side context with boundary markers to facilitate information passing that aligns with the multi-agent topology, thereby avoiding redundant accumulation of historical data. The authors evaluated SLC on seven benchmark datasets covering mathematical reasoning, scientific reasoning, and code generation, using Qwen3-8B and Qwen3-14B. Experimental results show that SLC achieves up to 4.1× online inference acceleration while maintaining performance comparable to text-based communication methods, and significantly reduces token usage and memory overhead.

Critical strengths
  1. The method design is simple and has good deployment potential. SLC does not introduce complex additional neural networks. Instead, it achieves self-regression generation in the latent space through a linear mapping matrix obtained through a one-time calibration. At the same time, the method utilizes the existing KV Cache mechanism of Transformer to achieve information transmission, without having to re-perform the complete text encoding process for each message. This design has certain engineering practicality and avoids additional costs.
  2. The experiments cover multiple tasks and the efficiency gains are quite obvious. The paper uses seven Benchmarks, two different-sized LLMs, and compares with various representative communication strategies, yielding impressive experimental results.
Critical weaknesses
  1. Due to the lack of validation for heterogeneous Agents and different model architectures, the authors only used the Qwen3-8B and Qwen3-14B models, and each experiment's Agent was based on the same Backbone. This setting ensures that the KV Cache of each Agent has a compatible representation space. However, in actual multi-agent systems, different Agents may use different models, with different Hidden Dimensions, Attention Head configurations, and positional encoding methods. In these cases, directly passing the KV Cache may not work properly and cannot guarantee semantic consistency.
  2. The existing experiments are not sufficient to prove that the performance improvement truly comes from more efficient multi-agent collaboration. For example, in Figure 5(a), SLC does not demonstrate a sustained performance improvement brought by the addition of Agents. Therefore, the existing experiments more fully prove efficient communication rather than a more effective collaboration mechanism. It is recommended that the authors design experiments to analyze the actual gains of each Agent.
  3. In line 80, how to reflect "collaboration gains" from Figure 2? If it is accuracy rate, the (a) sub-figure should show different numbers of Agents and compare the accuracy rates of the methods.
  4. Minor issues: The letter "K" in the KV cache overlaps with the "K" in formulas (3) and (4). It is recommended to change a variable name and the explanation of W_s in formula 3 is not clear enough.
Questions for authors

The theoretical analysis of the paper relies on the Shared Semantic Coefficients assumption, which states that the actual hidden state can be represented as (h=\beta\sigma W_{out}). However, the token probabilities can be calculated from the hidden state. But merely relying on the token probabilities does not necessarily allow for an accurate restoration of the original hidden state. This is because the hidden state may contain some information that is not fully expressed by the top-k token probabilities? Can the author explain the rationale for this hypothesis in the context of practical LLMs?

1 comment
AuthorsOpenReview ↗

Response to Reviewer YJ2m

We thank the reviewer for the careful assessment and constructive suggestions. We clarify SLC's scope, experimental design, and theoretical assumptions below.

W1: Heterogeneous Model Communication

SLC currently assumes compatible KV representations, with all agents within each multi-agent system sharing the same backbone. Our work focuses on generating, transmitting, and integrating latent responses along the prescribed collaboration topology, rather than aligning representations across heterogeneous models. Cross-model communication requires additional alignment of representation spaces, attention configurations, and positional encodings, as explored by Cache-to-Cache (C2C; Fu et al., 2026) through learned projection and fusion mechanisms.

To evaluate applicability beyond Qwen3, we additionally ran three seeds on AQuA with Llama-3.1-8B-Instruct under the four-agent Chain topology. SLC achieves 58.40 ± 0.61% accuracy versus 59.19 ± 2.17% for TextMAS, with 75.9% fewer textual output tokens. Each system still uses a shared backbone. We appreciate the suggestion and will consider heterogeneous multi-agent communication in future work.

W2 & W3: Collaboration Effectiveness and Agent-Scaling Experiments

We agree that our original wording of "collaboration gains" could suggest that adding agents should consistently improve accuracy, which was not our intended claim. As illustrated in Figure 1, latent context sharing induces star-like information flow without newly generated intermediate responses, while latent response accumulation creates complete-graph-like dependencies by repeatedly forwarding accumulated history, even under a Chain topology. Varying agent count within each topology allows us to examine how the effects of these communication mechanisms evolve with collaboration scale. Alongside comparisons across topologies, these experiments assess task performance and communication overhead, rather than a monotonic accuracy-scaling law.

Figure 2(a) summarizes each method's accuracy across the evaluated agent counts, rather than the incremental contribution of each agent. Quantifying individual contributions would require targeted agent-level ablations. We have revised the Introduction (Lines 77–80) to clarify this motivation and distinguish communication efficiency from task effectiveness.

W4: Notation and the Alignment Matrix

We have replaced KK, denoting the total number of collected calibration hidden states, with QQ in the corresponding formulations to avoid confusion with keys in KV caches. Section 4.1 now explicitly defines WsW_s as the mapping from final-layer hidden states to the Transformer input embedding space. Both revisions are included in the updated manuscript.

Q1: Rationale for the Shared Semantic Coefficients Assumption

We agree that top-k token probabilities may not fully capture practical LLM hidden states. The shared semantic-coefficient assumption is an idealization for deriving the alignment objective, not a claim that hidden states can be exactly reconstructed from top-k probabilities. In Appendix B.1–B.2, shared coefficients pair corresponding input and output representations, enabling Proposition B.1 to upper-bound their squared Wasserstein distance. This motivates an alignment objective that accounts for the semantic coefficient distribution.

In practice, Appendix B.3 constructs empirical coefficients from top-k token probabilities to represent context-dependent semantic alternatives during calibration. During inference, SLC directly maps the complete hidden state through Ws⋆W_s^\star; it does not reconstruct hidden states from token probabilities. The ablation in Table 2 further supports the empirical value of semantic-aware alignment: AQuA accuracy on Qwen3-14B increases from 82.28% without this component to 88.58% with it.

Reviewer_uAxMOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

SLC is a latent communication protocol for multi-agent systems built on a single shared model. Each agent maps output hidden states to input embeddings with a weighted ridge regression, solved in closed form on TextMAS trajectories. It then passes on only the KV of its last prompt position plus m latent steps, and the receiver splices these into its own cache between boundary tokens. On Qwen3-8B/14B and seven benchmarks, SLC roughly matches TextMAS accuracy with far fewer tokens and reports up to 4.1× faster inference.

Critical strengths

Forwarding only the incremental KV, instead of the accumulated cache, is a simple and sensible fix to LatentMAS-style forwarding. In Table 6, memory overhead stays around 0.6 GiB from 2 to 4 agents, and accuracy holds up while LatentMAS drops from 80.31 to 55.12.

Critical weaknesses
  1. The main claim is that SLC turns more agents into collaboration gains, and the paper's own data does not support it. In Table 6, SLC goes 87.40 → 86.22 → 86.22 from 2 to 4 agents. KVComm, which the paper says adds no new reasoning, beats TextMAS at both 2 and 4 agents. Without a single-agent baseline, all the results show is that SLC degrades less than LatentMAS.

  2. The 4.1× speedup is measured on inference time. End-to-end time in the same row gives 2.5× (13371.80 vs 5341.36 s). There is an unexplained gap of 659–2053 s between SLC's end-to-end and inference time, and it grows with the number of agents; for every other method the gap is about 30–50 s. On end-to-end time, SLC with 4 agents is slower than KVComm (4216 s) and LatentMAS (4480 s). The offline calibration cost is also left out.

Some baseline numbers also look off. LatentMAS on 14B averages 41.00, below its 8B result. KVComm on 8B ARC-C scores 62.23 while producing only 11 output tokens.

Questions for authors
  1. What accounts for SLC's end-to-end vs inference time gap in Table 6?
  2. How does a single-agent baseline with a matched budget do on AQuA?
  3. Were k, m, ρ and the 50 calibration inputs selected on or drawn from the test sets?
  4. When sender KV is inserted, how are position IDs handled? Are the keys re-rotated?
  5. Can the LatentMAS 14B numbers be reproduced with the official code and configs?
2 comments
AuthorsOpenReview ↗

Response to Reviewer uAxM

We thank the reviewer for the constructive feedback.

W1 & Q2: Main Claim and Single-Agent Baseline

Our main claim concerns topology-aligned latent communication, not monotonically increasing accuracy with more agents. As stated in the Abstract (Lines 17–19) and Introduction (Lines 40–83), we study how latent information flow can diverge from the prescribed topology and alter the propagation of intermediate reasoning. Figure 1 shows that context sharing and response accumulation can induce star-like and complete-graph-like flows even under a Chain topology. We vary agent count within each topology to examine how these effects evolve with collaboration scale. SLC transmits new latent responses along prescribed edges without forwarding inherited caches. KVComm can still help the final Solver through prefill-derived context, even without newly generated intermediate responses.

We compare a single Qwen3-8B agent on AQuA with four-agent Chain systems. Total output counts textual tokens and latent steps per sample:

MethodAccuracy (%) ↑Latent stepsText tokensTotal output ↓
Single Agent83.860587587
TextMAS86.2201,7121,712
SLC87.80120529649

SLC improves accuracy over the single agent by 3.94 percentage points, using 649 total output positions versus 587 for the single agent and 1,712 for TextMAS (62.1% fewer than TextMAS). Table 6 also reports SLC accuracy above the single-agent result with 2–4 agents under CompleteGraph (87.40%, 86.22%, 86.22%).

W2 & Q1: Online and End-to-End Time

The 4.1× speedup refers to online inference; Section 5.2 (Lines 411–413) also reports 2.5× end-to-end acceleration. The 659–2053 s gap primarily reflects model initialization and one-time semantic alignment. The reported E2E time includes model loading, calibration-response generation, hidden-state extraction, and alignment fitting; online inference excludes these setup costs but includes KV handling. More intermediate agents require more alignment matrices (Appendix C.2, Lines 874–879). With four agents, SLC has higher reported E2E time than KVComm and LatentMAS, but lower online inference time than both.

Q3: Hyperparameters and Calibration Data

We selected kk, mm, and ρ\rho based on AQuA evaluation performance and reused the selected configurations on other benchmarks, rather than tuning each benchmark separately. Settings and sensitivity analyses appear in Sections 5.1 and 5.4 and Appendix E. AQuA calibration used the first 50 evaluation inputs in dataset order, also included in the full 254-example evaluation. Their TextMAS trajectories were used to fit intermediate-agent alignment matrices without accessing reference answers. The final Solver was not calibrated, and backbone weights remained unchanged (Appendix C.2, Lines 874–879).

Q4: Position IDs and RoPE

We do not re-rotate sender keys. Each KV block retains its RoPE encoding and bounds [s,e)[s,e), with ss the minimum position ID and ee one past the maximum. A receiver prefix of LL valid tokens, including the opening boundary, receives IDs s−L,…,s−1s-L,\ldots,s-1, without clamping negative values. Incoming KV is inserted unchanged; the closing boundary and suffix begin at ee. With multiple senders, KV blocks and validity masks are concatenated in workflow order, using s=min⁡jsjs=\min_j s_j and e=max⁡jeje=\max_j e_j for receiver positioning. Overlapping or noncontiguous sender position ranges are preserved. In our implementation, causal visibility follows cache sequence indices and validity masks, separately from numerical RoPE IDs.

Q5: LatentMAS Implementation and Reproducibility

Our reported LatentMAS results come from a paper-based reimplementation, rather than direct execution of the official repository. Our implementation of working-memory transfer (Section 3.2) computes the receiver's local-prompt KV separately before prepending predecessor KV for latent generation. The official code instead passes predecessor KV through past_key_values during receiver prompt prefill (methods/latent_mas.py, models.py). This changes whether receiver prompt representations attend to inherited context. The reported Qwen3-14B results were obtained with our reimplementation under the common evaluation protocol.

Baseline Verification: KVComm

Inspection of the original logs revealed failed KVComm runs. We reran the affected experiments and updated Table 1. Qwen3-8B ARC-C changed from 62.23% accuracy / 11 output positions to 94.03% / 327. Since KVComm transfers prefill-derived representations without generating intermediate responses, its generated output primarily comes from the final Solver.

Thanks for the response. The timing and RoPE details help, and I’ll bump my score to 2. I’m still not convinced on two points, though: the hyperparameters and calibration samples come from the test set, and the gain over a single agent is within the seed noise.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Overall Response: Clarifying the Key Contributions of SLC and Addressing Reviewers' Concerns

We sincerely thank all reviewers for their constructive feedback and for recognizing SLC's clear problem formulation, methodological simplicity, deployment potential, and efficiency improvements. We summarize the motivation, methodological contributions, and experimental design below.

1. Rethinking latent communication from the perspective of collaboration topology.

A key insight of our work is that the information flow induced by latent communication may differ from the intended multi-agent collaboration topology. The topology specifies how agents build upon their predecessors' reasoning, but the communication mechanism determines what information is propagated and how it is incorporated downstream. In the Chain example in Figure 1, latent context sharing transfers prefill-derived representations without newly generated intermediate responses, inducing star-like information flow. Latent response accumulation instead forwards generated responses together with accumulated history, inducing complete-graph-like information flow. These observations motivate us to study how latent communication can support the propagation and composition of intermediate reasoning along the prescribed topology, beyond efficient representation exchange alone.

2. A unified protocol for topology-aligned latent collaboration.

SLC integrates semantic-aware latent response generation with topology-aligned latent context composition. Semantic-aware input-output distribution alignment incorporates top-k token distributions into a regularized objective with a closed-form solution, yielding a reusable hidden-state mapping for latent response generation. Context-augmented KV transfer communicates only each agent's newly generated latent response and a single sender-context position, rather than recursively forwarding inherited caches. Boundary-mediated latent integration enables downstream agents to condition subsequent generation on both upstream responses and local instructions without re-encoding the received messages. Together, these mechanisms connect response generation, transmission, and integration along the intended collaboration workflow while avoiding repeated text encoding and recursive history accumulation.

3. Structural advantages of topology-aligned communication.

Our communication complexity analysis characterizes the distinction between SLC and cumulative forwarding. Over one execution round on the considered directed acyclic graphs, with fixed-size local response blocks, SLC transmits one locally generated message per edge, making communication volume proportional to the number of edges. Cumulative forwarding without deduplication or truncation repeatedly propagates inherited information, with transmission volume determined by the number of nonempty directed paths. Under these assumptions, SLC reduces communication complexity from quadratic to linear for chains and from exponential to quadratic for complete forward DAGs. This explains the structural benefit of avoiding recursive history replication, beyond replacing textual messages with KV caches.

4. Motivation for agent-scaling experiments and empirical findings.

Motivated by the information-flow differences in Figure 1, our agent-scaling experiments examine how communication paradigms affect task performance and communication efficiency as the collaboration scale varies. Alongside comparisons across topologies, varying agent count within each topology reveals how these differences evolve as more intermediate agents participate. Across seven benchmarks and two Qwen3 model scales, SLC achieves competitive accuracy with TextMAS while substantially reducing online inference and communication overhead. On AQuA with four Qwen3-8B agents under CompleteGraph, it achieves 4.1× online inference acceleration and 2.5× end-to-end acceleration over TextMAS. In the agent-scaling experiments, SLC maintains relatively stable accuracy compared with the evaluated LatentMAS baseline. Additional experiments and technical clarifications concerning statistical reliability, calibration protocols, timing, and implementation are provided in our reviewer-specific responses.

Overall, SLC contributes a topology-aligned perspective on latent communication and a unified protocol for progressively generating, transmitting, and integrating latent responses across collaborating agents. This perspective connects the design of latent communication mechanisms to the collaboration structures they are intended to support.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.