- 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
-
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.
-
The proposed alignment step is simple and efficient.
-
The efficiency study is detailed, showing results with time, memory, TFLOPs, and input/output tokens.
Critical weaknesses
-
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.
-
It is not clear how , , and 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.
-
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.
-
KV states computed at the sender's positions are inserted at new positions in the receiver. It is unclear how positional encoding stays consistent.
-
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
-
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.
-
Could you clarify how , , 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.
-
Could you explain the gap between E2E time and inference time? Does it include fitting the alignment matrix or KV handling?
-
Could you clarify my questions regarding positional encoding and SLC across models in W4 and W5.
2 comments
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.
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:
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 k, m, and ρ 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} latent steps and k∈{16,32,64} token alternatives, with ρ=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=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):
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), where s is the minimum position ID and e is one past the maximum. A receiver prefix of L valid tokens, including the opening boundary, receives IDs s−L,…,s−1, without clamping negative values. Sender KV is inserted unchanged; the closing boundary and suffix begin at e. With multiple senders, KV blocks and validity masks are concatenated in workflow order, using s=minjsj and e=maxjej 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.