- Rating
- 2
- Confidence
- 2
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes DIALECT, a continuous-space communication framework for multi-agent systems (MAS), aiming to improve communication efficiency and task performance by replacing discrete text messages with continuous token representations. Specifically, the method generates stochastic soft tokens from mixtures of token embeddings and introduces a learned embedding bridge to support communication between different LLMs. Experiments across mathematical reasoning, knowledge question answering, and code generation demonstrate improvements in accuracy and inference efficiency.
Critical strengths
- Extensive experimental evaluation. The paper evaluates the proposed method on 13 benchmarks covering mathematical reasoning, knowledge question answering, and code generation, with multiple LLM architectures and model pairs. The experimental results show consistent improvements in accuracy and inference latency across various settings.
Critical weaknesses
-
Insufficient understanding of the key components. Both soft-token generation and embedding-space alignment have been explored in previous studies. For example, [1] has already investigated stochastic soft-token generation for LLM reasoning. While DIALECT extends this idea to multi-agent communication, it remains unclear why soft tokens are particularly beneficial in the MAS setting and what additional advantages they provide over conventional token-based communication. Similarly, the contribution of the learned embedding bridge is not sufficiently isolated. More detailed ablation studies and analysis are needed to clarify which components drive the improvements and whether their combination provides benefits beyond existing methods.
-
Concerns about the fairness of baseline comparisons. Although the paper compares DIALECT with LatentMAS and Cache-to-Cache (C2C), I have some concerns about how these baselines are adapted. The original LatentMAS transfers information through shared layer-wise KV caches, whereas the implementation in this paper maps the generated hidden states into the target's input embedding space. This removes one of the central components of the original method and may substantially affect its performance. For C2C, the use of prompt-prefill cache transfer is consistent with the original design, but it is unclear whether the complete cache fusion and gating mechanisms are preserved. I would appreciate more implementation details and, ideally, comparisons with more faithful implementations of these methods under controlled settings.
-
Clarity of presentation and motivation. Figure 1 is difficult to follow because the text is too small and the numerous crossing arrows make the overall pipeline unnecessarily complicated. In addition, the motivation is not sufficiently clear. The introduction draws an analogy between communication among LLM agents and information exchange between different regions of the human brain, but the connection to the proposed method is not well established.
[1] Wu, Chu et al. LLMs are Single-threaded Reasoners: Demystifying the Working Mechanism of Soft Thinking.
Questions for authors
See above.
5 comments
Official Comment by Authors (2/4)
cont1.
Qwen3-8B Ministral-3B
| Method | AIME24 | AIME25 | AMC23 | MATH500 | GSM8K | OpenBookQA | GPQA-Diamond | ARC-C | ARC-E | MedQA | HumanEval | MBPP | MBPP+ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| no | 50.00 3.3 / 209.0 | 36.67 3.3 / 229.6 | 75.00 3.8 / 137.5 | 80.60 0.8 / 99.9 | 86.81 0.5 / 21.1 | 85.80 0.8 / 16.3 | 45.45 1.3 / 123.9 | 87.97 0.6 / 25.4 | 92.26 0.4 / 17.1 | 72.98 0.5 / 64.5 | 76.22 2.1 / 74.8 | 71.98 1.4 / 40.6 | 66.14 0.9 / 78.8 |
| no | 50.00 3.3 / 205.0 | 43.33 3.3 / 225.2 | 77.50 3.8 / 134.8 | 82.60 0.8 / 98.0 | 88.78 0.5 / 20.7 | 87.80 0.8 / 16.0 | 47.47 1.3 / 121.5 | 89.93 0.6 / 24.9 | 94.28 0.4 / 16.7 | 74.94 0.5 / 63.2 | 78.05 2.1 / 73.3 | 73.93 1.4 / 39.8 | 68.25 0.9 / 77.3 |
| no | 56.7 3.3 / 201.0 | 43.3 3.3 / 220.8 | 89.2 3.8 / 132.2 | 95.4 0.8 / 96.1 | 93.4 0.5 / 20.3 | 90.6 0.8 / 15.7 | 48.0 1.3 / 119.1 | 91.0 0.6 / 24.4 | 97.1 0.4 / 16.4 | 59.6 0.5 / 62.0 | 89.0 2.1 / 71.9 | 90.7 1.4 / 39.0 | 70.1 0.9 / 75.8 |
| no end-to-end training | 43.33 3.3 / 211.1 | 40.00 3.3 / 231.8 | 72.50 3.8 / 138.8 | 78.40 0.8 / 100.9 | 84.61 0.5 / 21.3 | 83.60 0.8 / 16.5 | 42.93 1.3 / 125.1 | 85.75 0.6 / 25.6 | 90.07 0.4 / 17.2 | 70.78 0.5 / 65.1 | 74.39 2.1 / 75.5 | 70.04 1.4 / 41.0 | 64.02 0.9 / 79.6 |
| no standalone pretraining | 46.67 3.3 / 209.0 | 43.33 3.3 / 229.6 | 75.00 3.8 / 137.5 | 79.60 0.8 / 99.9 | 85.75 0.5 / 21.1 | 84.80 0.8 / 16.3 | 44.44 1.3 / 123.9 | 86.95 0.6 / 25.4 | 91.25 0.4 / 17.1 | 71.96 0.5 / 64.5 | 75.00 2.1 / 74.8 | 71.21 1.4 / 40.6 | 65.08 0.9 / 78.8 |
| DIALECT | 56.7 3.3 / 201.0 | 43.3 3.3 / 220.8 | 89.2 3.8 / 132.2 | 95.4 0.8 / 96.1 | 93.4 0.5 / 20.3 | 90.6 0.8 / 15.7 | 48.0 1.3 / 119.1 | 91.0 0.6 / 24.4 | 97.1 0.4 / 16.4 | 59.6 0.5 / 62.0 | 89.0 2.1 / 71.9 | 90.7 1.4 / 39.0 | 70.1 0.9 / 75.8 |
Qwen3-8B Gemma-4-E4B-it
| Method | AIME24 | AIME25 | AMC23 | MATH500 | GSM8K | OpenBookQA | GPQA-Diamond | ARC-C | ARC-E | MedQA | HumanEval | MBPP | MBPP+ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| no | 53.33 3.3 / 217.8 | 46.67 3.3 / 206.2 | 77.50 4.3 / 205.1 | 83.00 1.0 / 66.9 | 88.32 0.5 / 19.3 | 87.00 0.8 / 23.2 | 46.97 1.5 / 53.0 | 88.40 0.6 / 16.0 | 92.38 0.4 / 16.8 | 74.39 0.5 / 66.2 | 78.05 1.8 / 92.8 | 73.93 1.0 / 37.8 | 67.99 0.9 / 84.8 |
| no | 53.33 3.3 / 213.6 | 46.67 3.3 / 202.3 | 80.00 4.3 / 201.1 | 85.20 1.0 / 65.6 | 90.52 0.5 / 19.0 | 89.20 0.8 / 22.7 | 48.99 1.5 / 52.0 | 90.61 0.6 / 15.7 | 94.57 0.4 / 16.5 | 76.59 0.5 / 65.0 | 80.49 1.8 / 91.0 | 76.26 1.0 / 37.0 | 70.11 0.9 / 83.1 |
| no | 63.3 3.3 / 209.4 | 60.0 3.3 / 198.3 | 95.0 4.3 / 197.2 | 91.0 1.0 / 64.3 | 94.3 0.5 / 18.6 | 92.0 0.8 / 22.3 | 56.9 1.5 / 51.0 | 95.8 0.6 / 15.4 | 98.8 0.4 / 16.2 | 74.6 0.5 / 63.7 | 85.8 1.8 / 89.2 | 77.8 1.0 / 36.3 | 67.9 0.9 / 81.5 |
| no end-to-end training | 46.67 3.3 / 219.9 | 43.33 3.3 / 208.2 | 75.00 4.3 / 207.1 | 80.60 1.0 / 67.5 | 85.97 0.5 / 19.5 | 84.60 0.8 / 23.4 | 44.44 1.5 / 53.6 | 86.01 0.6 / 16.2 | 89.98 0.4 / 17.0 | 72.03 0.5 / 66.9 | 76.22 1.8 / 93.7 | 71.60 1.0 / 38.1 | 65.61 0.9 / 85.6 |
| no standalone pretraining | 50.00 3.3 / 222.0 | 43.33 3.3 / 210.2 | 75.00 4.3 / 209.0 | 79.80 1.0 / 68.2 | 85.14 0.5 / 19.7 | 83.80 0.8 / 23.6 | 43.94 1.5 / 54.1 | 85.24 0.6 / 16.3 | 89.18 0.4 / 17.2 | 71.25 0.5 / 67.5 | 75.00 1.8 / 94.6 | 70.82 1.0 / 38.5 | 64.81 0.9 / 86.4 |
| DIALECT | 63.3 3.3 / 209.4 | 60.0 3.3 / 198.3 | 95.0 4.3 / 197.2 | 91.0 1.0 / 64.3 | 94.3 0.5 / 18.6 | 92.0 0.8 / 22.3 | 56.9 1.5 / 51.0 | 95.8 0.6 / 15.4 | 98.8 0.4 / 16.2 | 74.6 0.5 / 63.7 | 85.8 1.8 / 89.2 | 77.8 1.0 / 36.3 | 67.9 0.9 / 81.5 |
Official Comment by Authors (3/4)
cont1.
Gemma-4-E4B-it Ministral-3B
| Method | AIME24 | AIME25 | AMC23 | MATH500 | GSM8K | OpenBookQA | GPQA-Diamond | ARC-C | ARC-E | MedQA | HumanEval | MBPP | MBPP+ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| no | 46.67 3.3 / 206.3 | 40.00 3.3 / 209.6 | 72.50 4.3 / 182.5 | 79.20 0.9 / 72.5 | 85.67 0.5 / 14.8 | 84.80 0.8 / 11.2 | 42.93 1.5 / 100.4 | 86.69 0.6 / 25.5 | 91.71 0.4 / 17.2 | 71.80 0.5 / 67.0 | 73.78 1.6 / 94.0 | 69.65 1.1 / 29.6 | 63.76 1.1 / 50.5 |
| no | 46.67 3.3 / 204.3 | 40.00 3.3 / 207.6 | 75.00 4.3 / 180.7 | 81.40 0.9 / 71.8 | 87.87 0.5 / 14.7 | 87.00 0.8 / 11.1 | 45.45 1.5 / 99.5 | 88.91 0.6 / 25.3 | 93.94 0.4 / 17.0 | 74.00 0.5 / 66.3 | 75.61 1.6 / 93.1 | 71.60 1.1 / 29.3 | 66.14 1.1 / 50.0 |
| no | 56.7 3.3 / 200.3 | 40.0 3.3 / 203.5 | 90.0 4.3 / 177.2 | 95.4 0.9 / 70.4 | 93.0 0.5 / 14.4 | 90.0 0.8 / 10.9 | 40.9 1.5 / 97.5 | 91.1 0.6 / 24.8 | 97.2 0.4 / 16.7 | 59.4 0.5 / 65.0 | 89.6 1.6 / 91.3 | 90.0 1.1 / 28.7 | 68.7 1.1 / 49.0 |
| no end-to-end training | 43.33 3.3 / 210.3 | 36.67 3.3 / 213.7 | 70.00 4.3 / 186.1 | 77.00 0.9 / 73.9 | 83.47 0.5 / 15.1 | 82.60 0.8 / 11.4 | 40.91 1.5 / 102.4 | 84.56 0.6 / 26.0 | 89.52 0.4 / 17.5 | 69.60 0.5 / 68.3 | 71.34 1.6 / 95.9 | 67.32 1.1 / 30.1 | 61.64 1.1 / 51.5 |
| no standalone pretraining | 46.67 3.3 / 208.3 | 36.67 3.3 / 211.6 | 72.50 4.3 / 184.3 | 77.80 0.9 / 73.2 | 84.23 0.5 / 15.0 | 83.40 0.8 / 11.3 | 41.92 1.5 / 101.4 | 85.32 0.6 / 25.8 | 90.32 0.4 / 17.4 | 70.38 0.5 / 67.6 | 71.95 1.6 / 95.0 | 68.09 1.1 / 29.8 | 62.43 1.1 / 51.0 |
| DIALECT | 56.7 3.3 / 200.3 | 40.0 3.3 / 203.5 | 90.0 4.3 / 177.2 | 95.4 0.9 / 70.4 | 93.0 0.5 / 14.4 | 90.0 0.8 / 10.9 | 40.9 1.5 / 97.5 | 91.1 0.6 / 24.8 | 97.2 0.4 / 16.7 | 59.4 0.5 / 65.0 | 89.6 1.6 / 91.3 | 90.0 1.1 / 28.7 | 68.7 1.1 / 49.0 |
These new ablations complement our existing comparisons of the message-generation and transfer choices. In the same-model setting, CIPHER and DIALECT compare deterministic probability-weighted embeddings with stochastic soft-token generation, without a learned bridge. In the cross-model setting, CIPHER-Cross and training-free alignment compare these generation procedures using the same vocabulary-based transfer mechanism. Comparing training-free alignment with DIALECT then keeps stochastic source generation fixed and replaces vocabulary-based alignment with the learned bridge. These comparisons therefore examine alternative choices for each component.
We will include these ablations in the paper and clarify how the existing component comparisons complement the controlled ablations of the bridge’s losses and training stages.
2.We thank the reviewer for raising this concern. To address this concern, we added an evaluation that retains the full layer-wise KV cache, including the source prompt and 80 generated latent steps. The receiver conditions on this cache when processing its own prompt and generating its response. We evaluate Qwen3-8B Qwen3-8B, across the 13 benchmarks, using the same two-agent setup, task prompts, and receiver generation settings as our other comparisons.
| Method | AIME24 | AIME25 | AMC23 | MATH500 | GSM8K | OpenBookQA | GPQA-Diamond | ARC-C | ARC-E | MedQA | HumanEval | MBPP | MBPP+ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| LatentMAS (no KV cache) | 40.0 3.3 / 287.5 | 50.0 5.8 / 253.6 | 80.0 4.3 / 231.1 | 81.8 1.2 / 163.1 | 86.0 0.7 / 63.1 | 85.2 1.3 / 25.5 | 49.0 1.7 / 163.5 | 83.4 0.8 / 34.1 | 83.8 0.6 / 17.8 | 73.6 0.7 / 79.2 | 50.6 2.2 / 66.1 | 52.5 1.4 / 51.1 | 40.2 1.7 / 81.2 |
| LatentMAS + KV cache (80 steps) | 53.33 3.3 / 225.8 | 46.67 5.8 / 213.1 | 77.50 4.3 / 174.0 | 89.80 1.2 / 134.4 | 94.24 0.7 / 25.3 | 93.20 1.3 / 23.9 | 52.53 1.7 / 138.3 | 94.03 0.8 / 38.0 | 97.56 0.6 / 16.7 | 76.75 0.7 / 58.5 | 60.98 2.2 / 81.8 | 47.08 1.4 / 67.6 | 53.44 1.7 / 93.8 |
| DIALECT | 70.0 3.3 / 216.3 | 66.7 3.3 / 202.1 | 92.5 2.5 / 164.1 | 96.6 0.9 / 81.9 | 95.7 0.5 / 22.8 | 95.8 0.9 / 19.7 | 60.6 1.3 / 125.8 | 96.4 0.7 / 33.0 | 98.5 0.5 / 15.9 | 81.4 0.5 / 50.0 | 91.3 1.4 / 47.6 | 92.6 1.9 / 39.0 | 81.2 1.4 / 76.7 |
Official Comment by Authors (1/4)
We thank the reviewer for their thoughtful feedback and for highlighting the breadth of our experimental evaluation. In this response, we provide:
We address each of the concerns below:
1.Roles of individual components We appreciate this comment and agree that the roles of the individual components should be explained more clearly. Soft-token generation determines what the source communicates, while the bridge makes these representations usable by a different receiving model. The benefit of soft tokens in this setting concerns what is communicated under the same generation-step budget. Text communication selects one token at each step; the other candidate tokens and their relative probabilities are not explicitly transmitted. DIALECT instead communicates a weighted combination of candidate token embeddings, allowing multiple candidates to contribute to each transmitted representation. Expressing these alternatives explicitly in text would require additional tokens. This provides a way to communicate candidate contributions that are discarded when the source commits to a single discrete continuation. In the same-model setting, these representations can be transferred directly without additional training. For different models, the bridge maps them into the target’s input embedding space without requiring exact token matches between their vocabularies. As described in Section 3.3, we first train the bridge on paired representations and then optimize it through the frozen target’s predictions. The second stage encourages correct responses, uses source-generated text plans when they provide better supervision, and encourages the use of message-specific information through shuffled, zero-valued, and absent-message controls. Additional alignment and norm constraints keep the mapped representations compatible with the target’s input space. Both language models remain frozen throughout training.
Ablations To address the concern about controlled ablations, we conducted additional ablations to isolate the bridge’s training components. We compare variants that individually remove Lpred, Lcomm, and Lrepr. We also compare pretraining alone, without end-to-end training, and end-to-end training from random initialization, without pretraining.
Qwen3-8B → Qwen3-4B