- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes NegoLat, a training-free communication protocol that allows downstream agents to inspect upstream messages and request receiver-specific revisions before latent transfer. Receivers either accept a textual semantic probe or issue structured requests identifying missing, ambiguous, or conflicting information. The sender selectively regenerates the message, and the resulting hidden states are transferred through the existing StateBridge interface. Experiments on ten benchmarks covering different tasks report generally improved performance under sequential and hierarchical agent structures, with ablations supporting the main negotiation components.
Critical strengths
- The studied problem is well motivated and valuable. Allowing receivers to identify missing or conflicting information before using an upstream message is a natural and reasonable approach to improving agent collaboration.
- The proposed protocol is simple, training-free, and interpretable. Structured requests and selective regeneration provide a practical way to tailor messages to different receivers.
- The experiments cover three task families, four backbones, and two communication structures. Ablations of receiver roles, request structure, additional generation, shared payloads, and joint auditing provide useful evidence for the proposed design.
Critical weaknesses
-
I feel that the methodological novelty and comparative advantage are relatively limited. The latent transfer mechanism is inherited from StateBridge, making receiver-guided feedback the main contribution. However, comparisons are limited to Single, TextMAS, and StateBridge, without testing text-based negotiation or direct receiver-role conditioning without an audit. The unconditioned two-pass ablation is useful, but does not establish an advantage over stronger feedback-based refinement. Comparisons with communication optimization methods such as AgentPrune and G-Designer, under comparable inference budgets and with optimization costs reported separately, would further clarify the performance–cost advantage.
-
The theoretical analysis provides very limited insight into the proposed method. Theorem 1 is a direct expected-value decomposition, while Theorem 2 assumes independent recovery opportunities and preservation of recovered information, which are not established for complete message regeneration. Moreover, Appendix C.4 measures defects mentioned in requests rather than defects actually repaired. The practical justification for the request budget therefore remains limited.
-
It seems that the claim of limited overhead is not sufficiently supported by the runtime results. In Table 5, sequential MBPP+ takes 149.97 seconds with NegoLat, compared with 38.52 seconds for StateBridge and 54.62 seconds for TextMAS. Short feedback and selective regeneration do not eliminate auditing costs. The authors are suggested to qualify the efficiency claim and compare performance under matched end-to-end inference budgets.
-
The robustness and generality of the improvements require stronger evidence. All evaluated backbones have only 3B–4B parameters (especially for a training-free method), with agents sharing the same backbone and using manually designed task vocabularies. Several gains are small, and some settings show regressions. Although selected experiments include repeated runs, the main results lack uncertainty estimates, making the reliability of these improvements difficult to assess.
Questions for authors
- How does NegoLat compare with text-based negotiation, directly conditioning the sender on the receiver’s role, or feedback-based refinement, given the same inference budget?
- Could you check whether the defects identified in requests are actually fixed after regeneration, and whether new errors are introduced? Does the request-budget finding also hold for naturally occurring errors?
- Could you report uncertainty estimates for the main results and test a larger backbone? It would also help to explain where the additional latency in Table 5 comes from.