acceptodds
Withdrawn from ICLR 2027

Good Agentic Friends Do Not Just Give Verbal Advice: They Can Update Your Weights

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multi-agent LLM systems usually collaborate by exchanging natural-language messages. This interface is simple and interpretable, but it forces each sender's intermediate computation to be serialized into tokens and then reprocessed by the receiver, thereby increasing the generated-token cost, prefill overhead, and KV-cache memory. We study an alternative communication interface: instead of appending a sender's message to the receiver's context, compile the sender's hidden states into a transient, receiver-specific weight perturbation. We introduce **TFlow** (Thought Flow), a weight-space communication framework for a known and fixed receiver architecture. For each query, frozen role-prompted sender agents process the input, and a learned parameter generator maps their internal activations into low-rank LoRA perturbations targeting the receiver's modules. These perturbations are fused and applied only during the receiver's generation, enabling instance-level adaptation without permanently changing the model or enlarging the receiver's text context. With three Qwen3-4B agents, **TFlow** improves over a standalone receiver by up to 8.5 accuracy points across five benchmarks while reducing processed tokens by up to 32.69%. Compared with a text-based three-agent baseline, it reduces total processed tokens by up to 83.27% and the wall-clock inference time by up to 4.6, while maintaining competitive accuracy on four of five benchmarks. These results suggest that transient low-rank weight perturbations can serve as an executable communication medium for efficient multi-agent LLM collaboration.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multi-Agent Systems, Large Language Models, Weight-Space Communication, Parameter Generation, Agent Collaboration

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_tc2XOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TFlow (Thought Flow), a weight-space communication framework for multi-agent LLM collaboration that replaces conventional natural-language message passing with instance-specific parameter perturbations. Specifically, frozen sender agents extract hidden-state representations of the input, which are transformed into LoRA perturbations by a trainable parameter generator. These perturbations are fused and temporarily injected into a frozen receiver to guide its generation without expanding the textual context. The framework is trained with a next-token prediction loss and a diversity regularizer. Experiments across five reasoning, knowledge, and coding benchmarks show that TFlow consistently improves accuracy over a single-agent baseline while substantially reducing token consumption and inference latency compared to a text-based multi-agent system, albeit with generally lower accuracy.

Critical strengths

(1) The paper revisits multi-agent collaboration from the perspective of communication interfaces, shifting information exchange from explicit text or hidden-state transfer toward dynamic modulation of the receiver's computation. This perspective is conceptually interesting and offers a promising direction for exploring deeper forms of inter-model collaboration. (2) The methodological design is well aligned with the underlying motivation. By translating sender representations into instance-conditioned parameter perturbations, the framework enables dynamic information injection without modifying the frozen backbone or expanding the receiver's textual context, offering a practical approach to efficient collaboration. (3) Beyond evaluating downstream task performance, the paper investigates the role of dynamic weight generation through static adaptation, mismatched perturbation injection, and representation similarity analyses. These experiments provide useful empirical insights into instance-conditioned parameter modulation, extending the evaluation beyond conventional performance comparisons.

Critical weaknesses

(1) The paper frames instance-conditioned weight perturbations as a multi-agent communication mechanism, yet the sender agents merely perform forward passes on the same input under different role prompts. The distinction between this process and conventional conditional parameter adaptation remains unclear, as does the independent or complementary information contributed by different agents. Consequently, the conceptual foundation of the proposed multi-agent collaboration paradigm is insufficiently established. (2) The method largely builds upon existing techniques in hypernetworks, dynamic LoRA, and structured parameter generation, while its substantive technical novelty remains unclear. Moreover, several design choices lack sufficient justification. For instance, the diversity regularizer encourages orthogonal perturbations across samples without considering that semantically similar tasks may benefit from shared update directions, leaving its underlying rationale and necessity insufficiently established. (3) The experimental evidence does not adequately support the paper's central claims. Existing ablations primarily demonstrate the effectiveness of dynamic parameter adaptation without isolating the contribution of multi-agent communication. In particular, a single-agent dynamic LoRA baseline is missing, and the supervision conditions differ between TFlow and its main baselines. Furthermore, the absence of direct comparisons with existing latent communication methods leaves the distinctive advantages of the proposed mechanism unclear.

Questions for authors

(1) What evidence demonstrates that TFlow benefits from genuine multi-agent communication rather than conventional query-conditioned parameter adaptation? In particular, how would it compare with a similarly trained parameter generator conditioned on a single agent's or the receiver's hidden states? Do different sender roles contribute meaningful complementary information? (2) How can the benefits of the proposed communication mechanism be disentangled from those of additional supervised training? Given that TFlow's parameter generator is trained while the main baselines remain frozen, would the conclusions hold under comparable supervision? Can the authors also provide direct comparisons with existing latent-space communication methods? (3) What constitutes the key technical advancement over existing hypernetwork-based and dynamic LoRA generation approaches, particularly HY-WU? Beyond changing the conditioning signal, which architectural components are essential to the observed improvements, and what motivates their specific design? (4) Although the paper reports end-to-end inference latency, can the authors further break down the computational costs of sender forwarding, parameter generation, and dynamic LoRA injection? How do the reported efficiency advantages hold up when considering memory consumption and batched inference throughput?

Reviewer_Gr5hOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TFlow, a novel multi-agent framework that converts sender agents' hidden states into query-specific LoRA for a frozen receiver model. The receiver model loads this LoRa and answers the original query. Experiments on five benchmarks demonstrate consistent improvements over a single-agent baseline and substantial reductions in token consumption and inference latency compared to text-based multi-agent systems.

Critical strengths
  1. The paper introduces weight-space communication as an interesting alternative to text-based and latent-space communication

  2. TFlow reduces token consumption and achieves speedups over TextMAS while maintaining competitive accuracy on most benchmarks.

Critical weaknesses
  1. I am not quite convinced why this setup could help improve agent scores on benchmarks. it is unclear whether the improvements stem from meaningful multi-agent communication or simply from effective input-conditioned parameter generation

  2. The single-agent baseline performs substantially worse than published Qwen3 results particularly on MATH. I understand the reason might be that qwen official evaluates on k-shot and the users use zero-shot (if not, please correct me). Could the authors provide k-shot eval for their multi-agent as well?

  3. All experiments use agents sharing the same Qwen3-4B backbone and a fixed receiver architecture. It remains unclear whether the approach generalizes to heterogeneous models or different receiver architectures. Additionally, TFlow consistently underperforms TextMAS in accuracy.

Questions for authors

See weaknesses.

Reviewer_5rAcOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

TFlow converts hidden states from two role-prompted senders into temporary LoRA updates for a receiver sharing the same frozen Qwen3-4B backbone. A receiver-specific hypernetwork generates and combines these updates, trained with answer supervision and a diversity regularizer, so the receiver answers without consuming textual sender messages. The paper reports improved accuracy over a standalone model and substantially lower latency than text-based collaboration, with useful adaptation ablations. This is a promising application of conditional weight generation, but discrepancies between the manuscript and released evaluation pipeline, together with insufficient isolation of the multi-role contribution, limit the strength of its conclusions.

Critical strengths

1. A concrete and potentially useful communication design. The method removes sender autoregressive decoding and receiver prefill of sender messages while preserving a frozen backbone and an explicitly fixed receiver architecture (Sections 3.1–3.4). Table 1 reports 2.31–4.58× lower wall-clock inference time than TextMAS, with accuracy gaps of 1.32–4.53 points on four tasks and 9.76 on HumanEval+. The paper also acknowledges that TFlow is slower than the standalone model on four tasks. This makes the work a potentially useful exploration of the quality–cost tradeoff, even without a fundamentally new hypernetwork mechanism.

2. Meaningful evidence that the generated adaptations are useful. The static-LoRA comparison controls training data, backbone and target modules (Figure 4); mismatched-update injection tests whether the conditioning matters (Figure 5); and removing the diversity regularizer reduces average accuracy by 4.29 points on three tasks (Table 3). Together, these provide more substantive evidence than benchmark improvements alone, although they do not fully establish the authors' multi-agent interpretation.

Critical weaknesses

1. The headline efficiency result and evaluation protocol are not yet supported by a consistent, auditable comparison. In the released code, TFlow executes both sender prefills but returns only the receiver's token usage, whereas TextMAS sums all three agents' prompt and completion tokens. This cannot substantiate comparable “total processed tokens” or the stated 83.27% reduction. Wall-clock measurement does include sender work, so this finding does not itself invalidate the speedups. However, the receiver prompts also differ: TFlow explicitly encourages concise reasoning and discourages code tracing, confounding the attribution of shorter outputs to weight modulation. Moreover, Section 4.1 specifies zero-shot Hendrycks MATH, while the documented evaluator defaults to four-shot Minerva Math; Eq. (2)'s learned fusion gate becomes fixed 0.5/0.5 weights in the release. These are verified implementation discrepancies, not proof of which pipeline produced Table 1. Establishing that provenance and evaluating with matched prompts and complete accounting are consequential requirements, rather than cosmetic corrections.

2. The evidence does not isolate the proposed multi-agent contribution from ordinary conditional adaptation. Generative Adapter (Chen et al., 2024), absent from the submitted paper's references, already generates low-rank updates from frozen-LM activations; Doc-to-LoRA (Charakorn et al., 2026), which the submitted paper cites, does so through cross-attention. TFlow's distinctive application is therefore the fusion of role-conditioned query encodings for one answer, rather than activation-to-weight conversion itself. Yet the senders only prefill the same query, and no control isolates the benefit of their distinct roles relative to a comparably trained generator using one neutral or receiver encoding. Static LoRA is valuable but does not match hypernetwork capacity; moreover, summing two rank-4 updates can produce rank up to 8. Figure 5's matched update scores 92.12 versus 90.83 for a same-task mismatch: this supports some instance dependence, but most improvement over 84.99 remains under mismatch. The synthesis could be worthwhile, but its ambitious communication interpretation requires more direct evidence; the existing ablations do not resolve that central uncertainty.

References

Charakorn, R., Cetin, E., Uesaka, S., & Lange, R. T. (2026). Doc-to-LoRA: Learning to instantly internalize contexts. arXiv preprint arXiv:2602.15902. https://doi.org/10.48550/arXiv.2602.15902

Chen, T., Fang, H., Xia, P., Liu, X., Van Durme, B., Zettlemoyer, L., Gao, J., & Cheng, H. (2024). Generative Adapter: Contextualizing language models in parameters with a single forward pass. arXiv preprint arXiv:2411.05877. https://doi.org/10.48550/arXiv.2411.05877

Questions for authors
  1. Which exact code revision, dataset IDs/revisions and sample counts, shot counts, prompts, and fusion rule produced Table 1? How do those runs relate to the four-shot Minerva loader and fixed sender weights in the supplement?

  2. Do Table 1's TFlow token totals include both sender prefills? With complete all-agent accounting, identical receiver prompts, and disclosed batch sizes, what accuracy/token/latency tradeoff remains?

  3. What evidence establishes that two distinct sender roles outperform a single neutral/receiver-conditioned generator under matched training, generator capacity and effective update rank?

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

Related papers

Discussion (0)

Sign in to comment.