{"paper":{"title":"Good Agentic Friends Do Not Just Give Verbal Advice: They Can Update Your Weights","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Multi-agent LLMs can collaborate by mapping sender activations directly into transient low-rank weight updates on the receiver instead of passing text messages.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Huan Wang, Jian Wang, Kai Wang, Wenrui Bao, Yuzhang Shang, Zhangyang Wang","submitted_at":"2026-05-13T17:58:32Z","abstract_excerpt":"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 communica"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"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.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That a learned parameter generator, trained once, can map arbitrary sender activations into effective, stable, receiver-specific LoRA perturbations for every new query without overfitting or degrading generation quality.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"TFlow enables multi-agent LLMs to collaborate via transient low-rank LoRA perturbations derived from sender activations, yielding up to 8.5 accuracy gains and 83% token reduction versus text-based baselines on Qwen3-4B models.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Multi-agent LLMs can collaborate by mapping sender activations directly into transient low-rank weight updates on the receiver instead of passing text messages.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"a612ec25da6fa95aef53f882f09dc8efb546770245bd71d1e72fa60790ec43e2"},"source":{"id":"2605.13839","kind":"arxiv","version":1},"verdict":{"id":"b63380da-48f3-4603-80fa-d63178e0ec61","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T18:57:09.200161Z","strongest_claim":"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.","one_line_summary":"TFlow enables multi-agent LLMs to collaborate via transient low-rank LoRA perturbations derived from sender activations, yielding up to 8.5 accuracy gains and 83% token reduction versus text-based baselines on Qwen3-4B models.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That a learned parameter generator, trained once, can map arbitrary sender activations into effective, stable, receiver-specific LoRA perturbations for every new query without overfitting or degrading generation quality.","pith_extraction_headline":"Multi-agent LLMs can collaborate by mapping sender activations directly into transient low-rank weight updates on the receiver instead of passing text messages."},"references":{"count":58,"sample":[{"doi":"","year":2023,"title":"Li, G., H. Hammoud, H. Itani, et al. Camel: Communicative agents for\" mind\" exploration of large language model society. InNeurIPS. 2023","work_id":"4d095bca-8e26-4f15-9003-bac7eafebb9c","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Hong, S., M. Zhuge, J. Chen, et al. Metagpt: Meta programming for a multi-agent collaborative framework. InICLR. 2023","work_id":"fbd9997d-1af3-4426-97df-90a48d7c0816","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Wu, Q., G. Bansal, J. Zhang, et al. AutoGen: Enabling next-gen LLM applications via multi- agent conversations. InCOLM. 2024","work_id":"544eb61c-a62f-4511-8b40-6b43090e3ef8","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Du, Y ., S. Li, A. Torralba, et al. Improving factuality and reasoning in language models through multiagent debate. InICML. 2024","work_id":"15f47422-cc65-4cb2-8f36-9938690c4dd3","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Liang, T., Z. He, W. Jiao, et al. Encouraging divergent thinking in large language models through multi-agent debate. InEMNLP. 2024","work_id":"e50ee9c8-7446-4b89-8f32-1e1571001335","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":58,"snapshot_sha256":"0ccd6656dfff0bfa309637122014ef8a4d5e70e610e52c34a145e15d75496ffa","internal_anchors":11},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}