TL;DR
RecursiveMAS transforms multi-agent collaboration from text to latent space using RecursiveLink modules, achieving 8.3% accuracy improvement, 1.2×-2.4× speedup, and 34.6%-75.6% token reduction across 9 benchmarks without parameter updates.
RecursiveMAS transforms multi-agent collaboration from text to latent space using RecursiveLink modules, achieving 8.3% accuracy improvement, 1.2×-2.4× speedup, and 34.6%-75.6% token reduction across 9 benchmarks without parameter updates.
31
已证实
0
证据不足
0
无法验证
N/A
可复现性
核心问题
Can treating multi-agent collaboration as recursive computation in continuous latent space, rather than text-based interaction, improve performance and efficiency across diverse reasoning tasks?
核心方法
{'approach': 'The framework connects heterogeneous LLM agents through lightweight RecursiveLink modules (inner links for within-agent latent thoughts, outer links for cross-agent communication) and trains them via an Inner-Outer Loop paradigm that freezes base LLM parameters while optimizing only the RecursiveLink modules. Agents are chained into recursive loops where each acts as an RLM layer, passing latent representations through multiple recursion rounds before final decoding.', 'key_components': ['RecursiveMAS is instantiated with four collaboration styles: Sequential, Mixture, Distillation, and Deliberation.', 'Heterogeneous agent compositions are constructed using LLMs from Qwen, Llama, Gemma, and Mistral families.', 'Single Advanced Agents baseline isolates individual LLM agents as standalone models with fine-tuning.', 'Recursion-based baselines include LoopLM and Recursive-TextMAS (text-based agent collaboration).', 'Representative Multi-Agent Frameworks such as TextGrad are included for comparison.', 'RecursiveMAS achieves 8.3% average performance improvement over the strongest baseline on each benchmark.', 'All methods are instantiated with identical backbone models and comparable training budgets for fair comparison.', 'Fine-tuning individual agents improves performance, but RecursiveMAS delivers further gains through system-level optimization.', 'Particularly strong gains are achieved on reasoning-intensive tasks: 18.1% on AIME2025 and 13.0% on AIME2026.', 'RecursiveMAS outperforms advanced architectures including TextGrad and LoopLM.'], 'section_ids': ['sec_15', 'sec_17']}
论点验证
已证实 (85%)
We call this new system-level agentic recursion framework RecursiveMAS. Without updating all model parameters, agents are connected and iteratively optimized solely via the lightweight RecursiveLink, a two-layer residual projection module for latent states transmission and refinement.
The paper provides complete architectural specification of RecursiveMAS with RecursiveLink detailed in paragraphs 20-24, including equations for the two-layer residual projection module. The claim that only RecursiveLink parameters are updated is exp
The paper provides complete architectural specification of RecursiveMAS with RecursiveLink detailed in paragraphs 20-24, including equations for the two-layer residual projection module. The claim that only RecursiveLink parameters are updated is exp
已证实 (85%)
An inner RecursiveLink within each agent first consolidates the model's ongoing latent thoughts between input and output spaces during auto-regressive generation. An outer RecursiveLink then bridges hidden representations across heterogeneous agents built on different model types and sizes, enabling seamless cross-agent interaction.
The paper provides complete specifications for both inner and outer RecursiveLink with mathematical formulations. Inner link equation in paragraph 21-22, outer link equation in paragraph 23. The cross-model transition capability is demonstrated with
The paper provides complete specifications for both inner and outer RecursiveLink with mathematical formulations. Inner link equation in paragraph 21-22, outer link equation in paragraph 23. The cross-model transition capability is demonstrated with
已证实 (80%)
Correspondingly, we pair RecursiveMAS with an Inner-Outer Loop training paradigm for progressive co-optimization.
The Inner-Outer Loop training paradigm is described in detail in paragraphs 28-32. The inner loop (paragraph 29) and outer loop (paragraph 31) training objectives are specified with equations. The progressive co-optimization is demonstrated through e
The Inner-Outer Loop training paradigm is described in detail in paragraphs 28-32. The inner loop (paragraph 29) and outer loop (paragraph 31) training objectives are specified with equations. The progressive co-optimization is demonstrated through e
已证实 (80%)
The inner loop provides a preliminary model-level warm start for each agent, by training its inner RecursiveLink to better align with latent thoughts generation. The outer loop then trains the outer RecursiveLink across agents at the system-level, with gradients recursively backpropagated through the full computation traces over recursion rounds.
The training methodology is fully specified with equations for both loops. Inner loop uses cosine similarity objective (paragraph 29), outer loop uses cross-entropy with gradient backpropagation through full recursive paths (paragraphs 31-32). The cl
The training methodology is fully specified with equations for both loops. Inner loop uses cosine similarity objective (paragraph 29), outer loop uses cross-entropy with gradient backpropagation through full recursive paths (paragraphs 31-32). The cl
已证实 (75%)
To justify why recursion should occur in latent space rather than text-mediated interaction, we provide two theoretical analyses on runtime complexity and learning dynamics.
The paper provides two theoretical analyses: Proposition 3.1 for runtime complexity (paragraph 26-27) and gradient propagation analysis for learning dynamics (paragraph 33-34, with full proof in Appendix A). The runtime complexity comparison shows Θ(
The paper provides two theoretical analyses: Proposition 3.1 for runtime complexity (paragraph 26-27) and gradient propagation analysis for learning dynamics (paragraph 33-34, with full proof in Appendix A). The runtime complexity comparison shows Θ(
已证实 (80%)
Compared with advanced recursive language models and MAS baselines, RecursiveMAS achieves an average accuracy improvement of 8.3%, while delivering 1.2×-2.4× inference speedup and reducing token usage by 34.6%-75.6%.
The paper provides concrete quantitative evidence: 8.3% average accuracy improvement in paragraph 41 and Table 3; 1.2×-2.4× speedup in paragraph 45 and Figure 5; 34.6%-75.6% token reduction in paragraph 46-47 and Figure 6. All three metrics have spec
The paper provides concrete quantitative evidence: 8.3% average accuracy improvement in paragraph 41 and Table 3; 1.2×-2.4× speedup in paragraph 45 and Figure 5; 34.6%-75.6% token reduction in paragraph 46-47 and Figure 6. All three metrics have spec
已证实 (80%)
We introduce RecursiveMAS, an end-to-end recursive framework that links heterogeneous LLM agents together to scale the entire system through efficient and seamless latent collaboration.
The framework is fully specified with architectural details (paragraphs 20-26), training methodology (paragraphs 28-32), and comprehensively evaluated on 9 benchmarks with 4 collaboration patterns. The end-to-end nature is demonstrated through the co
The framework is fully specified with architectural details (paragraphs 20-26), training methodology (paragraphs 28-32), and comprehensively evaluated on 9 benchmarks with 4 collaboration patterns. The end-to-end nature is demonstrated through the co
已证实 (80%)
The RecursiveLink R is designed to preserve and transmit this information from one embedding space to another. In RecursiveMAS, the transition arises in two cases: (i) Dense-to-Shallow Transition, where the previous step's last-layer embeddings are fed back as the next-step input embeddings during latent thoughts generation; and (ii) Cross-Model Transition, where one model's newly generated latent representations are passed as conditioning inputs to another model.
The two transition cases are clearly defined in paragraph 20 with detailed specifications following. Dense-to-Shallow transition is handled by inner RecursiveLink (paragraph 21-22), Cross-Model transition by outer RecursiveLink (paragraph 23). The de
The two transition cases are clearly defined in paragraph 20 with detailed specifications following. Dense-to-Shallow transition is handled by inner RecursiveLink (paragraph 21-22), Cross-Model transition by outer RecursiveLink (paragraph 23). The de
已证实 (80%)
The residual branch largely preserves the original semantics of the input, allowing the RecursiveLink network to focus on aligning distributional differences rather than learning the full projection from scratch. This leads to more stable and efficient training.
The design rationale is provided in paragraph 24, and empirically validated in Table 4 (paragraph 48) showing that the 2-layer residual design performs best. The specific example shows residual connection improving single-layer design from 63.2% to 6
The design rationale is provided in paragraph 24, and empirically validated in Table 4 (paragraph 48) showing that the 2-layer residual design performs best. The specific example shows residual connection improving single-layer design from 63.2% to 6
已证实 (75%)
Without RecursiveLink, a text-based Recursive MAS with the same collaboration structure requires runtime complexity of Θ(Nm|V|dh), while RecursiveMAS achieves Θ(Nmd²h).
Proposition 3.1 in paragraph 26 provides the runtime complexity analysis. The text-based approach has Θ(Nm|V|dh) while RecursiveMAS achieves Θ(Nmd²h). Remark 3.2 in paragraph 27 explains that since dh ≪ |V| in practice, the latent-space transformatio
Proposition 3.1 in paragraph 26 provides the runtime complexity analysis. The text-based approach has Θ(Nm|V|dh) while RecursiveMAS achieves Θ(Nmd²h). Remark 3.2 in paragraph 27 explains that since dh ≪ |V| in practice, the latent-space transformatio
已证实 (85%)
We instantiate RecursiveMAS with diverse agent collaboration patterns, including (i) Sequential Style, (ii) Mixture Style, (iii) Distillation Style, and (iv) Deliberation Style.
The four collaboration patterns are described in paragraph 16, paragraph 36, and detailed in Table 1. Each pattern is evaluated with results reported in Tables 2, 3 and Appendix tables 6, 7, 8. The patterns cover diverse MAS architectures.
The four collaboration patterns are described in paragraph 16, paragraph 36, and detailed in Table 1. Each pattern is evaluated with results reported in Tables 2, 3 and Appendix tables 6, 7, 8. The patterns cover diverse MAS architectures.
已证实 (85%)
We evaluate RecursiveMAS against (i) Single Advanced Agents, where individual LLM agents from each collaboration pattern are isolated as standalone models to solve problems, (ii) Recursion-based Methods, including single recursive language models, LoopLM, and Recursive-TextMAS, and (iii) additional Representative Multi-Agent Frameworks, including TextGrad.
The baseline categories are clearly defined in paragraphs 37-38 and detailed in paragraphs 74-80. Single agents (paragraph 75-76), recursion-based methods including LoopLM and Recursive-TextMAS (paragraph 79-80), and multi-agent frameworks including
The baseline categories are clearly defined in paragraphs 37-38 and detailed in paragraphs 74-80. Single agents (paragraph 75-76), recursion-based methods including LoopLM and Recursive-TextMAS (paragraph 79-80), and multi-agent frameworks including
已证实 (80%)
For fair comparison, we provide full supervised and LoRA fine-tuning for single models on the same training set.
The fair comparison methodology is mentioned in paragraph 38 and detailed in paragraphs 75-76, which describe LoRA and full supervised fine-tuning for single models using the same training set as RecursiveMAS.
The fair comparison methodology is mentioned in paragraph 38 and detailed in paragraphs 75-76, which describe LoRA and full supervised fine-tuning for single models using the same training set as RecursiveMAS.
已证实 (80%)
Overall, RecursiveMAS delivers a consistent whole-system advantage, achieving an average performance improvement of 8.3% over the strongest baseline on each benchmark.
The 8.3% average improvement is stated in paragraph 41 and supported by Table 3 results. The paper provides benchmark-by-benchmark accuracy numbers comparing RecursiveMAS against baselines. However, no error bars or statistical significance tests are
The 8.3% average improvement is stated in paragraph 41 and supported by Table 3 results. The paper provides benchmark-by-benchmark accuracy numbers comparing RecursiveMAS against baselines. However, no error bars or statistical significance tests are
已证实 (80%)
RecursiveMAS remains the performance advantage compared to advanced architectures such as TextGrad and LoopLM, especially on reasoning-intensive tasks (e.g., accuracy gains of 18.1% on AIME2025, 13.0% on AIME2026, and 5.4% on GPQA-Diamond).
Specific accuracy gains are provided in paragraph 42: 18.1% on AIME2025, 13.0% on AIME2026, and 5.4% on GPQA-Diamond compared to TextGrad and LoopLM. These are concrete numbers from the experimental results.
Specific accuracy gains are provided in paragraph 42: 18.1% on AIME2025, 13.0% on AIME2026, and 5.4% on GPQA-Diamond compared to TextGrad and LoopLM. These are concrete numbers from the experimental results.
已证实 (75%)
In Mixture-style, RecursiveMAS achieves an average improvement of 6.2% over the strongest domain specialist on each benchmark.
The 6.2% improvement in Mixture-style is stated in paragraph 44. Detailed results are referenced in Appendix D.1 (Tables 7, 6, 8). The claim is supported but detailed tables are in appendix.
The 6.2% improvement in Mixture-style is stated in paragraph 44. Detailed results are referenced in Appendix D.1 (Tables 7, 6, 8). The claim is supported but detailed tables are in appendix.
已证实 (75%)
In Deliberation-style, RecursiveMAS improves the original tool-calling agent by 4.8%.
The 4.8% improvement in Deliberation-style is stated in paragraph 44. Detailed results are in Appendix D.1. The claim is supported with specific numbers.
The 4.8% improvement in Deliberation-style is stated in paragraph 44. Detailed results are in Appendix D.1. The claim is supported with specific numbers.
已证实 (75%)
In Distillation-style, RecursiveMAS improves the learner by 8.0% while retaining 1.5× end-to-end speed advantage over the expert.
The 8.0% improvement and 1.5× speed advantage in Distillation-style are stated in paragraph 44. Detailed results are in Appendix D.1. Both performance and efficiency claims are supported.
The 8.0% improvement and 1.5× speed advantage in Distillation-style are stated in paragraph 44. Detailed results are in Appendix D.1. Both performance and efficiency claims are supported.
已证实 (80%)
At recursion round r = 1, RecursiveMAS already achieves a 1.2× speedup on average, and this advantage grows to 1.9× and 2.4× at larger recursion rounds of r = 2/3.
Specific speedup numbers are provided in paragraph 45 with supporting Figure 5: 1.2× at r=1, 1.9× at r=2, and 2.4× at r=3. The trend of increasing advantage with recursion depth is clearly demonstrated.
Specific speedup numbers are provided in paragraph 45 with supporting Figure 5: 1.2× at r=1, 1.9× at r=2, and 2.4× at r=3. The trend of increasing advantage with recursion depth is clearly demonstrated.
已证实 (80%)
RecursiveMAS reduces the token usage by 34.6% for the first recursion round, and the reduction scales to 75.6% at r = 3.
Specific token reduction numbers are provided in paragraph 46-47 with supporting Figure 6: 34.6% reduction at r=1, scaling to 75.6% at r=3. The explanation for the reduction (avoiding intermediate text decoding) is provided.
Specific token reduction numbers are provided in paragraph 46-47 with supporting Figure 6: 34.6% reduction at r=1, scaling to 75.6% at r=3. The explanation for the reduction (avoiding intermediate text decoding) is provided.
... 共 31 个论点
可复现性评估
较低可复现性 (0%)
缺失的复现细节
- 学习率（Learning rate）- AdamW优化器的学习率未指定
- 训练轮数/步数（Number of epochs or training steps）- 完全未提及
- 随机种子（Random seeds）- 未提及，无法复现随机性
- 具体模型版本和规模（Specific model versions/sizes）- 表1被引用但未在提供文本中展示，只提及模型家族名称
- LoRA微调超参数（LoRA fine-tuning hyperparameters）- 提及使用LoRA但未提供具体配置（rank、alpha等）
- 递归深度（Recursion depth）- 提及作为匹配参数但未指定具体值
- 权重衰减（Weight decay）- AdamW优化器的权重衰减参数未指定
- 预热步数/学习率调度（Warmup steps/learning rate schedule）- 未提及
- 训练数据划分（Training/validation splits）- 未说明数据如何划分
- 提示模板/输入格式（Prompt templates/input formats）- 未提供具体的提示设计
局限性（作者自述）
论文中未明确列出局限性。
本分析由 PDF 阅读助手 自动生成，仅供参考，不构成学术评审意见。验证结论和可复现性评估基于论文文本自动分析，可能存在偏差。原始论文请参阅 arXiv。
分析时间：2026-04-30T01:23:24+00:00 · 数据来源：Paper Collector