Paper deep dive
DeAR: Decentralized Agentic Reasoning via Capability Grounding and Collaborative Thought Navigation
Xing Wei, Changmeng Zheng, XiaoYong Wei, Xiufen Ye, Qing Li
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 8/19/2026, 4:39:30 AM
Summary
The paper introduces DeAR (Decentralized Agentic Reasoning), a framework that replaces centralized multi-agent protocols with autonomous peer-to-peer collaboration. DeAR utilizes three core mechanisms: decentralized capability grounding (dynamic agent specialization based on benchmark scores), thought map navigation (local graph traversal using a collaboration propensity matrix), and topology update (adaptive error correction via progressive backtracking). Evaluations on multimodal and text-based QA benchmarks demonstrate that DeAR outperforms centralized baselines like AutoGen and single large models by enhancing accuracy in knowledge-intensive reasoning tasks.
Entities (9)
Relation Signals (8)
DEAR → evaluatedon → MathVista
confidence 95% · We evaluate our method on four widely used multimodal reasoning datasets: ... MathVista...
DEAR → usesmechanism → Thought Map Navigation
confidence 95% · DeAR is built on three mechanisms: (2) thought map navigation...
DEAR → usesmechanism → Topology Update
confidence 95% · DeAR is built on three mechanisms: (3) topology update...
DEAR → usesmechanism → Decentralized Capability Grounding
confidence 95% · DeAR is built on three mechanisms: (1) decentralized capability grounding...
Collaboration Propensity Matrix → enables → Thought Map Navigation
confidence 90% · Based on the collaboration matrix C, agents construct a collaboration path... The navigation begins with a randomly selected seed agent...
DEAR → outperforms → Qwen2-VL-72B
confidence 90% · the score of DeAR is 18.07 points higher than Qwen2-VL-72B on MathVista
DEAR → outperforms → AutoGen
confidence 90% · DeAR consistently outperforms established multi agent platforms including AutoGen... across all four benchmarks.
DEAR → usesmodel → Deepseek-VL-7B-Chat
confidence 90% · For multimodal reasoning tasks, we assign a different Multimodal Large Language Model (MLLM) to each agent: DeepSeek-VL-7B-Chat...
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Existing agentic reasoning systems typically rely on centralized protocols. This design introduces routing bottlenecks and static role allocations that often fail when handling complex multimodal queries. We propose DeAR (Decentralized Agentic Reasoning), a framework that shifts from central control to autonomous peer-to-peer collaboration. DeAR is built on three mechanisms: (1) decentralized capability grounding for query-dependent agent specialization, (2) thought map navigation for targeted peer interactions, and (3) topology update for adaptive error correction. Evaluations across 9 diverse multimodal reasoning and text-based QA benchmarks indicate that DeAR consistently outperforms recent baseline methods, validating that decentralized and adaptive collaboration among agents enhances accuracy in knowledge-intensive reasoning tasks. The source code will be available at https://open_upon_acceptance.
Tags
Links
- Source: https://arxiv.org/abs/2608.17282v1
- Canonical: https://arxiv.org/abs/2608.17282v1
Trouble viewing inline? Open PDF directly →
Full Text
47,537 characters extracted from source content.
Expand or collapse full text
DeAR: Decentralized Agentic Reasoning via Capability Grounding and Collaborative Thought Navigation Wei Xing 1 , Zheng Changmeng 2 , Wei Xiaoyong 2 , Ye Xiufen 1 , Li Qing 2 1 College of Intelligent Systems Science and Engineering, Harbin Engineering University 2 Department of Computing, The Hong Kong Polytechnic University m18324618517@163.com, yexiufen@hrbeu.edu.cn, changmeng.zheng@polyu.edu.hk, cs007.wei@polyu.edu.hk, qing-prof.li@polyu.edu.hk Abstract Existing agentic reasoning systems typically rely on central- ized protocols. This design introduces routing bottlenecks and static role allocations that often fail when handling com- plex multimodal queries. We propose DeAR (Decentralized Agentic Reasoning), a framework that shifts from central con- trol to autonomous peer-to-peer collaboration. DeAR is built on three mechanisms: (1) decentralized capability ground- ing for query-dependent agent specialization, (2) thought map navigation for targeted peer interactions, and (3) topology update for adaptive error correction. Evaluations across 9 di- verse multimodal reasoning and text-based QA benchmarks indicate that DeAR consistently outperforms recent base- line methods, validating that decentralized and adaptive col- laboration among agents enhances accuracy in knowledge- intensive reasoning tasks. The source code will be available at https://open_upon_acceptance. Introduction Complex reasoning represents the frontier of Large Lan- guage Model (LLM) applications, demanding capabilities that extend far beyond simple pattern matching. To address this, Agentic Reasoning has emerged as a proven and effec- tive paradigm (Zheng et al. 2023). By decomposing intricate problems into manageable sub-tasks handled by specialized agents, agentic workflows have achieved remarkable profi- ciency in multi-step problem solving (Zheng et al. 2024). Despite these successes, current agentic reasoning sys- tems (Zhou et al. 2025) frequently encounter significant stumbling blocks, particularly when handling complex multi- modal queries. Consider the task in Figure 1, which asks how many people in a specific image were born after World War I. A standard single agent often fails here because it cannot simultaneously manage fine-grained visual recognition and historical knowledge retrieval. Furthermore, in a centralized multi-agent framework, a planner rigidly assigns visual ex- traction to one agent and factual querying to another. If an agent misidentifies a person or hallucinates a birth year, the central "Judge" node becomes a severe bottleneck. It blindly aggregates these flawed intermediate results without peer verification, resulting in an incorrect final answer. We argue that these failures stem from structural flaws within the Centralized Collaboration Protocol rather than deficient model capabilities. In these topologies, a central co- ordinator monopolizes routing and aggregation (Wang et al. (b) CentralizedMultiagent Framework Agent_1 Elizabeth Warren June 22, 1949 Agent_2 Agent_3 ElizabethWarrenwasbornin 1949(Agent_1). BarackObamawas bornin1961(Agent_2),Cordray wasbornin1941(Agent_3).” Judge Agents 2 people Centralized methods depend heavily on judge agent LLM Single Agent struggles to obtain knowledge domains. (a) Single Agent 1 person Obama,1961 (c) Decentralized Multiagent Framework Agents Read Guidelines and Corpus Agent(a) Agent(b) Agent(d) Agent(c) Evaluation of agents capability Thought Map Navigation Construction Warren, 1949 Warren,1949 Cordray,1969 Cordray,1959, Warren,1949, Obama,1961 3 people Howmanypeopleintheimagewere bornaftertheendofWorldWarII? Judge_Agent Howmanypeopleinthe imagewerebornafterthe endofWorldWarII? Barack Obama August 4, 1961 Richard Cordray May 3, 1941 Howmanypeopleinthe imagewerebornafterthe endofWorldWarII? Figure 1: Framework comparison on a multimodal query ("How many people in the image were born after World War I?"). (a) Single agent fails due to restricted knowledge domains. (b) Centralized framework fails as the judge bottle- neck propagates incorrect inputs. (c) DeAR (Ours) succeeds by grounding capabilities and navigating an adaptive thought map to collaboratively verify facts and yield the correct an- swer. 2024b), inducing severe information loss and strict single- point bottlenecks (Owens et al. 2025). Consequently, this rigid role allocation (Jia et al. 2024) fails to adapt to fluid, query-dependent reasoning, forcing dynamic problems into static workflows that misroute tasks and propagate unverified errors downstream. In the real world, complex problem-solving follows a dif- ferent topology. It is inherently decentralized. Interdisci- plinary experts do not wait for a single omniscient manager to dictate every micro-interaction; they engage in peer-to- peer discourse, negotiate boundaries, and self-organize to bridge knowledge gaps. However, this decentralized paradigm remains largely un- explored in the context of agentic reasoning.To bridge this gap, we propose DeAR (Decentralized Agentic Reasoning), arXiv:2608.17282v1 [cs.AI] 18 Aug 2026 a framework designed to bypass the bottlenecks of central- ization by enabling agents to reason and collaborate au- tonomously. The evolving reasoning context is maintained not by a hidden system layer, but as a dynamic message payload passed directly through peer-to-peer token streams. Realizing robust decentralized reasoning requires address- ing three challenges, which we tackle by restructuring role allocation, collaboration mechanisms, and reasoning paths: • Dynamic capability vs. Static role: Centralized systems often rely on pre-defined roles (Fang et al. 2025), allowing the coordinator to route requests by labels. But role labels are a weak proxy for real competence and are precisely what causes coordinators to hallucinate capabilities in open-ended scenarios. We address this by replacing static labels with a dynamic capability grounding mechanism. Agents actively assess their own suitability for a spe- cific query fragment based on verifiable linguistic bench- marks, ensuring that role allocation is query-dependent and driven by actual competence rather than rigid titles. • Local graph navigation vs. Central routing: In central- ized pipelines (Wu et al. 2025a), the coordinator acts as a global router that dictates a linear execution path. To eliminate this central bottleneck while maintaining selec- tive collaboration, we propose a collaboration propensity matrix, which intuitively functions as a dynamic adja- cency matrix defining the edges of our reasoning graph. Analogous to sociolinguistic alignment, this matrix es- tablishes a decision boundary. Instead of a central judge computing a global sequential route, the currently active agent performs a local graph traversal, independently se- lecting its optimal downstream peer based on the current context. • Progressive reasoning vs. Restarting: In centralized systems, if a reasoning chain breaks, the “Judge” often has to discard the entire trajectory and restart, which is highly inefficient. We propose that decentralized reason- ing should be viewed as collaborative navigation over an agentic thought map. Rather than fragile linear chains, our agents maintain a shared topological map of the rea- soning space. This allows for progressive reasoning: if a path proves invalid, agents do not retreat to the starting line. Instead, they efficiently pivot from the current node, utilizing the thought map to explore alternative effective chains without losing the progress made thus far. More details are provided in Section . Related work Recent studies show that multi-agent systems improve the robustness and interpretability of Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) by coordinating agents with complementary capabilities. These frameworks facilitate complex task decomposition, in- termediate state sharing, and dynamic workflow management (Hong et al. 2023). Consequently, multi-agent collaboration has been applied successfully across diverse domains, in- cluding mathematical reasoning (Xie et al. 2024) and visual question answering, where iterative planning and reversible reasoning enhance overall performance (Xinjie et al. 2025). To address complex multi-step reasoning tasks, recent frameworks explore distributing cognitive loads across spe- cialized agents (Jiang et al. 2025a; Liu et al. 2025). While debate-based approaches(Zheng et al. 2024; Liang, Wei, and Zheng 2026; Liang et al. 2026) and linear reasoning paths allow for the refinement of intermediate states (Hu et al. 2025), they heavily rely on rigid, centralized coordination protocols. For instance, AutoGen (Wu et al. 2024) depends on static topologies governed by central managers, Agent- Verse (Chen et al. 2024a) dictates synchronous pipelines under central evaluators, and DyLAN (Liu et al. 2024b) requires a global ranker for layer-wise filtering. Even dy- namic routing frameworks like AgentNet (Yang et al. 2025b) lack query-dependent capability grounding and adaptive er- ror correction. This persistent centralized bottleneck limits adaptability, incurs redundant computational overhead, and leaves systems vulnerable to hallucinated competencies and compounding reasoning errors, highlighting the critical need for a decentralized collaborative architecture. DeAR Framework In this section, we present Decentralized Agentic Reasoning framework in which agents autonomously decide when and with whom to collaborate during question answering. The execution of DeAR proceeds in three phases, corresponding to the three challenges identified in the introduction. Overview The core of DeAR is to navigate an Agentic Thought Map, where nodes represent reasoning states and edges represent collaboration probabilities. This navigation is governed by a dynamic Collaboration Propensity Matrix C ∈R N×N , which encodes the likelihood of beneficial interaction be- tween agent pairs based on their grounded capabilities. As agents traverse the map, they generate a sequence of knowl- edge states K. If a path leads to a dead end (i.e., fails to generate a valid answer y), the system employs a progressive refinement strategy, updating the matrix C to prune invalid edges and steer the navigation toward effective reasoning chains. The systemS is defined as: S = (Q,A,C,K).(1) As illustrated in Figure 2, the algorithm consists of three parts. Decentralized Capability Grounding To address the limitations of static role pre-definition, DeAR implements Decentralized Capability Grounding. At initial- ization, agents are not assigned rigid labels (e.g., “Math Agent”). Instead, each agent a i grounds its identity in verifi- able linguistic benchmarks derived from its underlying offi- cial technical reports or model cards, denoted asP i . Rather than relying on externally assigned task labels or groundless self-assessments, we characterize each agent’s capability along a set of fundamental cognitive dimensions commonly required in question answering. Specifically, each agent maintains an intrinsic capability profile: τ i = τ cr i ,τ ea i ,τ nc i ,τ fr i ,τ cm i ,(2) Agent (b) 휏 푐 휏 푒 휏 푛푐 휏 푓푐 휏 푐 Agent Prior Deepseek Qwen Gemini Llama FinalAnswer풚:Thetwo peopleinthepictureareDe GaulleandChurchill. Theage differencebetweenthemis 16 years. Agent (a) 휏 푐 휏 푒 휏 푛푐 휏 푓푐 휏 푐 Agent (d) 휏 푐 휏 푒 휏 푛푐 휏 푓푐 휏 푐 푲 ퟏ 푲 ퟐ Agent (c) 휏 푐 휏 푒 휏 푛푐 휏 푓푐 휏 푐 푲 ퟑ 푲 ퟒ Decentralized Capability Grounding Collaboration Matrix 푪∈푵×푵 failure Success Thought Map Navigation Construction Topology Update Agent푎 1 Agent푎 4 Agent푎 3 Agent푎 2 푐 4,2 푐 푖,푗 =α⋅푐 푖,푗 Collaborative Navigation with Progressive Refinement Ω (푄,퐾 푖 ) failure Success Question푸: What is the agegap between these two peoplein image? 푎 푖+1 =푎 푢 푖 (푎 푖 ) Agent푎 1 Agent푎 4 Agent푎 3 Agent푎 2 푐 4,2 Agent푎 1 Agent푎 4 Agent푎 3 Agent푎 2 푐 4,2 Figure 2: An overview of the Decentralized Agentic Reasoning framework. whereτ cr i ,τ ea i ,τ nc i ,τ cm i ,andτ fr i denote the agent’s proficiency in commonsense reasoning, explanatory ability, numerical computation, natural language comprehension ,and factual reliability, respectively. The detailed prompt templates and alignment protocols employed for this calibration mechanism are provided in Appendix. Crucially, these continuous values do not originate from arbitrary text reading. They are based on explicit mappings of quantitative benchmark data extracted from P i . Let s d i represent the raw accuracy score of agent a i on a standard dataset corresponding to dimensiond. For instance,s nc i is de- rived from the GSM8K benchmark, s cr i from MMLU, and s fr i from TruthfulQA. To resolve the inconsistency of reporting standards across different foundational models, we apply a calibration mechanism during the system initialization phase. The raw scores are projected into the [0, 1] interval and nor- malized across all N agents using a Softmax operation: τ d i = exp(s d i ) P N k=1 exp(s d k ) , d∈cr, ea, nc, fr, cm. (3) This mapping mechanism ensures that the capability values τ i of all agents are rigorously quantified and strictly compa- rable on the same scale. Upon receiving a query q, agent a i dynamically contex- tualizes its profile based on the query requirements and its technical report P i , producing a Query-Dependent Capa- bility State: τ i (q) = Φ i (q,P i ,τ i ),(4) where Φ i (·) denotes the agent’s internal reasoning operator, implemented via prompting, which evaluates the task context against the agent’s calibrated benchmark profile. Finally, the knowledge or intermediate reasoning gener- ated by agent a i is conditioned on the query and this capa- bility state: K i = Θ i (q |τ i (q)).(5) Here, Θ i (·) denotes the LLM’s generative reasoning process driven by its parametric knowledge, yielding the intermediate results K i . By evaluating peer benchmark reports, agent a i quantitatively estimates their strengths, forming the basis to compute the collaboration propensity c i,j and construct the decentralized thought map. Self-Organized Collaboration To replace centralized routing with local consensus, we intro- duce a mechanism for Self-Organized Collaboration. Each agent a i autonomously evaluates the potential utility of peer- ing with every other agent a j under the current query con- text.To initiate the decentralized reasoning trajectory, the first active agent is selected uniformly at random from the agent pool. This evaluation forms a Collaboration Propensity Vec- tor c i : c i = [c i,1 , c i,2 , ..., c i,N ],(6) where c i,j ∈ [0, 1] is the collaboration preference towards agent a j . Given a query mapping to a cognitive dimension d, c i,j is calculated via a temperature-scaled Softmax over the calibrated proficiency τ d j defined in Equation 3: c i,j = exp(τ d j /T ) P N k=1,k̸=i exp(τ d k /T ) , ∀j ̸= i(7) where T > 0 controls routing sharpness. Since τ d j ∈ [0, 1] is bounded, this formulation inherently prevents exponen- tial overflow, guaranteeing absolute numerical stability. By definition, self-collaboration is disallowed: c i,i = 0, ∀i∈1, 2,...,N.(8) Aggregating the collaboration preference vectors from all agents yields a collaboration matrixC: C = 0 c 1,2 · c 1,N c 2,1 0 · c 2,N . . . . . . . . . . . . c N,1 c N,2 ·0 (9) This matrix encodes pairwise and directed collaboration preferences among agents, reflecting both agent hetero- geneity and task-specific requirements. Unlike static rout- ing strategies or centralized coordination mechanisms, the Collaboration Propensity Coefficient Success Failure (a)Thought Map Navigation Construction Failure Success (c)Final Thought Map Navigation Collaboration Propensity Coefficient Failure 훂Decay factor (b)Topology Update Collaboration Propensity Coefficient Figure 3: Thought Map Navigation Construction among Agents collaboration structure here is dynamically shaped by each agent’s local assessment under the current query. Although the base ranking of c i,j reflects target capabilities, it avoids global ranking degeneration by activating agents on an as- needed basis. The cumulative context K i from operator Θ i imbues the payload with source-specificity, achieving context-dependent collaboration. Thought Map Navigation construction Based on the col- laboration matrixC, agents construct a collaboration path in a sequential manner. The objective of this process is to deter- mine an ordered sequence of agents such that the knowledge generated by each selected agent is incrementally propagated and refined through interaction. The navigation begins with a randomly selected seed agent (or the most relevant initial agent). At stepi, the current agent a i acts as a local navigator, selecting the next node (agent a j ) that maximizes the collaboration propensity. The incremental utility of transitioning to an unvisited agent a j is: u i (a j ) = c i,j , a j /∈V i , 0, a j ∈V i . (10) The next agent a i+1 is chosen by a i+1 = arg max a j ∈A u i (a j )(11) The selected agent then generates its knowledge condi- tioned on the input query and the accumulated context formed by the union of knowledge generated by previously selected agents, i.e., S i k=1 K k . This creates a chain of thought that exploits the strongest collaborative edges in the map. Through sequential selection and knowledge propagation, agents autonomously construct a thought map that progres- sively integrates complementary reasoning capabilities. In- dependent of any centralized judge or predefined topology, the terminal agent generates the final answer y directly from the fully accumulated knowledgeK N .Dynamic Termination. This decentralized trajectory terminates dynamically based on real-time capability assessments rather than a fixed step limit. Navigation halts if the evaluated propensities of all un- visited peers (outside V i ) fall below a minimal operational threshold, or if all agents have been exhausted. Upon halting, the active agent serves as the terminal node, outputting its accumulated context as the final answer. Collaborative Navigation with Progressive Refinement Upon reaching the terminal state, the system attempts to generate a final answer y. If the accumulated context K N is insufficient (a “dead end” on the map), we do not simply discard the effort. Instead, we perform a Topology Update Ψ(·) to refine the thought map. y = Ω (q,K N ),(12) where Ω(·) denotes the answer generation function of the ter- minal agent a N conditioned on the full accumulated context. Specifically,DeAR leverages the terminal agent rather than an external judge to verify the final contextK N . If the final context fails to match the initial query and its requirements, a refusal is triggered. To prune the target edge setE fail without over-penalizing valid upstream transitions, DeAR employs a localized, progressive backtracking strategy instead of full- path mitigation. Initially, E fail contains only the immediate terminal edge, prompting alternative routing at the current depth. If local choices are exhausted, E fail expands step-by- step to preceding edges in the reasoning path. Finally, these active edges are penalized in the Collaboration Propensity Matrix: C ← Ψ(C),(13) where the transformation applies a decay factor α (e.g., 0.5) to the failed connections: c i,j = α· c i,j , (a i → a j )∈E fail , c i,j ,otherwise. (14) This update prunes failing edges, backtracking to preceding layers step-by-step if local routes fail. The agents then re- navigate the adjusted topology. This process of exploration, sequential backtracking, and refinement allows DeAR to pro- gressively converge on a valid reasoning path, bypassing the single-point vulnerabilities typical of centralized systems. Experimental Setup Datasets We evaluate our method on four widely used multimodal rea- soning datasets: MMMU (Yue et al. 2024), MathVista (Lu et al. 2024b), ChartQA (Hegde, Fazli, and Seifi 2025), and ScienceQA (Lu et al. 2022). We adopt Accuracy (ACC) as the primary evaluation metric and conduct experiments on the full test splits for all benchmarks. Specifically, beyond the overall accuracy for MMMU and ChartQA, we report detailed fine-grained metrics for MathVista across different question types (e.g., FQA, GPS, MWP) and reasoning skills (e.g., ALG, ARI, GEO). For ScienceQA, we report the accu- racy across various subjects (NAT, SOC, LAN), modalities (TXT, IMG, NO), and grade levels (G1-6, G7-12), alongside the average score. Furthermore, to demonstrate the generalizability of DeAR, we additionally conduct experiments on 5 text- based QA benchmarks, encompassing both single-hop (e.g., NQ(Kwiatkowski et al. 2019), TriviaQA(Joshi et al. 2017), PopQA(Mallen et al. 2023)) and multi-hop (e.g., WikiMul- tihopQA(Ho et al. 2020), and HotpotQA(Yang et al. 2018)) datasets. For these pure text tasks, we employ Exact Match (EM) and F1 as evaluation metrics, with 1,000 examples randomly sampled for the multi-hop evaluations. Implementation Details In our implementation, the thought map uses a dynamic multi-agent design where up to four agents can participate de- pending on the query’s complexity. For multimodal reason- ing tasks, we assign a different Multimodal Large Language Model (MLLM) to each agent: DeepSeek-VL-7B-Chat (Lu et al. 2024a), LLaVA-1.5-7B (Liu et al. 2023), Qwen2-VL- 7B-Instruct (Wang et al. 2024a), and MiniCPM-V-2.6 (Yao et al. 2024). Using different models provides diverse visual reasoning perspectives. For text-based QA evaluations, the agents are powered by four distinct LLMs (Qwen3-8B (Yang et al. 2025a), Gemma-3-1B-IT (Team et al. 2025), Llama- 3.2-3B (Dubey et al. 2024), and DeepSeek-LLM-7B-Chat (Liu et al. 2024a)). We integrate these text models with the FlashRAG toolkit (Jin et al. 2025), using E5-base-v2 as the dense retriever to fetch the top k = 10 passages from a shared Wikipedia corpus. We keep all generation and system hyperparameters consistent across the multimodal and QA experiments. Finally, all models are evaluated in a zero-shot setting without any task-specific fine-tuning. Experimental Results Comparison with Baselines Evaluating DeAR against large single models shows that im- provements come from the structured architecture instead of parameter scaling. Despite relying on a cooperative system of four base models, DeAR maintains a smaller total param- eter count than 72B or 78B models while showing higher accuracy. Specifically, the score of DeAR is 18.07 points higher than Qwen2-VL-72B on MathVista (59.41 vs. 41.34), as shown in Table 1. Compared to recent chain of thought and multi agent frameworks such as MAD-Vote(Liang et al. 2024), MUG(Liang, Wei, and Zheng 2025), C2R(Jang et al. 2025), Cache of Thought(Wu et al. 2025b), Corvid(Jiang et al. 2025b), and Insight-V(Dong et al. 2025), DeAR reports the highest overall scores on MathVista (59.41), ChartQA (89.43), and ScienceQA (62.45) in Table 1.It also consis- tently outperforms established multi agent platforms includ- ing AutoGen(Wu et al. 2024), AgentVerse(Chen et al. 2024a), and DyLAN(Liu et al. 2024b) across all four benchmarks. For instance, DeAR surpasses these three baselines on ChartQA by over 2.3 and achieves a higher score on MMMU (55.58), showing the advantage of a decentralized topology over con- ventional centralized routing. Evaluations on five text QA benchmarks show the per- formance of DeAR beyond multimodal tasks, with detailed metrics provided in Table 2. The framework ranks first on NQ and obtains the highest F1 score on TriviaQA for single hop tasks. For multi hop queries, it obtains the highest score on HotpotQA and maintains stable precision and complete- ness on 2Wiki (Table 2), showing that decentralized thought navigation applies to pure NLP domains. Table 1 decouples architectural gains from ensemble ef- fects using identical backbones. DeAR (Fixed Role), serving as the centralized routing baseline, drops sharply (-5.76 on MMMU). DeAR (w/o Topology Update), acting as the en- semble control without backtracking, also degrades (-2.37 on MMMU). These margins mathematically prove our superi- ority stems from decentralized navigation rather than vanilla model aggregation. Analysis of the Thought Map Navigation This section evaluates the effectiveness of the proposed thought map navigation by comparing it against a variant without (w/o) the thought map navigation. To provide deeper insights beyond aggregate metrics, we conduct a fine-grained analysis of the performance across diverse question types and reasoning skills within the MathVista dataset. The re- sults demonstrate that dynamic agent routing consistently outperforms static setups across all sub-categories, confirm- ing that adaptive collaboration is essential for complex visual reasoning. Overall, the thought map navigation increases the com- prehensive MathVista accuracy from 51.68 to 59.41. The detailed breakdown reveals that the most substantial gains occur in tasks demanding intricate multi-step logic and spe- cialized visual processing. For example, accuracy on Sta- tistical Reasoning (STA) improves significantly from 62.9 to 80.9. Similarly, Logical Reasoning (LOG) sees a major boost from 19.6 to 31.0, and Figure Question Answering (FQA) rises from 69.1 to 80.3. These consistent enhancements val- idate that allowing agents to autonomously navigate the rea- soning space is far more effective than forcing a predefined execution order. Analysis of Topology Update This section evaluates the proposed topology update on multimodal benchmarks, including MMMU, MathVista, ChartQA, and ScienceQA. By correcting intermediate er- rors, the topology update consistently improves accu- racy across datasets with varying visual complexities. On MMMU, the score increases from 53.21 to 55.58, mitigat- ing errors in college level academic tasks, while MathVista accuracy rises from 55.65 to 59.41 in intricate mathematical settings. Similar gains are observed on ChartQA (88.01 to 89.43) and ScienceQA (59.92 to 62.45), reflecting a reduced accumulation of multiple step reasoning errors. MethodMMMU MathVista ChartQA ScienceQA FQA GPS MWP TQA VQA ALLNAT SOC LAN ALL Baseline Models (Single Agent) LLaMA-2-13B (Liu et al. 2023)-26.8 29.3 16.1 32.3 26.3 26.10-44.1 41.2 43.9 43.08 InternVL2.5-26B (Chen et al. 2024b)51.8038.0 35.0 30.0 40.0 37.0 36.03----- MiniCPM-V-2.6 (Yao et al. 2024) 45.1151.7 27.4 39.8 42.5 34.7 39.89-59.5 57.0 59.6 58.70 Qwen2-VL-72B (Wang et al. 2024a)46.2155.9 34.7 29.7 58.8 42.4 41.3488.30---- NVLM-H 1.0 78B (Dai et al. 2024)53.0865.0 48.0 45.0 68.0 53.0 55.91----- Baseline Models (Multi-Agent Frameworks) MAD-Vote (EMNLP 2024) (Liang et al. 2024)44.7049.0 38.0 35.0 51.0 42.0 43.1281.0255.0 52.1 54.0 53.70 MUG (AAAI 2025) (Liang, Wei, and Zheng 2025)50.3055.0 40.0 39.0 56.0 46.0 47.32-56.5 53.8 55.5 55.27 C2R (EMNLP 2025) (Jang et al. 2025)51.9260.1 38.2 45.4 62.0 49.3 51.0185.08---- Cache-of-Thought (EMNLP 2025) (Wu et al. 2025b)37.9254.2 35.4 30.0 58.3 44.6 44.51-59.0 56.5 59.8 58.44 Corvid (ICCV 2025) (Jiang et al. 2025b)52.2065.0 40.8 35.7 70.6 50.0 52.43----- Insight-V (CVPR 2025) (Dong et al. 2025) 53.7078.0 48.5 42.1 70.3 53.1 58.4083.0063.5 59.2 63.0 61.93 AutoGen (Wu et al. 2024)54.3177.6 49.2 42.7 68.7 54.5 59.0186.7962.7 58.7 62.1 60.24 AgentVerse (Chen et al. 2024a) 55.1076.4 47.1 41.6 69.1 54.4 58.7987.1261.9 58.4 62.9 61.45 DyLAN (Liu et al. 2024b)54.3977.2 48.9 42.9 67.1 53.9 59.1686.4662.4 59.1 61.3 60.52 Our Proposed Method DeAR (Fixed Role)49.8268.1 41.2 38.9 60.1 49.7 51.6085.7261.0 57.5 60.6 59.70 DeAR (w/o Thought Map)47.2068.2 41.3 39.0 60.2 49.7 51.6883.1255.5 51.2 55.5 54.07 DeAR (w/o Topology Update) 53.2175.0 45.0 41.0 70.5 46.7 55.6588.0161.2 57.5 61.0 59.92 DeAR (Ours) 55.5880.3 49.9 42.7 84.5 60.9 59.4189.4364.8 60.3 62.2 62.45 Table 1: Zero-shot evaluation on four multimodal benchmarks (fine-grained metrics reported for MathVista and ScienceQA). ’-’ denotes unreported data. Agent counts: Cache-of-Thought (2); MUG, C2R (3); MAD-Vote, Corvid, Insight-V, AutoGen, AgentVerse, DyLAN, ReConcile, and DeAR (4). Best results in bold. MethodNQTriviaQAPopQA2WikiHotpotQA EM∆EMF1∆F1EM∆EMF1∆F1EM∆EMF1∆F1EM∆EMF1∆F1EM∆EMF1∆F1 Single Agent Framework Vanilla Gen17.40 +29.79 26.27 +27.8656.60 +4.87 65.50 +7.5419.20 +20.54 23.33 +23.878.60 +42.53 16.50 +44.7416.80 +31.51 23.88 +29.86 Multi-Agent Frameworks MAD(Du et al. 2023)(ICML 2024)21.80 +25.39 33.11 +21.0256.40 +5.07 66.39 +6.6521.40 +18.34 28.28 +18.9218.20 +32.93 25.13 +36.1123.00 +25.31 32.79 +20.95 IRCoT(Trivedi et al. 2023)(ACL 2023)28.60 +18.59 37.36 +16.7747.20 +14.27 54.56 +18.4827.00 +12.74 33.02 +14.1832.80 +18.33 31.19 +30.0525.20 +23.11 34.40 +19.34 Iter-RetGen(Shao et al. 2023)(Arxiv 2023)40.80 +6.3952.31 +1.8263.00 -1.5372.23 +0.8139.60 +0.14 46.41 +0.7915.00 +36.13 24.75 +36.4927.80 +20.51 38.93 +14.81 FLARE(Jiang et al. 2023)(EMNLP 2023)19.40 +27.79 27.68 +26.4553.60 +7.87 63.05 +9.9921.60 +18.14 24.35 +22.859.20 +41.93 20.13 +41.1116.60 +31.71 23.74 +30.00 Self-RAG(Asai et al. 2024)(Arxiv 2024)44.00 +3.19 52.20 +1.9346.40 +15.07 58.37 +14.6722.00 +17.74 34.38 +12.8213.00 +38.13 26.63 +34.6114.80 +33.51 28.81 +24.93 DRAG(Hu et al. 2025)(ACL 2025)36.80 +10.39 50.38 +3.7560.80 +0.67 69.93 +3.1138.60 +1.1446.50 +0.7028.80 +22.33 36.97 +24.2730.80 +17.51 41.74 +12.00 DeAR (Ours)47.19–54.13–61.47–73.04–39.74–47.20–51.13–61.24–48.31–53.74– Table 2: Overall evaluation results of the proposed DeAR framework and other baselines on five text-only QA benchmarks. ∆EM and ∆F1 denote the performance gap between DeAR and each baseline (∆ = DeAR score−baseline score). Darker gray cells mark the best performance, and lighter gray cells mark the second-best performance among all methods. 0.10.20.30.40.50.60.70.80.9 Decay Factor 48 50 52 54 56 58 60 62 Performance Sensitivity Analysis of Decay Factor MMMU MathVista 2WikiMultihopQA Figure 4: Analysis of the topology update’s decay factor α. Effect of Decay Factor α Figure 4 investigates the topol- ogy update’s decay factor α ∈ [0.1, 0.9], which controls the penalty for failed paths. Small values (α = 0.1) cause overcorrection and unstable thought map paths, dropping the 2Wiki F1 score to 53.91. Conversely, large values (α = 0.9) insufficiently suppress ineffective paths, causing repeated in- valid computations and a 2Wiki F1 of 55.02. DeAR performs optimally in the [0.5, 0.8]. Specifically, a moderate α = 0.5 achieves peak performance on MMMU (55.58) and Math- Vista (59.41) by effectively selecting new paths without dis- carding previously successful interactions. Dynamic Agent Activation and Usage Frequency DeAR constructs reasoning paths dynamically based on query requirements and agent confidence. The framework does not mandate the participation of all N = 4 agents; uns- elected agents remain inactive if a smaller subset can resolve the query, which avoids unnecessary compute consumption. Table 3 presents the activation frequency of different agent Q1:Usingthemapprovided,whichistherightanswerto identifyeachpartoftheMongolEmpire?(Mathvistadataset) Q2:Howmanyzerosdoesthis functionhave?(MMMUQAdataset) (A).A .GreatKhan. B. Il-Khanate. C. ChagataiKhanate. D. GoldenHorde (B).A. GoldenHorde. B. ChagataiKhanate. C Il-Khanate. D. GreatKhan (C).A.GoldenHorde. B. Il-Khanate. C ChagataiKhanate. D. GreatKhan (D).A. ChagataiKhanate. B. GoldenHorde. C. GreatKhan. D. Khanate (A).One (B).Two (C).Three (D).Four DeAR Algorithm Agent_1 I knowthatD isGreatKhan.Agent_2 ...choose tocooperatewithAgent_2. Agent_2 I know...distinguished."Agent_3 hasstrong explanatoryability, soI chose...tohim. Agent_3 I observed...Bcoastline,andtheChagatai Khanateis ...andB representstheIlkhanate. Agent_4 I calculatedtheterritorialratioChagataiand theGreatKhan. Agent_4 Agent_1 Agent_2 Agent_3 I choose C Agent_1 Region A is top-left, ..., Region Cis central, Region Dis on the right. Agent_3 TheGoldenHorde...inthe...,the Southwest,...andGreatKhanin theEast. Judge_Agent Agent_2 OnlylettersA,B,C,Darepresent intheimage Agent_1 Agent_2 Agent_3 Judge_Agent I choose A Agent_1 is right,butAgent_2 has somemistakes,Agent_3’sresults hashallucination. Centralized Multiagent Agent_1 Iobservecurve.fourdots... cooperatewithAgent_2. Agent_2 CurvecrossestheX-axis...specifically atthe(1, 0). Agent_3 A0 isdefined...function'svalue evaluatestoy = 0. Agent_4 I set...tolog_2(x)=0. Solvingthisyieldsx = 1. DeAR Algorithm Agent_4 Agent_1 Agent_2 Agent_3 I choose A Agent_1 The fixed role ... 'Visual Extraction'. Agent_2 Equationisy = log2(x) forthenumberof0. Agent_3 I observethatthe bluecurvecrosses theX-axis... Judge_Agent The function has 4 zeros. Agent_1 Agent_2 Agent_3 Judge_Agent I choose B Centralized Multiagent Agent Capability Awareness Commonsense Reasoning Explanatory Ability Numerical Computation Factual Reliability Agent Capability Awareness Commonsense Reasoning Explanatory Ability Numerical Computation Factual Reliability Figure 5: Case Study of the Reasoning with Centralized and Decentralized Multi-Agent Collaboration on Q&A. subset sizes on MathVista. The results show that while com- plex problems require all 4 agents (37.5%), a majority of queries (62.5%) are resolved by fewer agents, confirming the adaptability and resource efficiency of the algorithm. Table 3: Frequency of activation for varying numbers of agents on the MathVista dataset. Number of Agents Activated Frequency (%) 1 Agent14.3 2 Agents21.8 3 Agents26.4 4 Agents37.5 Case Study Figure 5 compares DeAR and a centralized multi-agent base- line on historical map identification (Q1) and chart reasoning (Q2).DeAR resolves visual and logical ambiguities through dynamic, capability-driven hand-offs. In Q1, agents pro- gressively determine map assignments by combining factual knowledge, spatial commonsense, and historical geographic constraints, such as identifying landlocked regions. They fi- nalize the boundaries via numerical verification of area ra- tios. In Q2, rather than being misled by the four plotted dots, the agents translate the mathematical "zero" into a ge- ometric target on the x-axis and perform an algebraic check (log 2 (x) = 0 ⇒ x = 1) to confirm the single root.The cen- tralized framework restricts agents to isolated, static roles like visual extraction or text reading. All intermediate find- ings are forced through a single Judge Agent, creating an in- formation bottleneck. Without peer-to-peer verification, the Judge blindly aggregates fragmented reports. As a result, it hallucinates an incorrect map configuration in Q1 by forcibly merging vague hints. In Q2, the Judge directly misattributes the visual worker’s observation of "four prominent dots" to the mathematical query and incorrectly outputs "Four".These results confirm that decentralized navigation effectively re- solves complex logical dependencies and prevents the aggre- gation errors typical of centralized pipelines. Conclusion We propose DeAR, a decentralized framework replac- ing centralized multi-agent coordination with autonomous, capability-aware collaboration. DeAR integrates three core mechanisms: Decentralized Capability Grounding for query- dependent agent specialization, Thought Map Navigation for selective peer interaction, and Topology Update for adaptive error correction. These components allow agents to progres- sively refine reasoning trajectories without a central judge. Evaluations across 9 diverse multimodal and text-based QA benchmarks show DeAR consistently outperforms recent baselines, validating that decentralized, adaptive collabora- tion among agents enhances performance in complex tasks. References Asai, A.; Wu, Z.; Wang, Y.; Sil, A.; and Hajishirzi, H. 2024. Self-rag: Learning to retrieve, generate, and critique through self-reflection. Chen, W.; Su, Y.; Zuo, J.; Yang, C.; Yuan, C.; Chan, C.-M.; Yu, H.; Lu, Y.; Hung, Y.-H.; Qian, C.; et al. 2024a. Agent- verse: Facilitating multi-agent collaboration and exploring emergent behaviors. In International Conference on Learn- ing Representations, volume 2024, 20094–20136. Chen, Z.; Wu, J.; Wang, W.; Su, W.; Chen, G.; Xing, S.; Zhong, M.; Zhang, Q.; Zhu, X.; Lu, L.; et al. 2024b. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 24185–24198. Dai, W.; Lee, N.; Wang, B.; Yang, Z.; Liu, Z.; Barker, J.; Rintamaki, T.; Shoeybi, M.; Catanzaro, B.; and Ping, W. 2024. NVLM: Open Frontier-Class Multimodal LLMs. arXiv preprint. Dong, Y.; Liu, Z.; Sun, H.-L.; Yang, J.; Hu, W.; Rao, Y.; and Liu, Z. 2025. Insight-v: Exploring long-chain visual reason- ing with multimodal large language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, 9062–9072. Du, Y.; Li, S.; Torralba, A.; Tenenbaum, J. B.; and Mordatch, I. 2023. Improving factuality and reasoning in language mod- els through multiagent debate. In Forty-first International Conference on Machine Learning. Dubey, A.; Jauhri, A.; Pandey, A.; Kadian, A.; Al-Dahle, A.; Letman, A.; Mathur, A.; Schelten, A.; Yang, A.; Fan, A.; et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783. Fang, J.; Peng, Y.; Zhang, X.; Wang, Y.; Yi, X.; Zhang, G.; Xu, Y.; Wu, B.; Liu, S.; Li, Z.; et al. 2025. A comprehensive survey of self-evolving ai agents: A new paradigm bridg- ing foundation models and lifelong agentic systems. arXiv preprint arXiv:2508.07407. Hegde, S.; Fazli, P.; and Seifi, H. 2025. Chartqa-x: Generat- ing explanations for charts. arXiv e-prints, arXiv–2504. Ho, X.; Nguyen, A.-K. D.; Sugawara, S.; and Aizawa, A. 2020. Constructing a multi-hop qa dataset for com- prehensive evaluation of reasoning steps. arXiv preprint arXiv:2011.01060. Hong, S.; Zhuge, M.; Chen, J.; Zheng, X.; Cheng, Y.; Wang, J.; Zhang, C.; Wang, Z.; Yau, S. K. S.; Lin, Z.; et al. 2023. MetaGPT: Meta programming for a multi-agent collabora- tive framework. In The Twelfth International Conference on Learning Representations. Hu, W.; Zhang, W.; Jiang, Y.; Zhang, C. J.; Wei, X.; and Qing, L. 2025. Removal of hallucination on hallucination: Debate-augmented RAG. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 15839–15853. Jang, Y.; Choi, W. S.; Jung, M.; Lee, M.; and Zhang, B.-T. 2025. Confidence-guided Refinement Reasoning for Zero- shot Question Answering. In Proceedings of the 2025 Con- ference on Empirical Methods in Natural Language Process- ing, 6944–6961. Jia, C.; Luo, M.; Dang, Z.; Sun, Q.; Xu, F.; Hu, J.; Xie, T.; and Wu, Z. 2024. Agentstore: Scalable integration of hetero- geneous agents as specialized generalist computer assistant. arXiv preprint arXiv:2410.18603. Jiang, B.; Xie, Y.; Wang, X.; Yuan, Y.; Hao, Z.; Bai, X.; Su, W. J.; Taylor, C. J.; and Mallick, T. 2025a. Towards rationality in language and multimodal agents: a survey. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Lin- guistics: Human Language Technologies (Volume 1: Long Papers), 3656–3675. Jiang, J.; Ma, C.; Song, X.; Zhang, H.; and Luo, J. 2025b. Corvid: Improving multimodal large language models to- wards chain-of-thought reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 3034–3046. Jiang, Z.; Xu, F. F.; Gao, L.; Sun, Z.; Liu, Q.; Dwivedi-Yu, J.; Yang, Y.; Callan, J.; and Neubig, G. 2023. Active retrieval augmented generation. In Proceedings of the 2023 Confer- ence on Empirical Methods in Natural Language Processing, 7969–7992. Jin, J.; Zhu, Y.; Dou, Z.; Dong, G.; Yang, X.; Zhang, C.; Zhao, T.; Yang, Z.; and Wen, J.-R. 2025. Flashrag: A modular toolkit for efficient retrieval-augmented generation research. In Companion Proceedings of the ACM on Web Conference 2025, 737–740. Joshi, M.; Choi, E.; Weld, D. S.; and Zettlemoyer, L. 2017. Triviaqa: A large scale distantly supervised chal- lenge dataset for reading comprehension. arXiv preprint arXiv:1705.03551. Kwiatkowski, T.; Palomaki, J.; Redfield, O.; Collins, M.; Parikh, A.; Alberti, C.; Epstein, D.; Polosukhin, I.; Devlin, J.; Lee, K.; et al. 2019. Natural questions: a benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7: 453–466. Liang, D.; Gong, K.; Cai, Y.; Zheng, C.; and Wei, X.-Y. 2026. Mixture of Debaters: Learn to Debate at Architectural Level in Multi-Agent Reasoning. arXiv preprint arXiv:2606.29425. Liang, D.; Wei, X.-Y.; and Zheng, C. 2025. Multi-agent Undercover Gaming: Hallucination Removal via Counter- factual Test for Multimodal Reasoning. arXiv preprint arXiv:2511.11182. Liang, D.; Wei, X.-Y.; and Zheng, C. 2026. Multi-agent undercover gaming: Hallucination removal through counter- factual test for multimodal reasoning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, 6807–6815. Liang, T.; He, Z.; Jiao, W.; Wang, X.; Wang, Y.; Wang, R.; Yang, Y.; Shi, S.; and Tu, Z. 2024. Encouraging divergent thinking in large language models through multi-agent de- bate. In Proceedings of the 2024 conference on empirical methods in natural language processing, 17889–17904. Liu, A.; Feng, B.; Xue, B.; Wang, B.; Wu, B.; Lu, C.; Zhao, C.; Deng, C.; Zhang, C.; Ruan, C.; et al. 2024a. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437. Liu, H.; Li, C.; Wu, Q.; and Lee, Y. J. 2023. Visual Instruction Tuning. In NeurIPS. Liu, P.; Liu, X.; Yao, R.; Liu, J.; Meng, S.; Wang, D.; and Ma, J. 2025. Hm-rag: Hierarchical multi-agent mul- timodal retrieval augmented generation. arXiv preprint arXiv:2504.12330. Liu, Z.; Zhang, Y.; Li, P.; Liu, Y.; and Yang, D. 2024b. A dy- namic LLM-powered agent network for task-oriented agent collaboration. In First Conference on Language Modeling. Lu, H.; Liu, W.; Zhang, B.; Wang, B.; Dong, K.; Liu, B.; Sun, J.; Ren, T.; Li, Z.; Sun, Y.; Deng, C.; Xu, H.; Xie, Z.; and Ruan, C. 2024a. DeepSeek-VL: Towards Real-World Vision-Language Understanding. arXiv:2403.05525. Lu, P.; Bansal, H.; Xia, T.; Liu, J.; Li, C.; Hajishirzi, H.; Cheng, H.; Chang, K.-W.; Galley, M.; and Gao, J. 2024b. MathVista: Evaluating Mathematical Reasoning of Founda- tion Models in Visual Contexts. In International Conference on Learning Representations (ICLR). Lu, P.; Mishra, S.; Xia, T.; Qiu, L.; Chang, K.-W.; Zhu, S.-C.; Tafjord, O.; Clark, P.; and Kalyan, A. 2022. Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answering. In The 36th Conference on Neural Information Processing Systems (NeurIPS). Mallen, A.; Asai, A.; Zhong, V.; Das, R.; Khashabi, D.; and Hajishirzi, H. 2023. When not to trust language models: Investigating effectiveness of parametric and non-parametric memories. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 9802–9822. Owens, D. M.; Rossi, R.; Kim, S.; Yu, T.; Dernoncourt, F.; Chen, X.; Zhang, R.; Gu, J.; Deilamsalehy, H.; and Lipka, N. 2025. Multi-LLM Debiasing Framework. In Proceedings of the 15th International Conference on Recent Advances in Natural Language Processing-Natural Language Processing in the Generative AI Era, 843–853. Shao, Z.; Gong, Y.; Shen, Y.; Huang, M.; Duan, N.; and Chen, W. 2023. Enhancing retrieval-augmented large language models with iterative retrieval-generation synergy. arXiv preprint arXiv:2305.15294. Team, G.; Kamath, A.; Ferret, J.; Pathak, S.; Vieillard, N.; Merhej, R.; Perrin, S.; Matejovicova, T.; Ramé, A.; Rivière, M.; et al. 2025. Gemma 3 technical report. arXiv preprint arXiv:2503.19786. Trivedi, H.; Balasubramanian, N.; Khot, T.; and Sabharwal, A. 2023. Interleaving retrieval with chain-of-thought reason- ing for knowledge-intensive multi-step questions. In Pro- ceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers), 10014– 10037. Wang, P.; Bai, S.; Tan, S.; Wang, S.; Fan, Z.; Bai, J.; Chen, K.; Liu, X.; Wang, J.; Ge, W.; Fan, Y.; Dang, K.; Du, M.; Ren, X.; Men, R.; Liu, D.; Zhou, C.; Zhou, J.; and Lin, J. 2024a. Qwen2-VL: Enhancing Vision-Language Model’s Perception of the World at Any Resolution. arXiv preprint arXiv:2409.12191. Wang, Q.; Wang, T.; Li, Q.; Liang, J.; and He, B. 2024b. Megaagent: A practical framework for autonomous coopera- tion in large-scale llm agent systems. arXiv e-prints, arXiv– 2408. Wu, F.; Li, Z.; Wei, F.; Li, Y.; Ding, B.; and Gao, J. 2025a. Talk to right specialists: Routing and planning in multi-agent system for question answering. arXiv preprint arXiv:2501.07813. Wu, M.; Jiang, J.; Zheng, H.; Li, M.; Li, Z.; Tian, B.; Chen, B.; Park, Y.; Zhang, M.; Zhai, C.; et al. 2025b. Cache- of-thought: Master-apprentice framework for cost-effective vision language model inference. arXiv e-prints, arXiv– 2502. Wu, Q.; Bansal, G.; Zhang, J.; Wu, Y.; Li, B.; Zhu, E.; Jiang, L.; Zhang, X.; Zhang, S.; Liu, J.; et al. 2024. Autogen: Enabling next-gen LLM applications via multi-agent con- versations. In First conference on language modeling. Xie, W.; Liu, D.; Yan, H.; Wu, W.; and Liu, Z. 2024. Mathlearner: A large language model agent framework for learning to solve mathematical problems. arXiv preprint arXiv:2408.01779. Xinjie, Z.; Gao, F.; Song, X.; Chen, Y.; Yang, R.; Fu, Y.; Wang, Y.; Iwasawa, Y.; Matsuo, Y.; and Li, I. 2025. Reagent: Reversible multi-agent reasoning for knowledge-enhanced multi-hop qa. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, 4067– 4089. Yang, A.; Li, A.; Yang, B.; Zhang, B.; Hui, B.; Zheng, B.; Yu, B.; Gao, C.; Huang, C.; Lv, C.; et al. 2025a. Qwen3 technical report. arXiv preprint arXiv:2505.09388. Yang, Y.; Chai, H.; Shao, S.; Song, Y.; Qi, S.; Rui, R.; and Zhang, W. 2025b. Agentnet: Decentralized evolutionary co- ordination for llm-based multi-agent systems. arXiv preprint arXiv:2504.00587. Yang, Z.; Qi, P.; Zhang, S.; Bengio, Y.; Cohen, W.; Salakhut- dinov, R.; and Manning, C. D. 2018. HotpotQA: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 conference on empirical methods in natural language processing, 2369–2380. Yao, Y.; Yu, T.; Zhang, A.; Wang, C.; Cui, J.; Zhu, H.; Cai, T.; Li, H.; Zhao, W.; He, Z.; et al. 2024. MiniCPM- V: A GPT-4V Level MLLM on Your Phone. arXiv preprint arXiv:2408.01800. Yue, X.; Ni, Y.; Zhang, K.; Zheng, T.; Liu, R.; Zhang, G.; Stevens, S.; Jiang, D.; Ren, W.; Sun, Y.; Wei, C.; Yu, B.; Yuan, R.; Sun, R.; Yin, M.; Zheng, B.; Yang, Z.; Liu, Y.; Huang, W.; Sun, H.; Su, Y.; and Chen, W. 2024. MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI. In Proceedings of CVPR. Zheng, C.; Feng, J.; Cai, Y.; Wei, X.; and Li, Q. 2023. Re- thinking multimodal entity and relation extraction from a translation point of view. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 6810–6824. Zheng, C.; Liang, D.; Zhang, W.; Wei, X.-Y.; Chua, T.-S.; and Li, Q. 2024. A picture is worth a graph: A blueprint debate paradigm for multimodal reasoning. In Proceedings of the 32nd ACM International Conference on Multimedia, 419–428. Zhou, J.; Chen, J.; Lu, Q.; Zhao, D.; and Zhu, L. 2025. Shielda: Structured handling of exceptions in llm-driven agentic workflows. arXiv preprint arXiv:2508.07935.