Source-linked AI summary
DeAR: Decentralized Agentic Reasoning via Capability Grounding and Collaborative Thought Navigation
Xing Wei, Changmeng Zheng, XiaoYong Wei, Xiufen Ye, Qing Li
TL;DR
Centralized agentic reasoning protocols struggle with complex multimodal queries because rigid roles and central bottlenecks hinder adaptive, verified collaboration. DeAR enables autonomous peer-to-peer reasoning through capability grounding, thought-map navigation, and topology updates, consistently outperforming recent baselines across nine multimodal and text-based QA benchmarks.
Problem
Existing agentic reasoning systems lack decentralized, query-adaptive collaboration for complex multimodal queries, with rigid role allocation and centralized bottlenecks limiting reasoning workflows.
Method
DeAR replaces centralized coordination with capability grounding, thought-map navigation, and topology updates for autonomous peer collaboration and progressive reasoning refinement.
Results
Across nine multimodal and text-based QA benchmarks, DeAR consistently outperforms recent baselines, achieving the highest overall scores on MathVista (59.41), ChartQA (89.43), and ScienceQA (62.45).
Takeaways & Limitations
The results support decentralized, adaptive agent collaboration as an effective approach for complex knowledge-intensive reasoning tasks.
Abstract
from arXiv · showhide
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.
Introduction
The introduction argues that centralized agentic reasoning struggles with complex multimodal queries because rigid coordination creates routing bottlenecks, information loss, and error propagation. DeAR addresses this gap by enabling autonomous peer-to-peer collaboration through dynamic capability grounding and adaptive reasoning contexts.
- Motivation: Agentic reasoning decomposes intricate problems into specialized sub-tasks, supporting multi-step problem solving beyond simple pattern matching.The introduction frames complex reasoning as a frontier for LLM applications and agentic reasoning as an effective paradigm.
- Limitations of Existing Systems: Complex multimodal queries expose failures in both single-agent and centralized multi-agent systems when visual recognition and historical knowledge retrieval must be combined.A single agent may lack simultaneous fine-grained visual and historical capabilities, while centralized coordination can propagate incorrect inputs.
- Limitations of Existing Systems: Centralized collaboration introduces routing and aggregation bottlenecks, information loss, and rigid role allocations that misroute dynamic tasks and propagate unverified errors.The introduction attributes these failures to structural flaws in centralized topologies rather than deficient model capabilities.
- Motivation: Real-world complex problem-solving is decentralized, with interdisciplinary experts engaging in peer-to-peer discourse, negotiating boundaries, and self-organizing across knowledge gaps.This topology motivates exploring decentralized collaboration for agentic reasoning.
- DeAR: DeAR enables agents to reason and collaborate autonomously while maintaining evolving reasoning context as dynamic message payloads passed through peer-to-peer token streams.The framework is introduced to bypass bottlenecks of centralization and restructure role allocation, collaboration mechanisms, and reasoning paths.
- DeAR: Dynamic capability grounding replaces static role labels by having agents assess their suitability for specific query fragments using verifiable capabilities.The introduction identifies static labels as weak proxies for competence that can cause coordinators to hallucinate capabilities in open-ended scenarios.
Related work
Prior work shows that coordinating agents with complementary capabilities improves LLM and MLLM robustness and interpretability while supporting complex task decomposition and intermediate-state sharing. Specialized-agent frameworks address multi-step reasoning, but debate-based and linear approaches rely heavily on rigid, centralized coordination protocols.
- Multi-agent collaboration: Multi-agent systems improve LLM and MLLM robustness and interpretability by coordinating agents with complementary capabilities.These frameworks also support complex task decomposition, intermediate-state sharing, and dynamic workflow management.
- Specialized-agent reasoning: Specialized-agent frameworks distribute cognitive loads across agents to address complex multi-step reasoning tasks.The passage cites recent frameworks pursuing this design, including Jiang et al. (2025a) and Liu et al. (2025).
- Centralized coordination: Debate-based approaches and linear reasoning paths refine intermediate states but rely heavily on rigid, centralized coordination protocols.This limitation is identified for debate-based approaches and linear reasoning paths in complex multi-step reasoning.
DeAR Framework
DeAR replaces centralized routing and fixed agent roles with autonomous, query-dependent collaboration. Its framework grounds capabilities from benchmark evidence, navigates a dynamically constructed thought map, and updates topology through localized backtracking when reasoning reaches a dead end.
- Thought Map Navigation: Thought-map navigation sequentially selects high-propensity unvisited agents, propagates accumulated knowledge, and lets the terminal agent generate the final answer without centralized judging or predefined topology.The trajectory terminates dynamically based on real-time capability assessments rather than a fixed step limit.
- Decentralized Capability Grounding: DeAR agents ground their identities in verifiable technical-report or model-card evidence rather than rigid predefined roles.Each agent maintains proficiencies in commonsense reasoning, explanatory ability, numerical computation, natural-language comprehension, and factual reliability.
- Decentralized Capability Grounding: Benchmark-derived capability profiles are calibrated onto a common scale, then contextualized per query to condition each agent’s intermediate reasoning.The framework uses quantitative benchmark scores and produces a query-dependent capability state τ_i(q) before generating knowledge K_i.
- Self-Organized Collaboration: Agents autonomously evaluate peer utility and form a directed collaboration matrix whose preferences reflect agent heterogeneity and task-specific requirements.The first active agent is selected uniformly at random, while collaboration preferences are computed from calibrated proficiencies under the current query.
- Topology Update: When accumulated context cannot produce a valid answer, DeAR updates the collaboration topology and progressively backtracks from the failed terminal edge.The update applies a decay factor α, such as 0.5, to failed connections while avoiding full-path mitigation.
Experimental Setup
The experimental setup evaluates DeAR on four multimodal reasoning datasets and five text-based QA benchmarks using task-appropriate metrics and detailed subgroup analyses. Its dynamic thought map supports up to four query-dependent agents, with distinct MLLMs assigned to multimodal tasks.
- Benchmarks and Metrics: DeAR is evaluated on MMMU, MathVista, ChartQA, and ScienceQA using Accuracy (ACC) on the full test splits.MMMU and ChartQA use overall accuracy, while MathVista and ScienceQA include finer-grained analyses.
- Benchmarks and Metrics: MathVista reports accuracy by question type, including FQA, GPS, and MWP, and reasoning skill, including ALG, ARI, and GEO.These analyses complement the overall benchmark evaluation.
- Benchmarks and Metrics: ScienceQA reports accuracy across subjects, modalities, grade levels, and the average score.Reported categories include NAT, SOC, LAN; TXT, IMG, NO; and G1-6 and G7-12.
- Text-Based QA Evaluation: DeAR is additionally evaluated on five text-based QA benchmarks spanning single-hop and multi-hop datasets using Exact Match (EM) and F1.The multi-hop evaluations randomly sample 1,000 examples.
- Implementation: The thought map dynamically supports up to four agents depending on query complexity, assigning each multimodal agent a different MLLM.The multimodal models are DeepSeek-VL-7B-Chat, LLaVA-1.5-7B, Qwen2-VL-7B-Instruct, and MiniCPM-V-2.6, providing diverse visual reasoning perspectives.
Experimental Results
Across multimodal and text-only benchmarks, DeAR consistently outperforms centralized, multi-agent, and large-model baselines. Ablations and analyses show that thought-map navigation, topology updates, and adaptive agent activation drive these gains while improving efficiency.
- Multimodal Benchmarks: 18.07 points higher than Qwen2-VL-72B on MathVista, DeAR scores 59.41 versus 41.34 despite using fewer total parameters.The comparison attributes the gain to DeAR’s structured architecture rather than parameter scaling.
- Multimodal Benchmarks: DeAR achieves the highest scores among recent reasoning frameworks on MathVista (59.41), ChartQA (89.43), and ScienceQA (62.45), and scores 55.58 on MMMU.It also outperforms AutoGen, AgentVerse, and DyLAN across all four multimodal benchmarks, exceeding them by over 2.3 on ChartQA.
- Text QA Benchmarks: DeAR ranks first on NQ, achieves the highest TriviaQA F1 for single-hop tasks, and obtains the highest HotpotQA score while maintaining stable 2Wiki precision and completeness.These results extend decentralized thought navigation beyond multimodal reasoning to five text QA benchmarks.
- Ablation Analysis: Removing fixed roles decreases MMMU by 5.76 points, while removing topology updates decreases it by 2.37 points, separating decentralized navigation gains from ensemble effects.The ablations use identical backbones and compare centralized routing and ensemble-control variants.
- Thought Map Navigation: Thought-map navigation raises comprehensive MathVista accuracy from 51.68 to 59.41, with gains on Statistical Reasoning from 62.9 to 80.9, Logical Reasoning from 19.6 to 31.0, and Figure Question Answering from 69.1 to 80.3.The largest improvements occur on intricate multi-step logic and specialized visual-processing tasks.
- Topology Update: Topology updates improve MMMU from 53.21 to 55.58 and MathVista from 55.65 to 59.41 by correcting intermediate errors across multimodal datasets.The passage also reports similar gains on ChartQA, though its endpoint is truncated.
- Decay Factor Analysis: At α = 0.1, 2Wiki F1 drops to 53.91 from overcorrection, whereas α = 0.9 yields 55.02 from insufficient suppression; DeAR performs optimally in [0.5, 0.8].The decay factor controls penalties for failed reasoning paths.
- Adaptive Activation: 62.5% of MathVista queries are resolved by fewer than four agents, while 37.5% require all four, demonstrating adaptive activation and reduced unnecessary computation.DeAR dynamically constructs reasoning paths from query requirements and agent confidence.
Conclusion
DeAR replaces centralized multi-agent coordination with autonomous, capability-aware collaboration. Its three mechanisms support query-dependent specialization, selective peer interaction, and adaptive error correction without a central judge.
- Conclusion: DeAR replaces centralized multi-agent coordination with autonomous, capability-aware collaboration.The framework enables agents to collaborate without centralized control.
- Conclusion: Decentralized Capability Grounding enables query-dependent agent specialization.This mechanism grounds agents’ capabilities according to the query.
- Conclusion: Thought Map Navigation supports selective peer interaction, while Topology Update provides adaptive error correction.Together, these mechanisms allow agents to progressively refine reasoning trajectories without a central judge.