DeAR: how decentralized agents reason together without a central coordinator
The new DeAR framework shifts agentic reasoning to a peer-to-peer basis, eliminating the bottlenecks of centralized routing. By anchoring to capabilities and navigating a mind map, it consistently outperforms previous methods across nine benchmarks.

Introduction
Classic multi-agent systems are typically built around a "dispatcher": it decides which agent should handle a task, collects responses, and produces the result. This approach works reasonably well in simple scenarios, but the more complex the request, the more noticeable the weaknesses become. The central node turns into a bottleneck: routing slows down, agent roles are fixed in advance, and the system struggles to adapt to unconventional multimodal data.
A recent preprint, DeAR, proposes a different path — fully decentralized interaction. Agents negotiate among themselves on their own, without a single coordinator, and decide for themselves who is responsible for what. The idea is to make collaboration more flexible and resilient — and, judging by the results, it works.

Why centralization hinders agents
The familiar "router + executors" scheme looks logical, but it has systemic limitations. All requests pass through a single node, and its throughput determines the speed of the entire system. If a complex request arrives that requires the joint work of several specialized agents, the central planner must quickly figure out which part to assign to whom. In practice, this often leads to some agents sitting idle while others are overloaded.
Another problem is the static nature of roles. An agent assigned as "text-only" won't be able to help with an image, even if its underlying model is capable of more. Rigid boundaries prevent the system from using its real potential.
DeAR tries to move away from these limitations. Instead of a central planner, agents communicate in peer-to-peer mode. Each participant independently assesses how well its capabilities fit the current request and decides on its own in which direction to move the reasoning next.
How DeAR works
At the core of the framework are three mechanisms that together enable autonomous coordination.
Capability binding
The first step is decentralized capability binding. Each agent receives the request and "tries it on": it assesses what it can actually do and how relevant those skills are to the current task. No external planner assigns roles — specialization emerges naturally, based on the specific request. This makes the system more flexible: the same agent might work with text in one case and help analyze an image in another.
Thought map navigation
The second mechanism is thought map navigation. Once an agent has defined its area of responsibility, it needs to understand who to interact with next and how. Instead of randomly messaging everyone, agents build a targeted "map" of ideas and reasoning, along which they move toward the relevant participant. This approach reduces unnecessary traffic and makes collaboration meaningful: each step brings the system closer to an answer rather than merely creating the appearance of activity.
Topology update
Finally, the ability to correct mistakes on the fly matters. DeAR includes a topology update — an adaptive mechanism that revises connections between agents if the chosen reasoning path hits a dead end. The system doesn't wait for a central node to notice the problem; it restructures its own architecture right on the fly.

What the tests showed
The authors evaluated DeAR on nine diverse benchmarks covering both multimodal reasoning and text-based question answering. In all cases, the decentralized framework consistently outperformed recent baseline methods.
The gain is especially noticeable on tasks that require intensive knowledge use and the combined efforts of different specializations. This confirms the main thesis: when agents can freely adapt and connect with each other directly, reasoning accuracy improves.

Conclusions
DeAR is a step toward a more natural and resilient architecture for agent systems. Instead of a rigid hierarchy — an autonomous network where each participant understands its own capabilities and chooses how to contribute to solving the task. This approach not only removes the bottleneck problem but also makes the system ready for unexpected scenarios that cannot be anticipated with static role assignment.
For now, the framework exists as a research project: the source code is promised to be released after publication. But it's already clear that decentralized agent collaboration is a promising direction, and DeAR looks like one of its compelling examples.
Frequently asked questions
Related materials
All materials
Checking LLM Hallucinations via SQL-Based Source Representation Without Fine-Tuning
The study examines whether an LLM can apply a basic symbolic skill—building a SQL database from source documents—to ground its outputs in sources without supervision when detecting hallucinations. On RAGTruth and DiaHalu, this neuro-symbolic pipeline proves more accurate than direct prediction and comparable to state-of-the-art methods, without requiring domain-specific fine-tuning.
9 November 2026
MathAdv: What the Final Accuracy of Machine-Generated Proofs Conceals
A diagnostic suite of tasks spanning 13 undergraduate and graduate mathematics topics evaluates machine-proving systems in Lean 4 not only by their final results, but also by their subject-matter mastery, the quality of their informal reasoning, and their robustness to rephrased conditions. A component-by-component analysis shows that translating a problem into formal form remains the main obstacle, and that success depends heavily on the area of mathematics and how the problem is framed.
9 November 2026
AffordanceWAM: Affordances as an intermediary between scene prediction and manipulator control
The work describes a generative model that jointly predicts future RGB observations, scalar fields, affordance heatmaps, and continuous robot commands, using human videos to broaden interaction diversity without action annotations or retargeting. On RoboCasa, CALVIN ABC→D, and real-world tasks, the approach delivers consistent gains over RGB-only and robot-data-only baselines.
9 November 2026