Organizations: Laboratory of Brain Atlas and Brain-inspired Intelligence, Institute of Automation, Chinese Academy of Sciences (CASIA) · School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS) · University of the Chinese Academy of Sciences (UCAS) · College of Software (CS), Taiyuan University of Technology · School of Future Technology, University of Chinese Academy of Sciences (UCAS)
Large Language Models (LLMs) have demonstrated remarkable capabilities across a wide range of language tasks, yet complex multi-step reasoning remains a fundamental challenge. While Large Reasoning Models (LRMs) equipped with extended chain-of-thought mechanisms demonstrate improved performance over standard LLMs, both model types still suffer from accuracy collapse on sufficiently complex tasks, suggesting that scaling model-level reasoning alone is insufficient. Inspired by the global workspace theory of human cognition, we propose Brain-Inspired Graph Multi-Agent Systems (BIGMAS), in which specialized LLM agents are organized as nodes in a dynamically constructed directed graph and coordinate exclusively through a centralized shared workspace. A problem-adaptive GraphDesigner constructs task-specific agent topologies, while a global Orchestrator leverages the complete shared state for routing decisions, overcoming the local-view bottleneck of reactive approaches. Experiments on Game24, Six Fives, and Tower of London across six frontier LLMs demonstrate that BIGMAS consistently improves reasoning performance for both standard LLMs and LRMs, outperforming existing multi-agent baselines including ReAct and Tree of Thoughts, showing that multi-agent architectural design provides complementary gains orthogonal to model-level reasoning enhancements.
Figures & tables
Fig. 1: Three cognitive reasoning tasks used in evaluation. Left: Tower of London task requires planning optimal moves to reach target configuration. Middle: Six Fives requires constructing arithmetic expressions using exactly six 5s to reach a target value. Right: Game24 demands mathematical reasoning to combine four numbers reaching target value 24.
Fig. 2: Overview of the BIGMAS framework. (a) Graph Design : A GraphDesigner agent D analyzes the problem instance P=(x,C,y∗) and produces a task-specific directed agent graph G together with a Workspace contract κ . (b) Workspace B : A centralized shared workspace partitioned into read-only context Bctx , read-write working area Bwork , system metadata Bsys , and sink-only answer store Bans ; all agent nodes interact exclusively through B . (c) Graph Execution : Each active node vt produces a structured write instruction ωt=(π,α,δ) via an LLM call; the instruction is validated against B and G , with a self-correction loop on failure. A global Orchestrator routes execution based on the complete workspace state; the system halts when the sink node vsnk is reached or the step budget Tmax is exhausted.
Fig. 3: Grouped bar chart comparing the accuracy (%) of six LLMs on three reasoning benchmarks under two conditions: Base LLM (solid bars) and BIGMAS (hatched bars). The six models evaluated are DeepSeek-V3.2, DeepSeek-V3.2 (+thinking), Claude 4.5 Sonnet, Claude 4.5 (+thinking), Gemini 2.5 Pro, and GPT-5. The three tasks are Game 24 (arithmetic reasoning), Six Fives (constrained expression generation), and Tower of London (multi-step planning). For every model and task, BIGMAS consistently matches or improves upon the corresponding base-LLM accuracy, with particularly large gains on weaker base models (e.g., DeepSeek-V3.2 and Claude 4.5 Sonnet) and smaller but still positive gains on already strong models (e.g., GPT-5 and Gemini 2.5 Pro).
Method
Game 24
Six Fives
Tower of London
Base LLM
25.0
12.0
6.0
ReAct
26.0
18.0
10.0
Tree of Thoughts
30.0
25.0
18.0
BIGMAS (Ours)
36.0
30.0
20.0
TABLE I: Accuracy (%) of DeepSeek-V3.2 under four reasoning frameworks across three tasks.
Fig. 4: Distribution of graph complexity across three reasoning tasks, shown as violin plots overlaid with jittered individual observations. The left panel reports the number of nodes and the right panel the number of directed edges in each agent graph automatically designed by the GraphDesigner .
Fig. 5: Representative agent graph structures automatically designed for each of the three reasoning tasks. Each panel displays the highest-node-count graph produced by the system for that task; all three instances were solved correctly (indicated by ✓).
Fig. 6: Heatmap of agent node role distribution across three reasoning tasks. Each cell reports the proportion (%) of nodes assigned to a given functional role category within all graphs produced for that task; raw node counts are shown in italics below each percentage. Color intensity encodes the proportion on a white-to-navy scale.
Fig. 7: Absolute token consumption per architectural phase across three reasoning tasks, shown as grouped box plots (one box per model per phase, outliers omitted). Node Execution dominates in all tasks, accounting for 46.4%, 50.9%, and 55.5% of total tokens for Game 24, Six Fives, and Tower of London respectively. Orchestrator Routing consumes a bounded but non-trivial share (16.9%–25.2%), growing with task complexity. Graph Design incurs a fixed up-front cost that shrinks as a proportion of total tokens when execution steps are numerous.
Fig. 8: Distribution of orchestrator routing decisions per run across three reasoning tasks, displayed as violin plots overlaid with jittered individual observations. Routing count serves as a proxy for execution-time complexity. Game 24 concentrates near one decision per run (mean 1.3 ), Six Fives shows moderate spread (mean 2.6 ), and Tower of London yields the highest and most variable counts (mean 7.9 , maximum 49 ).
Fig. 9: Mean number of orchestrator routing decisions for correctly solved (green bars) versus incorrectly solved (red bars) instances, broken down by model and task. Error bars denote 95% confidence intervals. Incorrect runs consistently require more routing calls than correct ones across all three tasks.
Existing multi-agent learning approaches have developed interactive training environments to explicitly promote collaboration among multiple Large Language Models (LLMs), thereby constructing stronger multi-agent systems (MAS). However, during inference, they require re-executing the MAS to obtain final solutions, which diverges from human cognition that individuals can enhance their reasoning capabilities through interactions with others and resolve questions independently in the future. To investigate whether multi-agent interaction can enhance LLMs' independent problem-solving ability, we introduce ILR, a novel co-learning framework for MAS that integrates two key components: Dynamic Interaction and Perception Calibration. Specifically, Dynamic Interaction first adaptively selects either cooperative or competitive strategies depending on question difficulty and model ability. LLMs then exchange information through Idea3, an innovative interaction paradigm designed to mimic human discussion, before deriving their respective final answers. In Perception Calibration, ILR employs Group Relative Policy Optimization (GRPO) to train LLMs while integrating one LLM's reward distribution characteristics into another's reward function, thereby enhancing the cohesion of multi-agent interactions. We evaluate the effectiveness of ILR across three LLMs from two model families of varying scales on five mathematical, one coding, one general question answering, and one scientific reasoning benchmarks. Experimental results show that ILR consistently outperforms single-agent learning, yielding an improvement of up to 5% over the strongest baseline. We further discover that Idea3 can enhance the robustness of stronger LLMs during multi-agent inference, and dynamic interaction types can boost multi-agent learning compared to pure cooperative or competitive strategies.
Hehai Lin, Shilei Cao, Sudong Wang +5
1The Hong Kong University of Science and Technology (Guangzhou) · 2Sun Yat-sen University · 4National University of Singapore +2
While Large Language Models (LLMs) excel in certain reasoning tasks, they struggle in multi-agent games where the final outcome depends on the joint strategies of all agents. In multi-agent games, the non-stationarity of other agents brings significant challenges on the evaluation of the reasoning process and the credit assignment over multiple reasoning steps. Existing single-agent reinforcement learning (RL) approaches and their multi-agent extensions fail to address these challenges as they do not incorporate other agents in the reasoning process. In this work, we propose Strat-Reasoner, a novel RL-based framework that improves LLMs' strategic reasoning ability in multi-agent games. We introduce a novel recursive reasoning paradigm where an agent's reasoning also integrates other agents' reasoning processes. To provide effective reward signals for the intermediate reasoning sequences, we employ a centralized Chain-of-Thought (CoT) comparison module to evaluate the reasoning quality. Finally, we compute an accurate hybrid advantage and develop a group-relative RL approach to optimize the LLM policy. Experimental results show that Strat-Reasoner substantially improves strategic abilities of underlying LLMs, achieving 22.1% average performance improvements across various multi-agent games. Code is publicly available at https://github.com/ydhe1012/Strat-Reasoner.
Yidong He, Yutao Lai, Pengxu Yang +4
School of Software Engineering, South China University of Technology · Department of Computer Science, University of Oxford
We investigate whether explicit belief graphs improve LLM performance in cooperative multi-agent reasoning. Through 3,000+ controlled trials across four LLM families in the cooperative card game Hanabi, we establish four findings. First, integration architecture determines whether belief graphs provide value: as prompt context, graphs are decorative for strong models and beneficial only for weak models on 2nd-order Theory of Mind (80% vs 10%, p<0.0001, OR=36.0); when graphs gate action selection through ranked shortlists, they become structurally essential even for strong models (100% vs 20% on 2nd-order ToM, p<0.001). Second, we identify "Planner Defiance," a model-family-specific failure where LLMs override correct planner recommendations at partial competence (90% override, replicated N=20); Gemini models show near-zero defiance while Llama 70B shows 90%, and models distinguish factual context (deferred to) from advisory recommendations (overridden). Third, full-game evidence confirms inter-agent conventions (+128% over baseline, p=0.003) outperform all single-agent interventions, and individual belief-graph components must be combined to produce gains. Fourth, preliminary scaling analysis (N=10/cell, exploratory) suggests graph depth has diminishing returns: shallow graphs provide the best cost-benefit ratio, while deeper ToM graphs appear harmful at larger player counts (-1.5 pts at 5-player, p=0.029).
Yuqi Sun, Tianqin Meng, George Liu +4
Mindoverflow · University of Waterloo · Carnegie Mellon University +4