Existing Multi-Agent Systems (MAS) typically rely on homogeneous model configurations, failing to exploit the diverse expertise inherent in different post-trained architectures. We propose Team-of-Thoughts, a heterogeneous MAS framework that treats diverse models as specialized tools within an orchestrator-driven paradigm. Team-of-Thoughts introduces two novel components: (1) Orchestrator Calibration, which identifies models with superior coordination and synthesis capabilities, and (2) Agent Self-Assessment, a protocol where tool agents profile their own domain-specific strengths to guide selection. At inference, the orchestrator dynamically activates the most compatible agents based on these profiles to maximize capability coverage. Across five mathematical reasoning and code generation benchmarks, Team-of-Thoughts consistently outperforms individual models and existing MAS baselines. Notably, on AIME24 and LiveCodeBench, Team-of-Thoughts achieves 96.00% and 77.91% accuracy, respectively, significantly improving over homogeneous role-play baselines (80.00% and 65.93%).
Figures & tables
Figure 1: Pass@5 accuracy versus cost on MBPP+ for Team-of-Thoughts, single models, and AgentVerse. The dashed line denotes the Pareto front of single models.
Figure 2: Overview of Team-of-Thoughts. (Left) Unlike single-model reasoning or homogeneous MAS (role-play), Team-of-Thoughts leverages heterogeneous model priors to maximize solution space coverage. (Right) Our architecture utilizes a pre-inference pipeline—comprising orchestrator calibration and agent self-profiling—to enable dynamic agent selection. During inference, the orchestrator selectively invokes optimal tool agents and synthesizes their outputs into a high-confidence final response.
Figure 3: Schematic comparison of Chain-of-Thought (CoT) and Team-of-Thoughts. (Left) Standard agentic reasoning via CoT generates sequential intermediate steps to refine the prediction distribution heuristically. However, on complex tasks, CoT may fail to converge on distant targets. (Right) Team-of-Thoughts utilizes heterogeneous tool agents to explore a broader solution space. During inference, these agents refine their individual distributions through local reasoning, which an orchestrator then prioritizes and aggregates into the final output.
AIME 2024
AIME 2025
Humaneval+
MBPP+
LiveCodeBench v6
Model
pass@1
@3
@5
@1
@3
@5
@1
@3
@5
@1
@3
@5
@1
@3
@5
Claude Sonnet 4.5
84.00
86.67
86.67
74.67
87.67
90.00
95.61
96.34
96.34
81.80
84.29
85.45
48.24
55.33
57.69
GPT-5-mini
88.67
93.33
93.33
73.33
83.50
86.11
92.68
94.88
95.12
82.38
86.19
87.57
64.84
74.40
77.20
Gemini 3 Flash
86.67
90.67
93.33
71.33
81.33
86.67
96.46
96.95
96.95
84.44
85.13
85.45
72.53
75.71
77.47
Deepseek v3.2 Exp
94.67
96.67
96.67
94.00
98.67
100.00
91.26
95.88
96.65
80.20
84.59
85.83
76.48
86.65
88.46
GPT OSS 20b
75.33
88.00
90.00
73.33
83.50
86.11
89.23
95.12
95.63
79.23
84.88
86.29
46.04
54.34
58.24
Table 1: Performance comparison of Team-of-Thoughts across five benchmark tasks. Success rates (%) are reported using pass@1, 3, and 5. Team-of-Thoughts is evaluated against individual model baselines and state-of-the-art multi-agent methods. The Theoretical Limit represents the oracle upper bound achieved by selecting the best result among all individual models for each problem.
Model
AIME2024
MBPP+
Budget Cost ($)
0.03
0.02
Avg
0.03
0.02
Avg
Claude Sonnet 4.5
86.67
46.67
66.67
78.38
78.38
78.38
GPT-5 Mini
86.67
93.33
90.00
86.49
83.78
85.14
Gemini 3 Flash
86.67
40.00
63.34
83.78
83.78
83.78
DeepSeek v3.2
93.33
93.33
93.33
81.08
81.08
81.08
GPT-OSS 20B
80.00
80.00
80.00
75.68
75.68
75.68
Table 2: Orchestration performance across candidate models and budget constraints. Calibration accuracy (%) is evaluated on AIME2024 and MBPP+ benchmarks under two distinct budgetary tiers (USD). Avg denotes the mean accuracy across all evaluated settings.
Benchmark
Single
Self-Assess
Orch-Based
Random
AIME2024
94.67
96.00
92.67
94.00
MBPP+
82.38
85.93
85.77
84.44
Table 3: Performance evaluation of tool agent assessment strategies. Comparison of Pass@1 accuracy (%) with no tool agent (“Single”), Orchestrator-Led Assessment, Agent Self-Assessment, and Random Invocation across AIME2024 and MBPP+ benchmarks.
AIME 2025
MBPP+
Config
pass@1
@3
@5
pass@1
@3
@5
No traces
95.33
99.33
100.00
85.93
88.84
89.68
Include traces
78.67
93.33
96.67
80.90
85.61
87.04
Table 4: Performance of orchestrator aggregation strategies. Comparison of task accuracy (%) achieved when incorporating versus excluding the intermediate reasoning traces of tool agents.
Activated tools
Pass@1
Pass@3
Pass@5
2
82.75%
85.58%
86.77%
4
82.96%
85.32%
85.98%
8
60.32%
83.76%
86.24%
16
60.63%
83.10%
85.71%
Table 5: Performance under different numbers of activated tool agents, with 25 total tool agents available.
Activate 2
Activate 8
Agreement
Avail Tools
Pass@1
@3
@5
Pass@1
@3
@5
A (%)
10
82.70
86.32
87.57
74.97
85.26
85.98
90.74
25
82.75
85.58
86.77
60.32
83.76
86.24
91.80
50
82.43
86.59
87.83
54.76
82.17
86.51
61.64
75
82.49
87.09
88.36
63.65
84.21
86.24
75.13
100
82.49
86.56
87.57
61.43
84.13
86.51
71.43
Table 6: Orchestrator performance and tool selection consistency (Agreement) across varying ensemble sizes. Performance is compared between configurations activating two and eight agents per tool call.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Construction Method
pass@1
pass@3
pass@5
Random Sample
77.91%
85.05%
87.91%
Categorized Sample
81.65%
85.93%
86.81%
Appendix
Table 7: Performance in pass@k on LiveCodeBench v6 using different calibration construction strategies.
Figure 4: Comparing the times of tool agent invocations between two calibration dataset construction methods.
Orch context len
Pass@1
Pass@3
Pass@5
16K
60.32%
83.76%
86.24%
32K
59.79%
82.99%
85.71%
64K
59.42%
83.39%
86.51%
128K
60.32%
82.96%
85.45%
Appendix
Table 8: Performance under different orchestrator max context lengths, with eight of 25 available tool agents activated on each tool call.
Multi-Agent Systems (MAS) built on Large Language Models (LLMs) require effective orchestration to coordinate specialized agents, yet training such orchestrators is hindered by limited supervision and high computational cost. We propose Orchestration Reward Modeling (OrchRM), a self-supervised framework for evaluating orchestration quality without human annotations. OrchRM leverages intermediate artifacts from multi-agent executions to construct win-lose pairs for Bradley-Terry reward model training. Unlike existing MAS test-time scaling and orchestrator training frameworks that rely on costly sub-agent rollouts, OrchRM operates directly at the orchestration level, enabling efficient and high-performing reward-guided orchestrator training and MAS test-time scaling. OrchRM improves training efficiency by up to 10x in token usage while improving MAS test-time scaling performance by up to 8% in accuracy. These gains consistently transfer across multiple domains, including mathematical reasoning, web-based question answering, and multi-hop reasoning, demonstrating orchestration-level reward modeling as a scalable direction for robust multi-agent orchestration. Code will be available at https://github.com/Wang-ML-Lab/OrchRM.
King Yeung Tsang, Zihao Zhao, Vishal Venkataramani +5
Multi-agent systems (MAS) demonstrate clear advantages in tackling complex problems by coordinating diverse agents and external tools. However, most existing orchestration methods rely on static workflows or serial agent scheduling, and are further constrained by heterogeneous interface protocols between tools and agents. This leads to high system complexity and poor extensibility. To mitigate these issues, we propose Agent-as-Tool, a unified parallel orchestration paradigm that abstracts both agents and tools into a standardized, learnable action space with protocol normalization and explicit state feedback. Building on this paradigm, we train a lightweight orchestrator, ParaManager, which decouples planning decisions from subtask solving, enabling state-aware parallel subtask decomposition, delegation, and asynchronous execution. For training, we adopt a two-stage ParaManager training pipeline. It improves robustness by incorporating supervised fine-tuning (SFT) trajectories equipped with recovery mechanisms, and further applies reinforcement learning (RL) to achieve an optimal balance among task success, protocol compliance, diversity, and reasoning efficiency. Experiments show that ParaManager achieves strong performance across multiple benchmarks and exhibits robust generalization under unseen model pools.
Wenzhen Yuan, Wutao Xiong, Fanchen Yu +7
Shanghai Jiao Tong University · Sichuan University · Shanghai AI Lab +2
Multi-agent systems (MAS) can scale large language model agents on long-horizon tasks by running them in parallel, yet existing designs waste much of this parallelism in bubbles: agent time spent waiting on others or redoing a peer's work. These bubbles stem from how agents communicate. Independent agents share nothing and rediscover what their peers have already found; peer-communicating agents wait at synchronous rounds; and under centralized orchestration, the main agent blocks on its sub-agents while progress is relayed. We propose Decentralized Language Models (DeLM), a MAS framework on top of existing agent harnesses that squeezes out these bubbles by replacing the main agent with a shared context and a task queue. Agents asynchronously claim tasks, publish findings as soon as they are available, and build on or correct one another's progress, with every peer's status visible to all. On long-horizon tasks from Terminal-Bench 4.0 and DeepSWE v1.1, and on SWE-bench Verified, DeLM is both more accurate and faster than Codex, Claude Code, their native subagents, and AOrchestra in every setting, improving accuracy by up to 17.5 points over the strongest baseline and running up to 2.49x faster than the harness it builds on. On ProgramBench, where agents rebuild programs from scratch, DeLM makes faster progress than Claude Code and finishes a 120-minute budget up to 19.9 points higher in test pass rate. The code is available on our project website at https://yuzhenmao.github.io/DeLM/.