Autonomous agents powered by large language models (LLMs) perform complex tasks through long-horizon reasoning and tool interaction, where a fundamental trade-off arises between execution efficiency and reasoning robustness. Models at different capability-cost levels offer complementary advantages: lower-cost models enable fast execution but may struggle on difficult reasoning segments, while stronger models provide more robust reasoning at higher computational cost. We present AgentCollab, a self-driven collaborative inference framework that dynamically coordinates models with different reasoning capacities during agent execution. Instead of relying on external routing modules, the framework uses the agent's own self-reflection signal to determine whether the current reasoning trajectory is making meaningful progress, and escalates control to a stronger reasoning tier only when necessary. To further stabilize long-horizon execution, we introduce a difficulty-aware cumulative escalation strategy that allocates additional reasoning budget based on recent failure signals. In our experiments, we instantiate this framework using a two-level small-large model setting. Experiments on diverse multi-step agent benchmarks show that AgentCollab consistently improves the accuracy-efficiency Pareto frontier of LLM agents.
Figures & tables
Figure 1: Overview of the proposed AgentCollab framework. (a) The agent system first invokes the large model to warm up the interaction and plan the overall strategy. (b) After initialization, the small model performs a low-latency routine reasoning steps and (c) conducts self-evaluation on progress checks to determine whether escalation is required. (d) When stagnation is detected, control is temporarily transferred to a larger model within a predefined budget to resolve difficult reasoning segments. (e) Once the critical step is completed, control returns to the small model to continue the interaction.
Figure 2: Budget Allocation Strategy. As consecutive failures accumulate, the framework temporarily increases the budget for the larger model, allowing stronger models to provide additional support.
BrowseComp_zh
HLE-math
WritingBench
Agent
Method
Acc. (%)
#Steps
Speedup
Acc. (%)
#Steps
Speedup
Score
#Steps
Speedup
DDV2
Small
18.3
21.75
1.54×
8.0
35.55
3.38×
4.4
17.25
3.20×
Large
34.6
17.29
1.00×
23.3
28.41
1.00×
5.1
9.10
1.00×
Random
22.3
20.56
1.12×
9.2
32.82
1.33×
4.2
13.10
1.29×
RouteLLM
28.2
22.4
1.09×
15.2
33.42
1.62×
4.7
10.27
2.21×
FrugalGPT
27.2
20.2
1.29×
18.2
29.23
1.97×
4.5
12.22
1.94×
Table 1: Main results on three multi-turn agent benchmarks. Speedup is measured relative to the large-model baseline in each agent framework. AgentCollab substantially improves the performance of small-model agents while maintaining strong efficiency advantages. Results for WritingBench are not reported for WebSailor because it does not support long-form writing tasks.
Figure 3: Pareto frontier of DDV2 on BrowseComp_zh.
Figure 4: Illustrative execution trajectories. The first row shows a representative case, and the second row shows an extreme case. Panels (a) and (c) plot cumulative latency bars and instantaneous TPS curves over reasoning steps, with speedup ratios annotated relative to the Large baseline. Panels (b) and (d) show the corresponding cumulative latency staircase plots.
Method
Acc. (%)
Speedup
Large-model only
34.6
1.00×
Small-model only
18.3
1.54×
Role Assignment
w/ Large Planner
24.6
1.39×
w/ Large Executor
27.3
1.24×
Escalation Strategy
Table 2: Ablation study on DDV2. We examine the effect of planner/executor role assignment and escalation strategy. Accuracy and speedup are reported.
Figure 5: Pareto frontier of AgentCollab within WebSailor on BrowseComp_zh under different values of k .
Figure 6: Failure analysis. Box plots comparing successful and failed cases in terms of (a) large-token ratio (the proportion of tokens generated by the Large model), (b) number of reasoning steps, and (c) latency.
Configuration
Acc. (%)
#Step
Speedup
3B
9.7
9.44
1.37 ×
3B + 7B
12.5
11.40
1.08 ×
7B
14.2
10.38
–
3B + 32B
12.8
9.08
1.40 ×
7B + 32B
22.5
8.49
1.50 ×
32B
25.5
7.79
–
Table 3: WebSailor on BrowseComp_zh with different model size configurations. Configurations such as 3B + 7B denote collaboration between two models of different sizes. Speedup is computed against 7B in the upper block and against 32B in the lower block.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Speedup ratio as a function of task difficulty. Task difficulty is measured by the total latency of the Large (32B) baseline for each task (x-axis), and the y-axis reports the corresponding speedup ratio. Blue circles denote the speedup of Ours (7B–32B), while green crosses denote the speedup of the Small-only model (7B). Solid lines show linear trend fits. The red dashed line at y=1 indicates parity with the Large baseline. The proposed AgentCollab maintains a clear positive trend under a selective large-model allocation strategy.
Small Model
AgentCollab
Large Model
Let me analyze this task: The task requires searching for a historically significant document jointly signed by multiple national representatives, with a focus on the English copy stored in a Chinese museum. I should begin with a web search.
Let me analyze this task: I need to search for a multi-nation jointly signed historical document, stored in a Chinese museum, in English. Possible candidates include the UN Charter, Treaty of Versailles, etc. I need both Chinese and English keywords.
Let me analyze this task: Find an English document jointly signed by multiple national representatives with major historical significance, and stored in a Chinese museum. I should focus on international treaties, UN documents, and museum collections in China.
Analysis: All three systems begin with high-quality reasoning from the large model. The Small Model mode also starts with large model planning to compare them in a more fair configuration.
……
……
……
Now I have obtained detailed information on the Cairo Declaration… Let me search for other similar important documents.
I still have not found a multi-nation signed English document stored in a Chinese museum. Let me try searching for the “UN Conference Memorial Album” from the Museum of Overseas Chinese History. [Progress check: no significant progress made. Escalate.]
I found important information: UN Charter drafting documents (Chongqing Library), “UN Conference Memorial Album” (Museum of Overseas Chinese History), and other multi-nation historical materials.
Analysis: The left system enters a local loop with limited search breadth; the AgentCollab system encounters stagnation and triggers model escalation; the right system continues high-quality knowledge expansion.
……
……
……
Appendix
Table 4: Comparison of key THINK steps across three reasoning modes : small-model-only, AgentCollab, and large-model-only.
Enhancing the reasoning capabilities of large language model (LLM) agents requires effective orchestration of diverse expert models and tools. However, existing frameworks typically call APIs, based on coarse-grained matching between tasks and the functions of expert models or tools, while overlooking critical factors such as performance variability and cost efficiency among functionally similar alternatives. To address this, we propose Agora, a framework that uses a confidence-calibrated auction to dynamically allocate tasks to expert models and tools. By treating reasoning steps as tradeable items, Agora bases allocation on calibrated competence rather than raw confidence. Across five main benchmarks, Agora improves or remains competitive with single-model, routing, and cascade baselines under matched candidate pools.
Multi-agent large language model (LLM) systems enable complex, long-horizon reasoning by composing specialized agents, but practical deployment remains hindered by inefficient routing, noisy feedback, and high interaction cost. We introduce REDEREF, a lightweight and training-free controller for multi-agent LLM collaboration that improves routing efficiency during recursive delegation. REDEREF integrates (i) belief-guided delegation via Thompson sampling to prioritize agents with historically positive marginal contributions, (ii) reflection-driven re-routing using a calibrated LLM or programmatic judge, (iii) evidence-based selection rather than output averaging, and (iv) memory-aware priors to reduce cold-start inefficiency. Across multi-agent split-knowledge tasks, we show that while recursive retry alone saturates task success, belief-guided routing reduces token usage by 28%, agent calls by 17%, and time-to-success by 19% compared to random recursive delegation, and adapts gracefully under agent or judge degradation. These results demonstrate that simple, interpretable probabilistic control can meaningfully improve the efficiency and robustness of multi-agent LLM systems without training or fine-tuning.
Mohammad Parsa Hosseini, Ankit Shah, Saiyra Qureshi +3
Multi-agent large language model (LLM) systems have shown promise for solving complex tasks through agent collaboration. However, existing frameworks assign tasks based on predefined roles without considering whether an agent can accurately assess its own competence boundaries, leading to overconfident execution on tasks beyond its expertise. Inspired by metacognition theory from cognitive science, we propose MetaCogAgent, a multi-agent LLM framework where each agent is equipped with a Metacognitive Self-Assessment Unit that evaluates task-capability alignment before execution. The framework introduces three contributions: (1) a self-assessment mechanism that estimates per-task confidence by combining verbalized uncertainty with historical capability profiles; (2) an adaptive delegation protocol that routes low-confidence tasks to better-suited agents through cross-agent evaluation; and (3) a capability boundary learning module that iteratively refines each agent's competence model via cybernetic feedback. Experiments on our constructed MetaCog-Eval benchmark (700 tasks across 5 cognitive dimensions) demonstrate that MetaCogAgent achieves 82.4% task accuracy -- 8.7% above the best routing baseline -- while using 5% fewer API calls than AutoGen and 34% fewer than ensemble voting. Ablation studies confirm that each metacognitive component contributes to overall system performance.
Chenyu Wang, Yang Shu
School of Computer and Artificial Intelligence, Zhengzhou University, Zhengzhou, China · Zhejiang University, Hangzhou, China