Source-linked AI summary
Multi-Agent Relay Thinking
Xu Liu, Jungang Qiao, Guikun Chen, Xiaojuan Qi, Wenguan Wang
TL;DR
Existing multi-agent collaboration largely relies on parallel group thinking, while relay thinking transfers ownership of a shared reasoning trajectory for further refinement. MACR calibrates confidence and uses it to trigger handoffs and select the next agent. Across four multimodal benchmarks, it improves accuracy over single-model strategies and uses fewer tokens than multi-agent strategies while remaining comparable in results.
Problem
Most multi-agent collaboration methods use parallel group thinking, while relay-style transfer of an evolving reasoning trajectory is an alternative collaboration topology.
Source: §1 · pp. 1–2Method
MACR calibrates local confidence across heterogeneous models and uses confidence-guided bidding to initialize reasoning and relay generation when the leading model shows low confidence.
Source: §2.2 · p. 4Results
+3.3% on MathVista and +3.9% on MMStar over the strongest single-model strategies; MACR uses 6.0× fewer tokens than multi-agent strategies on HallusionBench and MMStar while achieving comparable accuracy.
Source: §1 · p. 2Takeaways & Limitations
MACR's relay collaboration integrates isolated single-model cues into a global reasoning chain and exploits complementary model capabilities across reasoning stages.
Source: §1 · p. 2
Abstract
from arXiv · showhide
Existing multi-agent collaboration (MAC) methods have primarily explored a group thinking collaboration topology, where agents independently generate solutions and aggregate their outputs. In this work, we investigate an alternative relay thinking topology, where reasoning ownership is dynamically transferred across agents, allowing subsequent agents to inherit and refine an evolving reasoning trajectory. We introduce Multi-Agent Confidence-guided Relay (MACR), a framework that enables dynamical relay collaboration among agents. The key challenge is determining when reasoning should be relayed and which agent should continue the process. method addresses this challenge through confidence-guided relay triggering and a bidding mechanism that selects the next agent based on its confidence. Experiments on four benchmarks show that method outperforms the strongest single-model strategies in accuracy, e.g., +3.3% on MathVista and +3.9% on MMStar, and remains on par with multi-agent strategies while consuming 6.0x less token overhead on HallusionBench and MMStar. Further analysis demonstrates that relay collaboration enables heterogeneous agents to progressively integrate complementary insights, combining heterogeneous capabilities across reasoning stages. Our source code is available at https://github.com/67L1/MACR.
1 INTRODUCTION
Most multi-agent collaboration uses parallel group thinking, while relay thinking offers a sequential alternative in which agents inherit and refine a shared reasoning trajectory. MACR uses calibrated confidence and bidding to manage handoffs, and the paper reports accuracy gains over single-model strategies with lower token overhead than multi-agent strategies.
1 INTRODUCTION
- Prior work has used multi-agent collaboration across mathematical reasoning, visual question answering, and multi-step decision making.The introduction frames collaboration as a way for multiple models to interact and exploit complementary capabilities.Source: §1 · p. 1
- Unlike group thinking, which combines parallel reasoning trajectories, relay thinking transfers an intermediate idea between participants for sequential refinement.Agents progressively inherit and extend a shared reasoning trajectory through ownership transitions.Source: §1 · pp. 1–2
- MACR calibrates confidence across heterogeneous models, then uses low confidence to trigger bidding and the highest-confidence bid to select the next agent.An initial bidding round selects the leading model; the process then monitors confidence and relays generation rights.Source: §1 · p. 2
- The paper identifies complementary model strengths across reasoning stages and bottlenecks that emerge at different stages across tasks as properties of heterogeneous collaboration.Its analyses also report that MACR integrates isolated single-model cues into a global reasoning chain.Source: §1 · p. 2
- +3.3% on MathVista and +3.9% on MMStar over the strongest single-model strategies; MACR uses 6.0× fewer tokens than multi-agent strategies on HallusionBench and MMStar.Experiments span four multimodal benchmarks.Source: §1 · p. 2
2 METHODOLOGY
MACR first calibrates confidence across heterogeneous models, then uses an online initiation–monitoring–bidding–relay workflow to transfer reasoning ownership. Its confidence signal is used to identify bottlenecks and select the next model.
2 METHODOLOGY
- MACR separates offline confidence calibration from online scheduling through initiation, monitoring, bidding, and relay.The workflow addresses when to transfer reasoning ownership and which model should continue.Source: §2 · pp. 2–3
2.1 HETEROGENEOUS MODEL CONFIDENCE SHIFT ANALYSIS AND CALIBRATION
- F = 13.31 indicates that between-model confidence differences outweigh within-model noise, making raw confidence mathematically incomparable.Local confidence is computed over a sliding window of generated log probabilities to identify reasoning bottlenecks.Source: §2.1 · p. 3
- F = 1.01 after standardization, with model confidence distributions aligned and correct trajectories showing higher local confidence.MACR uses each model’s historical expected mean and standard deviation for cross-model calibration during scheduling.Source: §2.1 · pp. 3–4
2.2 ONLINE BIDDING AND RELAY GENERATION
- At startup, candidate models generate w-token continuations from the initial context; MACR assigns generation to the model with the highest calibrated confidence, then monitors its local score.A drop below threshold τ is treated as a reasoning bottleneck and halts generation to trigger bidding.Source: §2.2 · p. 4
- When bidding, models generate candidate sequences from the shared context, and MACR selects the one with the highest calibrated confidence to append to the existing trajectory.Each model maintains its own KV cache, avoiding repeated prefilling.Source: §2.2 · p. 4
3.1 EXPERIMENTAL SETUP
The evaluation uses four multimodal benchmarks spanning broad multimodal abilities, hallucinations, visual logic, and visual-mathematical reasoning. MACR is compared with single-model and multi-agent baselines under shared evaluation settings and three random seeds.
3.1 EXPERIMENTAL SETUP
- Four benchmarks cover multimodal abilities, visual and knowledge hallucinations, structured visual reasoning, and joint visual-mathematical reasoning.The benchmarks are MMStar, HallusionBench, LogicVista, and MathVista.Source: §3.1 · p. 5
- Baselines include three 8B models, single-model prompting strategies, Majority Vote, and H-DMAD multi-agent collaboration.Single-model strategies are IO, CCoT, and DDCoT; H-DMAD assigns different reasoning strategies to agents and fuses their outputs.Source: §3.1 · p. 5
- The sliding window size is w = 16 and confidence threshold τ = −0.3, selected using 600 RealWorldQA instances disjoint from evaluation sets.Evaluations use identical settings and three random seeds; reported accuracy is mean ± standard deviation, alongside token use and IBC relative to each single model’s IO baseline.Source: §3.1 · p. 5
3.2 QUANTITATIVE COMPARISON RESULTS
MACR improves accuracy over the strongest single-model strategies while remaining comparable to multi-agent strategies with lower token overhead. Its reported IBC gains also favor MACR over H-DMAD.
3.2 QUANTITATIVE COMPARISON RESULTS
- +3.30% on MathVista (78.00% vs. 74.70%) and +3.87% on MMStar (72.00% vs. 68.13%) over the strongest single-model strategies.The comparison covers accuracy, token overhead, and IBC.Source: §3.2 · p. 5, Table 1 · p. 6
- MACR uses 6.00 × fewer tokens than multi-agent strategies on HallusionBench (4,414 vs. 26,375) and MMStar (4,150 vs. 24,677), despite their comparable accuracy.Source: §3.2 · p. 6
- MACR achieves a state-of-the-art IBC of 21.35, exceeding H-DMAD by +18.52 on MMStar (21.35 vs. 2.83) and +12.75 on MathVista (15.38 vs. 2.63).The table reports IBC in ×10−6.Source: §3.2 · p. 6, Table 1 · p. 6
3.3 QUALITATIVE ANALYSIS
Compared with single-model chain-of-thought reasoning, MACR integrates isolated cues into a more comprehensive global view. This integration is also visible when both approaches answer incorrectly.
3.3 QUALITATIVE ANALYSIS
- When single-model CoT fails, MACR synthesizes isolated cues into a more comprehensive global view; even when both are wrong, MACR still attempts broader integration.The examples contrast CoT’s focus on an isolated patch with MACR’s attempt to integrate cues across the image.Source: §3.3 · p. 6
3.4 ANALYSIS
The analysis examines model routing, generation-stage retention, task-dependent relay timing, and the effects of removing MACR components or model diversity. Ablations use 500 randomly sampled MMStar instances.
3.4.1 CAPABILITY ROUTING ANALYSIS
- MiniCPM-V 4.5 yields to InternVL 3.5 or Qwen3 VL in 40.3% of its bids, while InternVL–Qwen3 VL exchanges are bidirectional, indicating complementary routing.MiniCPM-V 4.5 has the highest self-transition rate (59.7%); InternVL 3.5 transfers to Qwen3 VL in 22.4%, while Qwen3 VL returns control in 26.9%.Source: §3.4.1 · p. 6
3.4.2 INFERENCE SURVIVAL ANALYSIS
- MiniCPM-V 4.5 has the highest early survival, InternVL 3.5 retains reasoning later, and Qwen3-VL has the poorest long-horizon retention.MiniCPM-V 4.5’s early number at risk is 3,151, roughly twice the other models’ 1,163 and 1,658; fixed sentence-boundary triggers degrade performance.Source: §3.4.2 · pp. 6–7
3.4.3 TASK-AGNOSTIC BOTTLENECK LOCATION ANALYSIS
- LogicVista and MathVista have denser early relays, whereas MMStar and HallusionBench concentrate relays later; bidding stabilizes in the final stage across tasks.Task-blind periodic relays degrade accuracy.Source: §3.4.3 · p. 7
3.4.4 ABLATION STUDY
- Fixed initiation reaches 71.6%, 71.2%, and 70.4% for MiniCPM-V 4.5, InternVL 3.5, and Qwen3 VL, respectively, versus 72.4% for MACR.The authors report that confidence-guided initialization outperforms each fixed pattern.Source: §3.4.4 · p. 7
- Fixed routing reduces accuracy from 72.4% to 69.0%, failing to exploit the models’ complementarity.The fixed order is MiniCPM-V 4.5 → InternVL 3.5 → Qwen3 VL.Source: §3.4.4 · p. 7
- Fixed sentence-boundary bidding reduces accuracy from 72.4% to 69.8%, truncating long-horizon reasoning and accumulating hallucinations from short-horizon models.The fixed trigger also cannot adapt to task-specific characteristics.Source: §3.4.4 · p. 7
3.4.5 ANALYSIS ON MODEL POOL HETEROGENEITY
- Homogeneous role-playing and temperature-decoded variants underperform MACR, supporting the authors’ conclusion that simulated diversity cannot substitute for complementary model heterogeneity.The results are reported as accuracy on MMStar.Source: Table 3 · p. 8, §3.4.5 · p. 8
3.4.6 LEXICAL PROFILING OF REASONING BOTTLENECKS
- Models show different lexical patterns before bidding: InternVL around evidence extraction, MiniCPM-V around spatial attributes, and Qwen3-VL around causal deduction.All three also frequently produce hesitation tokens such as “wait,” “look,” and “check.”Source: §3.4.6 · p. 8
4 RELATED WORK
Prior multi-agent debate work mainly organizes group-thinking workflows or diagnoses collaboration failures, while confidence research often evaluates sequence-level outputs. MACR instead uses calibrated local confidence to relay generation across heterogeneous models.
4 RELATED WORK
- Prior debate methods optimize collaboration workflows or diagnose limitations, but retain group thinking in which models generate complete responses that are aggregated.Workflow examples include role-based coordination, continuous-embedding communication, and topology or reward optimization; diagnosis work studies conformity, sycophancy, and system misalignment.Source: §4 · pp. 8–9
- Unlike rigid debate turns, MACR monitors confidence in real time and relays generation rights to the most confident candidate when a model struggles.Source: §4 · p. 9
- Prior confidence-guided collaboration evaluates sequence-level outputs and overlooks confidence comparability across heterogeneous models; MACR calibrates window-level local confidence for cross-model relays.The related work also reviews uncertainty quantification and confidence-based collaboration and reinforcement.Source: §4 · p. 9
5 CONCLUSION
MACR shifts multi-agent collaboration from group thinking to relay thinking, with reasoning transferred across heterogeneous models. It combines calibrated confidence and bidding to guide handoffs and reports stronger accuracy than single-model baselines with lower token overhead than multi-agent baselines.
5 CONCLUSION
- Across four multimodal benchmarks, MACR outperforms all existing single-model baselines in accuracy and achieves comparable results to multi-agent baselines with 6.0× less token overhead.The token-overhead comparison is reported across the benchmarks; the conclusion does not specify per-benchmark values.Source: §5 · p. 9
- MACR shifts multi-agent collaboration from group thinking to relay thinking, progressively transferring reasoning ownership across heterogeneous models.The next model inherits the evolving trajectory and continues its refinement.Source: §5 · p. 9
- To enable cross-model bidding, MACR calibrates local confidence and selects the most confident model for a handoff.It monitors the leading model and interrupts generation when it detects a reasoning bottleneck.Source: §5 · p. 9
- Further analyses indicate that relay collaboration integrates isolated cues into global reasoning chains and exploits model complementarity through task-agnostic selection.The conclusion characterizes the selection mechanism as task-agnostic.Source: §5 · p. 9