A 14B model's reasoning gets split across 1.7B, 4B and 8B collaborators, beating it on all ten metrics
Synopsis
A single large model that both generates and evaluates its own intermediate steps can favor its own outputs, letting errors propagate into later reasoning. A collaborative learning scheme replaces Qwen3-14B with three heterogeneous models of 1.7B, 4B and 8B under a 13.7B total parameter budget, using explicit task-state handoffs so different parameterizations reassess prior reasoning, then trains them with protocol distillation and step-level supervision to decide when to continue, hand off or terminate. Across ten metrics on five benchmarks, the heterogeneous system beats the original 14B model on all ten, with an average absolute gain of 2.38%, also surpassing a 24B homogeneous multi-agent baseline while using fewer model invocations than multi-agent frameworks.
Interpretation
Reasoning is split across several models with different parameters, so earlier steps are re-examined by another model instead of by the same model that produced them. Previously, a single model both generated and evaluated its own reasoning, tending to favor its own outputs, so errors in intermediate steps could go undetected and influence later reasoning. On ten metrics across MuSiQue, GSM-Hard, MATH, MultiPL-E and Conifer, the trained heterogeneous models beat the original Qwen3-14B on all ten, with an average absolute gain of 2.38%.
The heterogeneous models collaborate through explicit task-state handoffs, each deciding under a shared interaction protocol whether to keep control, hand the task state to another model, or terminate. A homogeneous multi-agent baseline uses three Qwen3-8B models totaling 24B parameters, while the heterogeneous system uses 13.7B yet beats it on all ten metrics. Against a homogeneous collaborative system trained with the same procedure, the heterogeneous system scores higher on all ten metrics, with an average absolute gain of 6.02%.
Training proceeds in two stages: the largest model slice generates demonstrations to initialize cross-model interaction, then the original large model scores each reasoning and interaction decision within complete trajectories. Outcome-level supervision alone gives insufficient guidance for intermediate reasoning and interaction decisions during collaboration. Three ablations, removing collaborative training, protocol distillation, and the step-level score, drop average scores by 9.37, 8.85 and 5.41 points respectively, each underperforming the full method on every metric.
The trained models need fewer invocations at inference and repeat their reasoning far less. Multi-agent frameworks typically require more model calls, and successive steps tend to overlap in their reasoning. Across five benchmarks the collaborative system uses 39.1% to 84.8% fewer model invocations than MAD, AgentVerse, AFlow and GPTSwarm; on MuSiQue the three models' reasoning overlap rates fall from 48.3%, 24.1% and 55.2% to 0.2%, 0.3% and 0.5%.
Perspective
The results cover three model sizes and five benchmarks, with a parameter budget comparable to the original Qwen3-14B. The work suits researchers and engineers working on multi-model collaborative reasoning and training methods. The source proposes future work on generalization to larger model pools, longer interaction horizons and unseen task distributions, and on more robust supervision through multiple inspectors or uncertainty-aware aggregation.
The source states that the study covers only three model sizes and five benchmarks, leaving generalization to larger model pools, longer interaction horizons and unseen task distributions unexplored. Pyramid supervision depends on the quality of the inspector and evaluator, whose biases may affect step-level feedback. Comparable parameter counts do not imply equal computational cost, since collaborative inference adds latency and memory overhead beyond the parameter-weighted output-token metric.