A 14B model split into 1.7B, 4B and 8B slices: 13.7B total beats the original on all ten metrics

Reasoning no longer runs end to end inside one model; several models with different parameters take turns, so each intermediate conclusion is re-examined under a different parameterization. Previously a single model both generated and evaluated its own solution, favoring its own outputs, so errors in intermediate steps could stay undetected and shape later reasoning. Across ten metrics on MuSiQue, GSM-Hard, MATH, MultiPL-E and Conifer, the 13.7B heterogeneous combination exceeds the original Qwen3-14B on every one, with an average absolute gain of 2.38%.

Training proceeds in two stages: models first learn when to keep control, hand the task state to another model, or stop, and are then adjusted by how much each step contributes to the whole collaboration trajectory. Outcome-level supervision alone gives insufficient guidance for intermediate reasoning and handoff decisions during collaboration. Removing collaborative training, protocol distillation, or the step-level score drops the ten-metric average by 9.37, 8.85 and 5.41 points respectively, and each variant trails the full method on every metric.

After training, the division of labor shifts with the task, repeated reasoning falls sharply, and the system needs fewer model invocations than multi-agent frameworks. An untrained homogeneous multi-agent system scores below the heterogeneous collaborative system on all ten metrics, by 6.02% on average. 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%, and invocations drop 39.1% to 84.8% against MAD, AgentVerse, AFlow and GPTSwarm.

The results cover three model sizes and five benchmarks, spanning multi-hop question answering, mathematical reasoning, code generation and constrained generation, and are aimed at researchers and engineers working on multi-model reasoning and training pipelines. 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 inspector and evaluator quality, 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.