Sign InOpen Brain
arXivPaperNeeds Review

Consolidating RLVR Capabilities Across Domains: A Deep Dive into Fusion Paradigms

Three ways to combine RLVR domain experts perform similarly on average but diverge by task. Choose Merge for cheap reuse, Mix RL for training from pooled data, and MOPD for preserving expert gains.

arXiv · Aug 27, 2026
Open Source Open MarkdownOpen JSON
Source Summary

The study compares **Merge**, **Mix RL**, and **MOPD** using shared experts and data. Average performance differs by at most **1.4 points**, while one benchmark shows an **8.6-point** gap; all improve single-sample accuracy.

Practical Implication

Choose Merge when trained experts already exist and fusion cost dominates. Use Mix RL to train one model from pooled data, tuning domain proportions; use MOPD when retaining domain-specific gains matters more than exceeding teachers or minimizing total cost.

Agent-Ready Context
The study compares **Merge**, **Mix RL**, and **MOPD** using shared experts and data. Average performance differs by at most **1.4 points**, while one benchmark shows an **8.6-point** gap; all improve single-sample accuracy.

Choose Merge when trained experts already exist and fusion cost dominates. Use Mix RL to train one model from pooled data, tuning domain proportions; use MOPD when retaining domain-specific gains matters more than exceeding teachers or minimizing total cost.

None of the three produced measurable gains in solution coverage, though held-out capabilities were not measurably reduced. Per-domain behavior can depart sharply from the close average, so aggregate scores are insufficient for selection.
Connected Context · Feed7 Judgment

This narrows RLVR consolidation from choosing a universal best method to matching deployment constraints: reuse cost favors Merge, pooled training favors Mix RL, and expert retention favors MOPD. Close averages conceal substantial task-level differences, and unchanged solution coverage means fusion should not be treated as evidence of broader reasoning exploration.

Agora: Enhancing LLM Agent Reasoning Via Auction-Based Task AllocationAgora preserves separate experts and allocates reasoning steps at runtime, while these fusion methods consolidate expertise into one model; they represent alternative cost and orchestration points for using specialization.Data Quality Is the Compute Multiplier — Ari Morcos, DatologyAIThe need to tune domain proportions for Mix RL makes data balancing and task matching direct implementation prerequisites rather than generic training hygiene.Ling 3.0 Flash Fin now available on AI Gateway for freeA finance-specialized gateway model exemplifies keeping a domain expert as a separate route; fusion offers alternatives when builders instead want one consolidated model, without showing that consolidation improves coverage.OctoLong: Mid-Training On Cross-Repository Code Contexts Enhances Long-Context ModelingOctoLong reinforces that domain gains depend on task-shaped training data, supporting this study’s warning that aggregate fusion scores can hide sharply different per-domain behavior.
Context Map
modelresearch#reasoning#model-selection
Uncertainty
None of the three produced measurable gains in solution coverage, though held-out capabilities were not measurably reduced. Per-domain behavior can depart sharply from the close average, so aggregate scores are insufficient for selection.