MoE Routing May Be the Confidence Signal Text Voting Was Pretending to Be

Majority voting works best on answers that behave like answers: short, canonical, easy to normalize, and preferably surrounded by a box. Code does not behave that way. Neither do patches, tool trajectories, debugging plans, or most useful agent outputs.

That is why Does the Same Token Mean the Same State? MoE Routing as Signal for Reasoning Control is interesting. The paper asks whether sparse Mixture-of-Experts routing patterns preserve information that the visible token stream erases. Its answer: yes. Even when two rollouts emit the same token, expert routing can still reveal different internal states, task histories, and reasoning modes.

The authors turn that observation into RAD, Routing Agreement Decoding. Instead of selecting a candidate by parsing answer strings or voting over final text, RAD represents each rollout by MoE routing states around fixed anchor windows and picks the densest route basin using Weighted-Jaccard k-nearest-neighbor agreement. It is not a verifier. It is a selector. That distinction is doing important work.

Text voting was always a leaky abstraction for code

The evaluation covers 10 sparse-MoE configurations across gpt-oss 20B/120B, Qwen3-30B-A3B, and Qwen3-Next-80B-A3B. The datasets span AIME24/25, BRUMO25, HMMT25, GPQA, and LiveCodeBench v5, totaling 4,850 model-problem cells and 310,400 generated runs. On math and GPQA, where string voting is well-posed, RAD is roughly comparable to textual majority: RAD scores 73.9, RAD+DC 74.2, and majority 73.6.

The more interesting result is SWE-bench Verified. For agentic patch selection, the paper samples 16 rollouts per 500 problems per model — 8,000 rollouts per model. On decidable SWE-bench subsets, RAD improves mixed selection by roughly four to five points: Qwen3.6-35B-A3B moves from 55.8 to 60.5, gpt-oss-120B from 58.7 to 63.6, and gpt-oss-20B from 52.0 to 57.0. End-to-end averages improve as well: Qwen from 49.2 to 51.4, gpt-oss-120B from 55.4 to 57.8, and gpt-oss-20B from 41.3 to 44.2.

Those are not miracle numbers. Good. Miracle numbers usually need a nap and a reproducibility audit. The value is that RAD attacks a real interface mismatch. If sixteen coding-agent rollouts produce sixteen diffs, what is the majority answer? The most common final sentence? The patch with the highest textual similarity? The one touching the most popular file? The one that compiles? At that point, “majority voting” becomes a polite fiction pasted over an open-ended artifact.

Routing agreement offers a different signal: are these rollouts internally moving through the same latent basin? Sparse MoE models route tokens through experts, and those routing decisions can act like a fingerprint of the computation path. If multiple candidates cluster in route space, RAD treats that cluster as meaningful agreement even when the output is not easy to canonicalize.

White-box control planes get another reason to exist

The deployment catch is not small. RAD needs access to MoE routing tensors and a patched serving path; the paper reports about 7% overhead during rollout collection. This is not available from most closed black-box APIs. It is a strategy for open or controllable sparse-MoE deployments where the runtime can capture internal routing windows without storing every token-level detail forever.

That limitation is also the strategic point. Open model stacks have often been sold on cost, privacy, and customization. This paper adds another advantage: internal observability. If routing tensors, activations, logits, cache behavior, and tool traces are available to the control plane, the runtime can make decisions that a black-box API cannot expose. Not all of those signals will be useful. Some will overfit. Some will be brittle across model families. But the direction is right: agent reliability will increasingly depend on runtime signals below the generated text.

Practitioners should resist turning RAD into magic confidence. The paper is explicit that a dense wrong basin can still win. Consensus is not truth; it is agreement. A model can converge internally around a wrong patch, just as text self-consistency can converge around a wrong answer. RAD belongs next to execution, tests, static analysis, human review, and cost-aware routing — not in place of them.

The immediate action item is to inventory where your agent selection policy still assumes outputs are easy to vote on. Math? Fine. Multiple-choice? Fine. Code patches, multi-step tool plans, data migrations, incident writeups, policy decisions? Much less fine. If you are already paying for multiple rollouts, you need a selector that matches the artifact. Sometimes that selector is unit tests. Sometimes it is a verifier. For sparse-MoE models, routing agreement may become one more cheap-ish triage signal: which candidate should be tested first, which traces deserve expensive validation, which basin looks unstable enough to sample again.

LGTM’s read: this is self-consistency growing up for agentic outputs. Text voting was always the wrong abstraction for code. MoE routing may give open models a better agreement signal — not because it proves correctness, but because it measures a kind of internal convergence that final strings cannot represent.

Sources: arXiv, SWE-bench, Qwen3 technical report, Self-consistency decoding