What votes throw away,
traces preserve.
When multiple LLM agents work on the same problem, the usual move is to take a vote. The answer with the most support wins. Researchers at Bioscope AI asked a different question: what gets lost when you do that? The answer turns out to be: sometimes, the correct solution.
First surfaced in Tandemly Briefing — 2026-05-30.
Voting compresses
what matters most.
Mixture-of-agents systems have a standard playbook: run N agents in parallel, collect their answers, and pick the winner by majority vote. The question is what gets dropped in that final step.
Mixture-of-agents (MoA) systems run multiple language model instances on the same problem and combine their outputs. The idea is borrowed from ensemble methods in classical machine learning: many estimators, properly combined, outperform any single one. For language models, the standard approach has been to treat each agent's final answer as a vote and count them up.
This works reasonably well. Empirically, running more agents and voting tends to beat running one agent. But the improvement has limits. Add more agents, inject more prompt diversity, try different samplers: the accuracy ceiling barely moves. The researchers behind this paper wanted to understand why that ceiling exists and whether it can be broken.
Their diagnosis: the ceiling is structural. Error correlations between agents are high, even when prompts vary. When one agent gets a question wrong, the others tend to get it wrong in similar ways. Diversity in surface prompting does not translate into diversity in failure modes. So the vote stays biased toward the wrong answer in exactly the cases that matter: hard questions where the correct reasoning path appears only in minority chains, not in the consensus.
What if the correct reasoning is buried inside a minority chain that majority voting discards? Can an aggregator that reads the full traces recover it, even in cases where every agent voted wrong?
An aggregator that
reads the whole story.
Instead of collecting final answers and voting, the team kept the full reasoning chains from all agents and gave them to a separate LLM to read. That reading model synthesizes across all the traces rather than picking a winner from a list of conclusions.
The core method is trace-level synthesis. Each agent in the pool produces not just a final answer but a complete chain of reasoning: every intermediate step, every pivot, every sub-conclusion that led to the result. The aggregator receives all of these chains and reads them together, not as a ballot box.
Because the aggregator sees everything, it can identify a correct intermediate step from an agent whose final answer was wrong. It can notice when two minority chains agree on a sub-step that the majority chains skipped. It assembles a response from the most credible reasoning across all the traces, rather than deferring to whichever conclusion received the most votes.
The paper also introduces Self-Consistent Mixture of Agents, or SC-MoA. This variant adds trace diversity through semantic-preserving input perturbations: slight rewordings of the original problem that should not change the correct answer but do change the reasoning path each agent takes. To protect against cases where the synthesizer introduces an error the majority did not have, SC-MoA uses anchored refinement. The synthesis is only accepted if it does not degrade below the majority baseline. This gives a provable non-degradation guarantee. The method always synthesizes, and it never gates on whether there is consensus first.
Agents with different prompts tend to make different intermediate errors, even when they converge on the same wrong final answer. The correct path exists somewhere in the trace pool more often than the correct answer appears in the answer pool. Trace-level synthesis exploits this gap: the signal is richer in the full chains than in the stripped conclusions.
The vote was wrong.
The traces knew better.
Three results define the paper. All three point in the same direction: the final answer is the wrong unit of analysis when aggregating across agents.
The trace-reading aggregator recovers correct answers from cases where all agents agreed on the wrong answer. This is the aggregation paradox: consensus in the answer pool does not mean consensus in the trace pool. The reasoning chains contain complementary information that the vote discards.
Across the five benchmarks evaluated, beneficial corrections by the trace-reading aggregator consistently outweighed harmful ones. The net effect is positive: reading traces does more good than harm, accounting for cases where the synthesis introduces an error the majority did not have.
The authors measured error correlations between agents across all experimental conditions: varied prompts, varied samplers, varied agent counts. The correlations remained high regardless. This means the ceiling that majority voting hits is not a product of insufficient diversity; it is a property of how agents share failure modes at the level of final answers. Adding more agents does not help past a point because the new agents fail in the same ways as the existing ones.
The Self-Consistent Mixture of Agents variant achieves the highest accuracy across all five benchmarks evaluated. The anchored refinement mechanism provides a formal guarantee that SC-MoA's output will not be worse than the majority baseline in expectation. Calibrated uncertainty estimates come as a byproduct of the synthesis process, without any additional computation beyond what the trace reading already produces.
The specific benchmark names, numbers, and full experimental conditions are detailed in the paper directly. The provable non-degradation guarantee applies to the anchored variant of SC-MoA under the conditions tested. How the aggregation paradox scales with larger agent pools, or generalizes across task types beyond those in the evaluation, remains an open question the authors acknowledge.
Replace the vote
before you scale N.
Anyone running multi-agent systems with majority voting or self-consistency should read this paper. The architectural change it proposes is contained: replace the aggregator. The reasoning behind it changes how you think about what a parallel rollout is actually for.
Where to go
from here.
If you want to dig deeper or start experimenting.