SafeRoute-MAS
A message routing method for multi-agent reasoning systems that uses coverage-aware and risk-calibrated two-stage decision-making to remove redundant expert messages while retaining key information, thereby reducing the risk of error propagation.
Timeline (last 90 days)
SafeRoute-MAS is a message routing method for multi-agent scientific reasoning that adopts a two-stage decision process: first assessing the coverage contribution of each expert message to the reasoning task, then combining it with a risk model to decide whether to retain or remove that message.
SafeRoute-MAS's routing decisions use only information from before message integration and do not depend on subsequent reasoning results, ensuring decision independence.
On 5,328 cached committees, SafeRoute-MAS achieved a gain of +1.80 percentage points at the source-compatible operating point (95% confidence interval [+1.13, +2.52]) while removing 12.76% of message tokens.
In 4,580 controlled integration experiments conducted on the gpt-oss-120b endpoint (stratified sampling from 458 committees), the sample estimate was -0.09 percentage points, and the design-weighted gain was +0.94 percentage points (95% confidence interval [-1.31, +3.50]), while reducing input by 18.72%.
The weighted token-matching comparison failed to establish an advantage for the selector (+1.68 percentage points, confidence interval [-1.27, +5.10]), indicating that SafeRoute-MAS's gains are unstable across different evaluation settings.
Of 88 cached non-neutral labels, only 13 (14.8%) maintained directional consistency under repeated-call majority voting.
Inconsistent label direction may stem from randomness in model outputs or uncertainty in evaluation criteria, posing a fundamental challenge to SafeRoute-MAS, which relies on labels for risk calibration.
The weighted upper confidence limit (UCL) of expected harm was 4.60%, but only 2 of 11 preset margins passed the support/risk rules.
The numerical qualification certification with Keep-All fallback proposed by the researchers was applied retrospectively rather than validated prospectively, so the actual deployment risk of SafeRoute-MAS has not yet undergone rigorous prospective testing.
All Facts (9)
SafeRoute-MAS is a message routing method for multi-agent scientific reasoning that adopts a two-stage decision process: first assessing the coverage contribution of each expert message to the reasoning task, then combining it with a risk model to decide whether to retain or remove that message.
50%UnverifiedSafeRoute-MAS's routing decisions use only information from before message integration and do not depend on subsequent reasoning results, ensuring decision independence.
50%UnverifiedOn 5,328 cached committees, SafeRoute-MAS achieved a gain of +1.80 percentage points at the source-compatible operating point (95% confidence interval [+1.13, +2.52]) while removing 12.76% of message tokens.
50%UnverifiedIn 4,580 controlled integration experiments conducted on the gpt-oss-120b endpoint (stratified sampling from 458 committees), the sample estimate was -0.09 percentage points, and the design-weighted gain was +0.94 percentage points (95% confidence interval [-1.31, +3.50]), while reducing input by 18.72%.
50%UnverifiedThe weighted token-matching comparison failed to establish an advantage for the selector (+1.68 percentage points, confidence interval [-1.27, +5.10]), indicating that SafeRoute-MAS's gains are unstable across different evaluation settings.
50%UnverifiedOf 88 cached non-neutral labels, only 13 (14.8%) maintained directional consistency under repeated-call majority voting.
50%UnverifiedInconsistent label direction may stem from randomness in model outputs or uncertainty in evaluation criteria, posing a fundamental challenge to SafeRoute-MAS, which relies on labels for risk calibration.
50%UnverifiedThe weighted upper confidence limit (UCL) of expected harm was 4.60%, but only 2 of 11 preset margins passed the support/risk rules.
50%UnverifiedThe numerical qualification certification with Keep-All fallback proposed by the researchers was applied retrospectively rather than validated prospectively, so the actual deployment risk of SafeRoute-MAS has not yet undergone rigorous prospective testing.
50%