Scholar Feed

Mixture-of-experts routing

SteerMoE

Steering MoE LLMs via Expert (De)Activation

Superseded baseline#29 of 1,370 most-superseded · first seen Sep 11, 2025

Superseded — cited as a baseline and beaten by newer methods

3 papers critique it · 3 beat it on benchmarks

What papers say

Verbatim critique sentences, each from a paper that cites SteerMoE as a baseline.

However, these approaches rely on observational analysis rather than proactive search: they depend on predefined unsafe/jailbreak datasets and are therefore constrained by the coverage of those sets. As a result, they typically reveal only modest shifts in harmful outputs while requiring prior data.
— Sparse Models, Sparse Safety: Unsafe Routes in Mixture-of-Experts LLMs
“It relies on a frequency-based analysis, assigning a Risk Difference (RD) score to each expert based on activation rate differences between prompt sets representing faithful and unfaithful responses.”
— MASCing: Configurable Mixture-of-Experts Behavior via Activation Steering Masks
“SteerMoE suppresses unsafe experts at inference time by modifying routing logits, but does not update expert parameters or repair unsafe representations.”
— RASA: Routing-Aware Safety Alignment for Mixture-of-Experts Models

Beaten on benchmarks

Head-to-head results where a newer method reports beating SteerMoE. Values are copied from the source paper's tables — verify against the cited paper.

What to use instead

Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.