Structural priors boost LLMs in-domain but collapse out-of-distribution across tasks

The key buried point is that this isn't primarily a finding about code security or vulnerability detection. It's a claim about the generalizability of a failure mode: structural priors that help models route to correct reasoning in one domain actively hurt them when the surface features change, and no amount of domain-specific scaffolding reliably fixes this.

This connects directly to the academic supervision study covered the same day ('Harnessing LLMs for Reliable Academic Supervision'), which argued that deterministic scaffolding around smaller models outperforms raw scale in high-stakes settings. That paper treated scaffolding as a solution; this paper complicates the picture by showing that certain scaffolding interventions, specifically structural priors injected as cheatsheets or domain hints, can introduce their own failure modes when the distribution drifts. Together, the two papers sketch a tension practitioners will need to resolve: scaffolding helps, but the wrong scaffolding can actively mislead a model that has already learned to rely on structural shortcuts.

Watch whether the SAIR benchmark gets adopted outside this team's own evaluations. If independent groups reproduce the routing ceiling on at least two additional code-adjacent domains within the next six months, the failure mode is robust; if replication stalls, the effect may be narrower than claimed.