black O2, inc.
ACTIVE
    FOR-001 / PeopleFIG-008

    Lauro Langosco di Langosco

    Active person

    CAUTION This record carries a caution marking. Read it before relying on it.

    title
    Lauro Langosco di Langosco
    Type
    person
    Identifier
    FIG-008
    Status
    Active
    Tags
    • figure
    • alignment
    • generalisation
    Aliases
    • Langosco et al.
    • Langosco
    connections
    • [[CON-005]]
    • [[SYS-001]]

    Lauro Langosco di Langosco

    Lead author of Goal Misgeneralization in Deep Reinforcement Learning (CIT-005, arXiv:2109.10430), cited for the finding that capability and objective can diverge out of distribution — an agent can retain the competence it was trained for while pursuing a different objective than the one intended.

    In the Corpus

    Langosco supports the L3 brittleness claim in SYS-001: training is not a durable binding because what generalises is the capability, not the specification. It is the empirical counterpart to the hierarchy's structural argument — the table says training fails to deliberate circumvention; Langosco supplies the failure that does not require any circumvention at all.

    The corpus pairs this with the sandbagging literature (van der Weij et al., named but not registered) to cover both directions: capability hidden deliberately, and objective drifting unintentionally.

    ⚠ not retrieved — identifier given by the discussion at L636, not checked. Register rates the risk moderate because the finding is widely replicated in the literature, but "widely replicated" is the discussion's assessment, not evidence gathered here.

    Connections

    • CON-005 — sandbagging, the deliberate counterpart.
    • SYS-001 — the L3 claim.

    Back to Shackled AIForests