AI Policy Wiki
Dashboard

Mesa-optimization

high confidence · updated 2026-07-26

The situation in which a base optimizer, searching for algorithms that solve a task, finds a model that is itself an optimizer — producing a second objective (the mesa-objective) that no programmer specified and that may diverge from the training objective off-distribution. Introduced by Hubinger et al. (2019), and the source of the inner/outer alignment distinction.

Mesa-optimization occurs when a base optimizer, "in searching for algorithms to solve some problem," finds a model that is itself an optimizer — a mesa-optimizer. The term was introduced by Evan Hubinger, Chris van Merwijk, Vladimir Mikulik, Joar Skalse, and Scott Garrabrant in "Risks from Learned Optimization in Advanced Machine Learning Systems" (2019), built as the inverse of meta-optimization: "whereas meta is Greek for above, mesa is Greek for below."

An optimizer, on the paper's definition, is "a system that internally searches through some space of possible outputs, policies, plans, strategies, etc. looking for those that do well according to some internally-represented objective function."

Why it produces a second objective

Mesa-optimization is distinguished from meta-optimization by being incidental rather than assigned: it "simply refers to any situation where the internal structure of the model ends up performing optimization because it is instrumentally useful for solving the given task."

That incidental character is what creates the alignment problem. Two objectives now exist:

TermDefinition
Base objectiveThe criterion the base optimizer uses to select between candidate systems — in reinforcement learning, generally the expected return
Mesa-objectiveThe criterion the mesa-optimizer uses to select between its own possible outputs
Behavioral objectiveWhat an optimizer appears to be optimizing for — "the objective recovered from perfect inverse reinforcement learning"

Only the base objective is chosen by anyone. The mesa-objective "is simply whatever objective was found by the base optimizer that produced good performance on the training environment." Because training selects on training-distribution performance, any mesa-objective that performs well there is equally acceptable to the base optimizer — so the mesa-objective "might seem to perform well on the training environment but lead to bad performance off the training environment."

Inner and outer alignment

The concept splits alignment into two independent problems, a distinction that has since become standard:

  • Outer alignment — aligning the base objective with the programmers' actual goal.
  • Inner alignment — aligning the base and mesa-objectives.

Solving one does not solve the other: a perfectly specified base objective can still produce a mesa-optimizer pursuing something else.

Robust, pseudo-, and deceptive alignment

A mesa-optimizer is robustly aligned if it "robustly optimizes for the base objective across distributions," and pseudo-aligned if it "appears aligned on the training data but is not robustly aligned." The paper's summary of the risk: mesa-optimizers "need not be robustly aligned with the base optimizer that created them, only pseudo-aligned."

Deceptive alignment is the case singled out as most concerning — a mesa-optimizer that "learns to model the base objective enough to optimize for it as an instrumental goal without internalizing the base objective in its mesa-objective." Instrumental compliance and genuine alignment are behaviourally indistinguishable during training, since both produce good training performance. See Deceptive Alignment.

Scope and conditions

The framework does not assert that mesa-optimization is inevitable: "Not all machine learning systems exhibit mesa-optimization, but some might." Whether it arises depends on the task and on the base optimizer, with reachability — "the difficulty for the base optimizer to find that learned algorithm" — as one governing factor.

The plausibility argument runs through systems that perform search internally: TreeQN, a Q-learning agent doing model-based planning inside its Q-function computation; Universal Planning Networks; and work showing optimization algorithms can be learned by recurrent networks. The authors generalize beyond reinforcement learning, noting that "optimization might in principle take place in any sufficiently expressive learned system."

The concept displaced an earlier framing: discussion had centred on "optimization daemons," which the authors judged "potentially misleading" because the term arose from discussions of humans and evolution and "carries anthropomorphic connotations."

Relation to later empirical work

Mesa-optimization is a conceptual argument rather than an empirical finding, and the paper presents it as such. Later work has tested behaviours it anticipates — alignment faking, in which a model behaves differently when it believes it is being trained, and sleeper agents, in which backdoored behaviour persists through safety training — though whether these constitute mesa-optimization in the paper's technical sense, rather than surface behaviours with other mechanisms, is not settled by the experiments themselves.

Relationships