2026-08-28·by Sijie Wang#cybernetics#theory

stages-gates-as-hard-optimization

Stages-and-gates, from the optimization angle

Given the ODE/optimizer lens: "one landscape, descend to the end" is the convex/well-conditioned world. On a hard landscape (non-convex, high-dim, ill-conditioned) direct end-to-end descent sticks in bad local minima or diverges. Stage-and-gating is not exotic — it's the classic hard-optimization toolkit, three things at once.

① stages = homotopy / continuation / annealing / curriculum

Optimize a sequence of gradually-hardening subproblems, each warm-starting the next: numerical continuation/homotopy, simulated annealing (temperature schedule = smooth→sharp landscapes), graduated non-convexity, curriculum learning. Stage = a subproblem on the path; gate = "converged here, warm-start the next." You shape the descent path because direct descent on a hard landscape sticks.

② decompose = splitting / coordinate descent / ADMM

Don't optimize all variables jointly — split into sub-landscapes, solve, then reconcile: ADMM (subproblems + consensus), block coordinate descent, Gauss–Seidel. Stage = a block/split; gate = the consensus/reconcile step = our decompose-and-reassemble. Joint optimization is intractable; splitting makes each piece tractable.

③ gate = trust-region / line-search acceptance

The most direct hit for "soft→hard gate": propose a step, then a gate decides accept-or-reject (enough decrease? model trustworthy? ratio ρ good?) — else reject/shrink/retry (trust-region; Armijo sufficient-decrease). The gate collapsing soft belief into a hard commit = the trust-region acceptance step.

So "necessity" re-read

Gates/stages are necessary for the same reason homotopy/trust-regions are necessary in hard optimization: direct end-to-end descent on a hard/ill-conditioned landscape fails; you must structure the process. necessity is the agent-level wording; "hard landscapes need continuation + splitting + trust-regions" is the optimization wording — same necessity.

The key twist: the LLM is itself an optimizer → this is bi-level optimization

  • LLM inference is a kind of optimization (in-context learning ≈ implicit gradient descent; OPRO — "LLMs as optimizers"). Inner loop = the LLM optimizer.
  • Stage-and-gating = a meta-optimizer (continuation schedule + trust-region acceptance) wrapped around the LLM inner optimizer. Bi-level.
  • And the inner optimizer is unsound (noisy/hallucinated gradient, gets stuck) → the outer structure is what makes the unreliable inner optimizer converge: the gate = trust-region acceptance on the LLM's proposed step (commit iff verifiably better); the stages = a curriculum shaping its descent around bad local minima.

One line

Stage-and-gating = the hard-optimization toolkit (homotopy/continuation + splitting/ADMM + trust-region acceptance) lifted to the agent level and wrapped around the LLM as an unsound inner optimizer. It looks like "partitioning the landscape" because that's exactly what continuation/splitting/multigrid do to hard landscapes — and it's needed more here because the inner gradient (LLM judgment) is unsound, so the trust-region acceptance isn't acceleration, it's survival. Necessity, optimization version: on a hard and unsound landscape, direct descent fails; stages-and-gates bolts on continuation + trust-regions to force a bad inner optimizer to converge.

about this entry

One of sijie's wiki entries. The AI on this site is grounded in the same corpus and answers in sijie's voice, with citations back to entries like this one — answering costs sijie money, so it waits behind a code: enter an access code →

stages-gates-as-hard-optimization