Reflection may improve sample efficiency under structured shifts

Type: kb/types/note.md · Tags: foundations, self-improving-systems

A reflective improvement pathway is one where changes route through the system's own causally connected self-representation. Among its expected advantages, reuse and transfer is the one that promises a payoff in sample efficiency. Here that narrows to fewer new target observations after a shift. The advantage itself is a hypothesis to be tested against built systems — since reflection buys addressability. The vocabulary is learned inductive commitments and their reach, fixed where reflection makes retained lessons second-order: a commitment shapes how the system generalizes, and reach is the scope it operates over — normally wider than the evidence that produced it. That is physicist David Deutsch's point in The Beginning of Infinity: a good explanation applies far beyond the problem that produced it. A theory with wide reach is already more accurate across that whole space, not only cheaper to adapt with when a shift arrives; this note narrows to the adaptation case specifically because it is the part that can be measured and falsified. Reach motivates the conjecture; it does not establish the statistical comparison this note stakes out.

The conjecture, narrowed to what could be measured and could be wrong: reflective routing can make a retained commitment's content and scope addressable. Addressability can support explicit, reusable hypotheses whose assumptions, applicability conditions, failures, and revisions are separately inspectable. A retained commitment may reduce the new target evidence required for adaptation when three things hold: it captures structure that remains stable across the task shift and its bet holds there; its reach covers the shift; and it is retrieved, applied, and validated economically. Each clause is a condition, and each can fail on its own: no stable structure, or a wrong commitment, is a failure of warrant; a shift outside the stated applicability conditions is a failure of reach; a hypothesis nothing surfaces is a failure of the retrieval wire.

A scoping note before the argument: the literature cited below supports the proposed transfer mechanism — reuse of causal, modular, compositional, or explicit structure — not computational reflection itself. The conjecture applies to reflective self-improvement only when the retained hypothesis participates in the causally connected self-representation through which the system changes its own behavior; explicit domain knowledge may transfer without making an improvement pathway reflective. The bridge from explicit representation to reflection is Commonplace's own, and it is conjectural end to end. It also assumes the evaluator has reach assessment: the capability to judge a candidate commitment's claimed reach as genuine rather than adaptive fit, which reflectivity does not supply on its own and which, in current practice, is observed in LLM-mediated evaluation without a theory for why.

The condition is a structured shift, not "off-distribution"

Generic "off-distribution" is too broad to predict anything: some shifts destroy every regularity a system could have retained, explicit or not. The expected advantage arises under structured shifts — the target cases differ enough that the existing parametric behavior would otherwise require adaptation, while some mechanism, rule, modular structure, or invariant remains stable across the shift and the retained hypothesis captures it. In commitment vocabulary: a structured shift is one the commitment's reach covers correctly — it operates there because its applicability conditions say so, and it helps there because the structure it names holds. The transfer literature supports exactly this conditional shape and no more: reusable causal mechanisms are proposed as what survives intervention-like change (Schölkopf et al. 2021), speed of adaptation to such change can be made a training signal (Bengio et al. 2019), and a growing library of explicit program components compounds across a task family (Ellis et al., DreamCoder, a program-synthesis system). None of this establishes a general superiority of explicit representations.

Generalization under shift stays difficult and assumption-dependent for every method: under the benchmarks and model-selection procedures studied by Gulrajani and Lopez-Paz (2020), the evaluated domain-generalization methods did not consistently outperform carefully implemented empirical risk minimization. In the settings analyzed by Rosenfeld, Ravikumar, and Risteski (2020), IRM (Invariant Risk Minimization) and related objectives can fail to recover the intended invariant predictor and need not improve over it. A harsher, bitter-lesson reading treats both results as an instance of a broader pattern — legible, addressable structure is usually exactly what a sufficiently scaled general method eventually absorbs and beats — rather than as generic difficulty. That reading is not rebutted here: naming a commitment's applicability conditions makes reach falsifiable per commitment, which answers a different worry (that the category is unfalsifiable) than the one just raised (that legible structure tends to lose to scale) — whether structured shifts as defined here are durable against that pattern, or just the next thing scale absorbs, remains open. Where no stable structure exists, or the hypothesis misses the structure that does, the conjectured advantage has nothing to bind to.

Parametric transfer is real; what is distinct is addressability

An earlier formulation said samples do not transfer and theories do. That is false as stated. Learned parameters and representations transfer, generalize, and support few-shot adaptation — features learned in one network transfer to related tasks (Yosinski et al. 2014), and meta-learned initializations adapt from a handful of examples (Finn, Abbeel, and Levine 2017) — with the transfer depending on learned invariances, prior training, and source–target similarity. What an explicit hypothesis offers is not transfer where parametric systems have none, but stronger, first-class addressability of the transferred thing: its applicability conditions, assumptions, failure modes, and revisions can each be considered separately; ordinary parametric transfer does not expose them by default as separately named, system-readable objects.

That default is not absolute, though: a capable base model conditioned on a prompt that states a mechanism, its applicability conditions, and its known failure modes in plain language is already operating on a separately named, inspectable object — the prompt text — without any commitment ever being persisted or reflected on. What persistence specifically adds over that transient case is availability across sessions without re-deriving or re-stating the commitment each time, and applicability outside any single context window. A third candidate advantage — that persisted structure can be checked by a machine, not only read by one — is plausible but untested here: nothing in this note establishes that persistence, rather than representational form, is what would make a commitment machine-checkable.

And scope is not warrant: a hypothesis can be general without being justified, and justified while having narrow scope. Generality is a property of what it says; warrant, of how it was checked. Either way the contest lives at the specific level of inductive bias — the generic bet that structure exists at all is held equally by every learner, so it is common ground neither pathway can exploit.

Two efficiencies, one ledger

The conjecture makes a sample-efficiency hypothesis: a reflective pathway may need fewer new target observations to reach a fixed performance level. Distinct from it is the economic qualification: that sample-efficiency advantage may shrink or disappear once the full cost is counted. Frame the total as amortized cost across a task family, and keep the ledger symmetric. The explicit pathway's ledger runs: hypothesis discovery, codification (writing the hypothesis into a persisted, checkable form), retrieval, applicability checking, validation, application, maintenance, and correction. The parametric ledger runs: pretraining, target adaptation data, optimization, and evaluation. Only the target-observation entries are the sample-efficiency claim; the rest is economics, and folding every ledger entry into "sample efficiency" conflates the two.

Three ledger entries deserve emphasis. Retrieval is a real discount on the explicit side, since retrieval failure is reflection failure: a hypothesis nothing surfaces contributes nothing. Parametric retention avoids a separate artifact-retrieval step because it is resident in the operative substrate — which does not guarantee that the relevant learned behavior will activate, generalize, or remain accessible in the current context; its discount is differently shaped, not absent.

The boundary must be drawn honestly: counting the historical evidence that produced the explicit theory as free, while charging the parametric learner for all its training data, biases the comparison before it starts.

And validation has more routes than criticism — proof, simulation, model checking, causal analysis, counterexample generation, targeted experiments. Criticism is the judgment-heavy route whose reliability bounds unattended use, since warranted autonomy is bounded by oracle domain — but it is not the only one.

What would test it

The comparison must isolate addressability, or it merely shows that one pathway received better prior knowledge. That comparison is also aspirational rather than available today: no full, autonomous, reflective self-improving system yet exists to run it on. The closest real instances are incomplete in different ways. The Gödel machine realizes the loop formally but was never implemented or run. Commonplace itself closes only some of its pathways this way: under its strictly computational boundary, a pathway is either reflective and thereby autonomous, or non-reflective and human-inclusive, never partially either — since a pathway cannot be both human-inclusive and reflective under that boundary — and most of Commonplace's own pathways are still the human-inclusive kind. Until a fuller instance exists, or a controlled harness substitutes for one, the design below states what evidence would settle the conjecture, not evidence available to run it against.

Closing that gap does not require waiting for a practical system to finish closing its remaining human-inclusive pathways. One route is to build a deliberately minimal toy system that is fully reflective and autonomous from the start, sized to prove the point rather than to be useful — the trade Commonplace itself declines by staying practical.

(Scaling-law models offer one illustration of an explicit retained model guiding decisions without exhaustive search — Hoffmann et al. 2022 — though their payoff is fewer exploratory experiments and better compute allocation, not necessarily fewer target examples, which is why they motivate the conjecture rather than evidence it.) Where practical, encode equivalent prior structure in an explicit, addressable form and in a parametric or distilled form, matching informational content, acquisition data, and compute as closely as possible. Then, over task families with controlled structured shifts, compare:

  • the explicit hypothesis with normal retrieval;
  • the same hypothesis with impaired retrieval;
  • a wrong but highly addressable hypothesis;
  • parametric or meta-learned adaptation;
  • a hybrid using retrieved hypotheses plus parametric adaptation.

Measure the new target observations required to reach a fixed performance level, alongside retrieval success, validation cost, correction cost, compute, and total amortized cost across the family.

The directional prediction, in sample-efficiency terms: for task families sharing a stable mechanism or compositional rule, a matched explicit-hypothesis pathway should reach a fixed target performance with fewer new target observations than a parametric adaptation baseline. That advantage increases with the number of tasks across which the retained structure is reused and decreases with retrieval and applicability errors; it disappears or reverses when the hypothesis is wrong or no relevant stable structure exists.

Open Questions

  • Whether a task family with controlled structured shift can be exhibited where the matched explicit-hypothesis pathway measurably reaches fixed performance on fewer target observations.
  • Whether one can be exhibited where it measurably fails to.
  • Whether hybrid pathways — parametric adaptation guided by retrieved explicit hypotheses — dominate both pure pathways, which would turn the contest into an engineering question about composition.
  • Whether a commitment's scope can be estimated from its addressable form before any shift tests it — whether legibility supplies evidence about reach, or only a handle on reach established some other way.
  • Whether discovery, codification, validation, and maintenance cost for a library of many retained commitments grows faster than the smooth cost curve of parametric scaling as the number of task families and commitments increases.
  • Whether "matching informational content" between an explicit and a parametric arm of the proposed test can be operationalized at all, or whether the comparison is untestable in the form proposed.
  • Whether validating a commitment's reach against an informally specified shift is itself target-data-free, or whether it quietly consumes the same target observations the conjecture claims to save.
  • Whether a full, autonomous, reflective self-improving system exists, or could be built cheaply enough to run this test on at all, given that the two closest real instances — the Gödel machine (unimplemented) and Commonplace (whose pathways split cleanly into reflective-and-autonomous versus non-reflective-and-human-inclusive, with most still the latter) — don't yet supply one.
  • Whether a deliberately minimal toy system (proposed above) can actually stay fully reflective and autonomous end to end, or whether the human ends up back inside the loop at a different point — designing the toy's objective, judging its results — the way it stays inside most of Commonplace's own pathways.

Relevant Notes: