How does the expert demonstration ceiling compare to the generation-verification gap bound?
This explores two different formal ceilings on AI learning — one set by the examples a curator demonstrates, the other by a model's ability to verify its own outputs — and asks whether they're describing the same wall from different sides.
This explores two different formal ceilings on AI learning — the expert-demonstration ceiling and the generation-verification gap — and whether they're really the same constraint wearing different clothes. The short version: they are two faces of one limit. Both say that competence is capped by some *external* source of truth, and neither can be escaped by the system thinking harder about itself.
The expert-demonstration ceiling is concrete and curatorial. When an agent trains only on static expert datasets, its competence is bounded by what the dataset's curator imagined — the agent never interacts with an environment, so it can't learn from its own failures or generalize past demonstrated scenarios Can agents learn beyond what their training data shows?. The generation-verification gap is the same wall described formally: a model can generate candidate improvements all day, but it can only *reliably* adopt the ones it can validate, and self-validation is circular. Pure self-improvement stalls without an external anchor — a past model version, a third-party judge, a tool, a human correction Can models reliably improve themselves without external feedback? What stops large language models from improving themselves?. Demonstrations are simply one way to supply that anchor up front; verifiers supply it continuously.
What makes the comparison interesting is that the demonstration ceiling is a *static* version of the gap and the verification gap is the *dynamic* version. A curated dataset is a frozen verifier — it can only certify what was anticipated. A live verifier can certify novel attempts, which is why it raises the ceiling. You can see the difference in how methods escape each wall. RLVR, surprisingly, escapes neither: pass@k analysis shows it narrows sampling toward solutions already inside the base model's distribution rather than expanding the boundary, so a verifiable reward mostly makes the existing ceiling easier to hit, not higher Does RLVR actually expand what models can reason about?. The Darwin Gödel Machine, by contrast, genuinely pushes the ceiling — but only by trading formal proof for *empirical* benchmarking, i.e. importing an external environment as its verifier Can AI systems improve themselves through trial and error?.
The sharpest place the two bounds diverge is verifiability itself. The demonstration ceiling actually becomes *useful* exactly where the verification gap is unbridgeable — in domains with no automated checker. Inverse-RL methods like RARO recover an implicit reward function *from* expert demonstrations, matching verifier-based RL on reasoning tasks while extending to domains that have no verifier at all Can reasoning emerge from expert demonstrations alone?. So demonstrations aren't merely a lower ceiling — they're a substitute anchor for when you can't build the higher one. And generative reward models that reason before judging show the verifier side of the gap is itself improvable with far less data than expected Can generative reasoning beat discriminative models with less training data?.
The thing worth walking away with: neither ceiling is about model capacity. Both are about where the truth signal comes from. Demonstrations cap you at someone's imagination; the verification gap caps you at what you can check. Raise the ceiling and you've done the same move in both cases — smuggled in an external source of correctness. The interesting research frontier isn't escaping the wall through pure introspection (that's provably circular); it's making the cheapest, broadest verifier you can, or, where none exists, recovering one from the demonstrations themselves.
Sources 7 notes
Agents trained on static expert datasets cannot learn from their own failures or generalize beyond demonstrated scenarios because they never interact with environments during training. Competence is capped by what curators imagined, not by agent capacity.
Pure self-improvement stalls due to the generation-verification gap, diversity collapse, and reward hacking. Reliable improvement methods succeed by smuggling in external anchors: past model versions, third-party judges, user corrections, or tool feedback.
Self-improvement in LLMs is formally bounded by the generation-verification gap, meaning every reliable fix requires something external to validate and enforce it. Models cannot escape this constraint through metacognition alone.
Pass@k analysis shows base models outperform RLVR models at high k, indicating RLVR doesn't expand solvable problems but rather narrows sampling toward solutions already in the base model's distribution. Distillation, by contrast, genuinely transfers new reasoning patterns.
DGM replaces formal proofs with empirical benchmarking and maintains an evolutionary archive of agent variants, achieving 2.5× improvement on SWE-bench and 2.2× on Polyglot by discovering capabilities like better code editing and context management.
RARO recovers implicit reward functions from expert demonstrations through adversarial co-training between a reasoning policy and relativistic critic. This approach matches verifier-based RL performance on reasoning tasks while extending to domains lacking automated verification.
GenPRM and ThinkPRM reframe process supervision as generative tasks with CoT reasoning before judgment, achieving superior performance on far fewer labels. A 1.5B GenPRM beats GPT-4o; ThinkPRM uses only 1% of PRM800K labels to surpass full-dataset discriminative verifiers.