If the goal you're training an AI toward shifts mid-run, what actually ensures earlier progress doesn't silently vanish?
How do epoch boundaries preserve self-improvement guarantees across objective changes?
This explores how a self-improving system can keep its improvement guarantees intact when the target it's optimizing for changes partway through — the corpus doesn't treat 'epoch boundaries' as a named mechanism, but it says a lot about what actually anchors a guarantee when the objective moves.
This reads as a question about keeping self-improvement honest across a moving target: if the objective itself is allowed to change (a new evaluator, a new reward, a new goal), what stops each change from quietly resetting or faking the progress you thought you'd banked? The corpus has no note on literal 'epoch boundaries,' but it converges hard on the underlying problem — and the answer is less about the boundary than about what you carry across it.
The deepest constraint is that self-improvement is bounded by the gap between generating a solution and verifying it What limits how much models can improve themselves?. A model can only reliably improve on tasks where it judges better than it produces; on factual tasks that gap vanishes, and pure self-improvement stalls into circularity, diversity collapse, and reward hacking Can models reliably improve themselves without external feedback?. So a guarantee is never really a property of the system alone — it's smuggled in from something external: a past model version, a third-party judge, a user correction, tool feedback What actually constrains large language models from self-improvement? What stops large language models from improving themselves?. That reframes your question: what preserves a guarantee across an objective change is that the anchor survives the change. If the boundary keeps a frozen external reference around, the new objective can't launder progress past it.
The Darwin Gödel Machine makes this concrete. It swaps formal proof for empirical benchmarking and keeps an evolutionary archive of every agent variant Can AI systems improve themselves through trial and error?. The archive is the boundary mechanism: each generation is re-validated against retained ancestors on a fixed benchmark, so a change that looks better but regresses on the old measure is caught rather than crowned. Compositional skill libraries do the same trick from the other direction — externalizing skills as retrievable, executable units means new learning composes on top of old capability instead of overwriting it, which is exactly how weight-update methods leak guarantees through catastrophic forgetting Can agents learn new skills without forgetting old ones?.
The hardest case is when the objective change is the whole point — when you deliberately co-evolve the evaluator alongside the agent. The Red Queen Gödel Machine does this to reach ungradable creative tasks that no static verifier can score, and it matches fixed-evaluator performance with fewer tokens Can evaluators improve alongside the agents they score?. But a moving evaluator is precisely where guarantees are most fragile: if the evaluator can drift to flatter the agent, you've rebuilt the circularity the mirage note warns about. What keeps it honest is a scoring rule that can't be gamed by the shift — binary correctness rewards, for instance, provably degrade calibration because they don't punish confident wrong answers, whereas adding a proper scoring rule (Brier) mathematically pins accuracy and calibration together so no reparameterization of the objective can trade one off against the other Does binary reward training hurt model calibration?.
The thing you didn't know you wanted to know: the guarantee doesn't live at the boundary — it lives in what refuses to move. Whether it's a retained ancestor in an archive, an externalized skill you can still call, or a scoring rule that's invariant to how you redraw the objective, self-improvement across changing goals is preserved exactly to the extent that something is held fixed for the new objective to be checked against. And there's a sharp practical corollary from long-horizon work: the systems that actually realize these guarantees aren't the ones that redraw objectives most cleverly — they're the ones that persist through the re-validation loop instead of terminating early or burning the budget What predicts success in ultra-long-horizon agent tasks?.
Sources 9 notes
Models can only improve themselves when they verify solutions better than they generate them. This gap scales with model size but vanishes entirely for factual tasks, predicting which domains benefit from self-improvement.
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.
LLMs cannot reliably improve themselves without external verification; metacognition must be externalized rather than learned. Alignment philosophy is shifting from preferentism to normative standards, but coherent values at scale include problematic self-valuation requiring utility engineering beyond output control.
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.
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.
Show all 9 sources
VOYAGER demonstrates that storing executable skills in an embedding-indexed library and composing complex skills from simpler ones allows agents to learn continuously while avoiding the forgetting that occurs with weight-update-based methods. Environmental feedback refines skills while an automatic curriculum drives continual exploration.
Red Queen Gödel Machine makes evaluation part of the improvement loop, allowing agents to optimize writing and proof generation without a static verifier. Co-evolved systems match fixed-evaluator performance while using fewer tokens, suggesting shared learning drives efficiency.
Binary correctness rewards incentivize high-confidence guessing because they don't penalize confident wrong answers. Adding the Brier score as a second reward term mathematically guarantees joint optimization of accuracy and calibration without trade-off.
Across 17 frontier models on 36 expert-curated optimization tasks, repeated benchmark-edit-incorporate cycles within a wall-clock budget proved the dominant success predictor. Most models terminated early or burned budget unproductively; Claude Opus 4.6 stood out as persistent.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Self-Improvements in Modern Agentic Systems: A Survey
- Mind the Gap: Examining the Self-Improvement Capabilities of Large Language Models
- Self-Improving Model Steering
- Hyperagents
- The Red Queen Gödel Machine: Co-Evolving Agents and Their Evaluators
- Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents
- Beyond Accuracy: The Role of Calibration in Self-Improving Large Language Models
- Continual Learning Bench: Evaluating Frontier AI Systems in Real-World Stateful Environments