Can an AI get better at remembering, at looking things up, and at rewriting itself — or only where it can check its own work?
How does self-improvement capability vary across memory, retrieval, and update tasks?
This explores whether a model's ability to improve itself changes depending on whether the job is recalling facts (memory), finding information (retrieval), or rewriting its own instructions and scaffolding (updates). The corpus has no head-to-head comparison of the three, so what follows is stitched together from adjacent findings.
This explores whether a model's ability to improve itself changes depending on whether the job is recalling facts (memory), finding information (retrieval), or rewriting its own instructions and scaffolding (updates). The corpus has no head-to-head comparison of the three, so this is stitched together from adjacent findings. The common thread is that self-improvement works where checking an answer is easier than producing one. It fails where the model has nothing independent to check against.
Memory is the weakest case. One formal result says self-improvement is capped by the gap between how well a model verifies and how well it generates, and that this gap self-improvement-is-bounded-by-the-generation-verification-gap-a-formal-quantity|vanishes entirely for factual tasks. If you don't know a fact, you can't tell that your recalled version is wrong. A related result is that models trained to imitate ChatGPT model-imitation-captures-style-not-factuality-a-substantial-capability-gap-persi|pick up its confident style but not its factuality, so they look improved while the capability ceiling stays where the base model put it. Wide & Deep is a recommender-systems paper rather than a self-improvement one. Its point still fits: wide-and-deep-models-combine-memorization-and-generalization-through-joint-train|memorizing rare specifics and generalizing across common cases are separate jobs, and each needs its own capacity.
Retrieval matters here as the fix for memory's blind spot. This part is my inference rather than a claim any note makes. One synthesis finds that reliable self-improvement always the-self-improvement-mirage-why-pure-self-improvement-is-circular-and-every-reli|smuggles in an external anchor: past model versions, third-party judges, user corrections, or tool feedback. Looking something up is the obvious way to give a model that outside check on a factual claim. Retrieval improves the answer because it supplies something the model couldn't verify from inside, not because the model got wiser.
Updates are where self-improvement demonstrably works, because the model can score its edits against a task. The survey evidence separates bounded-self-refinement-and-open-ended-recursive-self-improvement-are-different|bounded, evaluable self-refinement (today's industrial practice) from open-ended recursive self-improvement, which stays constrained. Even bounded updates need guardrails. SkillOpt found that bounded-textual-editing-with-rejected-edit-buffers-outperforms-uncontrolled-skil|a budget on edit size, a held-out validation gate, and a buffer of rejected edits beat letting agents rewrite their own instructions freely. Without constraints, harness edits automated-harness-self-improvement-can-overfit-the-training-tasks-constraining-e|memorize the training tasks and lose gains out of distribution. Well-designed selection can transfer: AIDE2's gains aide2s-gains-generalize-to-four-held-out-benchmarks-including-physics-based-weat|held on four unseen benchmarks, including weather forecasting. These notes cover updating skills and scaffolding, not updating stored facts.
Two caveats apply across all three. Reward hacking reward-hacking-can-arise-when-weights-are-updated-when-outputs-are-selected-and|shows up whether you update weights, select outputs, or revise prompts, since the cause is a scoring signal that only partly captures the real task. And today's improvement loops are truly-self-improving-agents-require-intrinsic-metacognition-current-methods-rely|fixed and human-designed, so they break when the domain shifts. One line of work tries to internalize the checking step, training models to post-completion-learning-uses-the-ignored-post-eos-space-to-internalize-self-eva|compute their own reward after finishing an output. That attacks the verification side of the gap directly, though nothing in the notes shows it helps on factual recall.
Sources 11 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.
Imitation models fool human evaluators by mimicking ChatGPT's confident, fluent style while failing to improve factuality or generalization on novel tasks. The ceiling is set by base model capability, not fine-tuning method—better fundamentals, not shortcuts, drive real improvement.
Wide & Deep models train memorization (cross-product features) and generalization (embeddings) together, allowing each component to specialize: the wide part becomes small because deep handles common cases, and deep doesn't overfit rare items because wide captures them. Ensembling requires both halves full-size.
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.
A 1,250-paper survey shows that bounded, evaluable self-refinement (current industrial practice) differs fundamentally from open-ended recursive self-improvement, which remains constrained by grounding requirements, collapse dynamics, and compute limits measurable today.
Show all 11 sources
SkillOpt's ablations show that adding a textual learning-rate budget, held-out validation gate, and rejected-edit buffer (retaining failed edits as negative feedback) produces more stable and generalizable skill improvement than allowing agents to freely rewrite their own instructions.
Recursive edits to agent scaffolding can memorize evolve-set tasks, causing in-distribution gains to shrink out-of-distribution. RRSI addresses this by constraining the proposal and selection stages to favor reusable mechanisms over benchmark-specific ones.
The paper reports that AIDE2's improvements transfer to four held-out benchmarks spanning machine learning, algorithm engineering, and physics-based weather forecasting—the last being outside the selection task distribution. This demonstrates transferable gains beyond overfitting to the selection set.
Reward hacking arises during weight training, output selection, and prompt revision through a shared failure: optimization against signals that incompletely represent the actual task. The substrate matters less than the misalignment between the scoring function and ground truth.
Current self-improvement methods use extrinsic, fixed metacognitive loops designed by humans that fail under domain shift or capability changes. True self-improvement requires agents to generate their own adaptive metacognitive knowledge, planning, and evaluation—a gap confirmed as a neglected research area across neuro-symbolic AI.
Post-Completion Learning exploits unused sequence space after model output to train self-assessment capabilities during training while maintaining zero inference cost. The model learns to compute its own reward functions, internalizing evaluation rather than relying on external reward models.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops
- Hyperagents
- Generalized Agent Iteration: One Formal Framework for Iterative Policy Improvement and Recursive Self-Improvement
- Self-Improvements in Modern Agentic Systems: A Survey
- Recursive self-improvement of AI research agents
- PAST-Bench: Benchmarking the Foundations of Recursive Self-Improvement in Personal Agents
- Mind the Gap: Examining the Self-Improvement Capabilities of Large Language Models
- RRSI: Regularized Recursive Self-Improvement of Agent Harnesses