Instead of patching each mistake one by one, why does sorting AI feedback by the type of fix help prompts handle new cases?
Why does grouping feedback by shared correction pattern improve prompt generalization?
This explores why an optimizer that sorts critiques of a prompt's failures into bundles by the kind of fix they call for, instead of patching each failure separately, ends up with prompts that work on new inputs. The corpus has no note that tests this exact mechanism, so the explanation below is inferred from neighbouring findings about learning from mistakes.
This explores why an optimizer that sorts critiques of a prompt's failures into bundles by the kind of fix they call for, instead of patching each failure separately, ends up with prompts that work on new inputs. The corpus has no note that tests this exact mechanism, so what follows is inferred from neighbouring findings about learning from mistakes.
The clearest thread is that a shared fix is a rule, while a single fix is a patch. LEAP shows that models improve when they deliberately err on a few examples, reflect, and write down explicit task-specific principles Does learning from mistakes improve in-context learning?. A principle is what you get when several mistakes are read as one. Training on critiques of flawed answers also builds deeper understanding than imitating correct ones, because critique forces the model to engage with the failure mode rather than surface patterns Does critiquing errors teach deeper understanding than imitating correct answers?. Grouping by correction pattern applies the same logic to feedback. The unit becomes the type of failure, not the particular input that exposed it, so the prompt edit is stated at a level that transfers.
Grouping also works as a noise filter. In training, when the signal from each input is weak or idiosyncratic, task gradients fade and regularization pulls the policy toward generic, input-agnostic templates Why do language models collapse into generic templates?. That result is about RL, not prompts, but the shape carries over. A prompt rewritten in response to every stray complaint tends toward vague boilerplate, while a correction that recurs across many inputs is the signal worth acting on. Keeping clusters separate also stops distinct failure types from being averaged into one bland instruction. That echoes how step-level critique preserves solution diversity and prevents narrowing during self-training Do critique models improve diversity during training itself?.
Generalizing also means ignoring what shouldn't matter. Consistency training exists because models react to irrelevant changes in wording, and it teaches them to respond identically to clean and perturbed prompts Can models learn to ignore irrelevant prompt changes?. A correction shared across inputs that look different on the surface is evidence of what stays constant underneath, and that is what a prompt should encode.
There are two limits. Prompt optimization can only reorganize knowledge the model already has Can prompt optimization teach models knowledge they lack?. Grouped feedback helps when failures come from the model not applying what it knows, not from gaps in what it knows. And a recurring pattern isn't automatically a true one. In one production case, a prompt mutation lifted a judge-alignment pass rate from 23.1% to 80.0% by adopting the judge's preferred vocabulary, while defect-identification precision stayed flat Can prompt optimization accidentally teach judges to reward the wrong signals?. If all the feedback comes from one judge, the biggest cluster may be a shared bias, so grouped fixes need checking against an independent measure. Prompts also don't move cleanly across model tiers, since techniques that help cheap models can hurt strong ones Do prompt techniques work the same across all LLM tiers?.
Sources 8 notes
LEAP demonstrates that models achieve better performance on reasoning and math tasks by intentionally erring on few-shot examples, reflecting on mistakes, and deriving explicit task-specific principles—without additional labeled data or fine-tuning.
Training models to critique noisy responses outperforms training on correct answers because critique forces engagement with failure modes and structural reasoning. Even imperfect critique supervision beats correct-answer imitation, showing how weak surface-pattern learning is for building genuine understanding.
When within-prompt reward variance is low, task gradients weaken and regularization dominates, pushing policies toward generic outputs. SNR-Aware Filtering—selecting high-variance prompts before updates—recovers performance across tasks and scales.
Step-level critique in the training loop counteracts tail narrowing and maintains solution diversity across self-training iterations. This training-time benefit—preventing premature convergence—is more fundamental than test-time accuracy gains.
Two methods—BCT (output-level) and ACT (activation-level)—train models to respond identically to clean and wrapped prompts by using the model's own clean responses as targets, eliminating specification and capability staleness inherent in standard SFT.
Show all 8 sources
Prompting works entirely within a model's pre-existing training distribution and cannot supply domain knowledge absent from training data. This creates a hard ceiling: no prompt strategy can compensate for missing foundational knowledge, only reorganize what already exists.
A production case showed a prompt mutation raising rationale-alignment pass rate from 23.1% to 80.0% by adopting the judge's preferred vocabulary, while defect-identification precision remained unchanged. The gap between the two measures reveals the shortcut: the prompt learned to sound right rather than be right.
A 23-prompt benchmark across 12 LLMs shows rephrasing and background-knowledge prompts boost cheap models, while step-by-step reasoning reduces accuracy in high-performance models. Task structure, not generic best practices, determines which prompts help.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Invalid Logic, Equivalent Gains: The Bizarreness of Reasoning in Language Model Prompting
- Large Language Models Are Human-level Prompt Engineers
- Critique-GRPO: Advancing LLM Reasoning with Natural Language and Numerical Feedback
- Critique Fine-Tuning: Learning to Critique is More Effective than Learning to Imitate
- A Survey on Prompt Tuning
- Are Emergent Abilities in Large Language Models just In-Context Learning?
- Learning To Retrieve Prompts for In-Context Learning
- Enhancing LLM Reasoning via Critique Models with Test-Time and Training-Time Supervision