When someone rejects an AI's suggested change, can that 'no' teach it more than just throwing the attempt away?

This explores whether a rejected edit can be used as a training signal, the way contrastive learning uses hard negatives (near-misses that look plausible but are wrong). The corpus has no note on edit rejection or on contrastive loss itself, so it can't confirm the exact analogy. It does have a lot on the underlying bet, which is that a "no" carries more information than most training pipelines take from it.

The strongest evidence is that negatives alone can do most of the work. Training only on wrong samples matches or beats full PPO and GRPO. It suppresses incorrect trajectories while keeping the model's range of plausible answers open, whereas positive-only training piles probability onto a few answers and hurts performance at higher k (Does negative reinforcement alone outperform full reinforcement learning?). That is close to what a hard negative does. It pushes down the specific wrong-but-tempting option without dictating a single right one. The caveat is that this result is about reinforcement learning on sampled answers, not embedding-space contrastive loss, so it supports the intuition rather than proving it.

A second cluster explains why the rejected item should be used and not just the accepted one. Supervised learning rewards the right response but never penalizes a contradiction, which is why persona consistency needed explicit punishment during training (Why does supervised learning fail to enforce persona consistency?). Accepting an edit teaches the model what worked. Rejecting one is the only place it learns what to avoid. Training a model to critique noisy, flawed responses also produced deeper understanding than imitating correct answers, because it forces engagement with how things fail (Does critiquing errors teach deeper understanding than imitating correct answers?). A rejected edit therefore works best as a negative paired with the reason it was rejected, not as a bare label. Critique inside the training loop also counteracts the narrowing of solution diversity across iterations (Do critique models improve diversity during training itself?).

The corpus also shows you can use a rejection without gradient updates or a contrastive loss. LLMs can rewrite a negative reaction like "doesn't look good for a date" into a positive preference like "prefer more romantic," and retrieval improves with no fine-tuning (Can language models bridge the gap between critique and preference?). Agents can turn plain success/failure feedback into written self-diagnoses kept in memory and improve across attempts without changing weights (Can agents learn from failure without updating their weights?). Deliberately making a model err on its few-shot examples, then distilling the mistakes into explicit principles, also improves in-context learning (Does learning from mistakes improve in-context learning?). In each case the rejection is converted into a positive preference, a stored lesson, or a rule, instead of being fed in as a raw negative.

The one part the corpus can't answer is what makes a hard negative "hard", meaning a rejected edit that is close to the accepted one and so teaches the fine distinction. Nothing here tests whether near-miss rejections beat random ones. That is the open question to chase.