Can post-hoc reranking actually fix popularity bias created during model training?

This explores whether post-hoc reranking — reshuffling the final list to demote popular items — can undo popularity bias that was created while the model was being trained. The corpus's consistent answer is skeptical: again and again, the bias turns out to be structural, encoded somewhere the reranker can't reach, which means reranking treats a symptom while the cause keeps regenerating it.

The clearest version of this comes from work on embedding size. When user and item embeddings are too low-dimensional, the model overfits toward popular items because that's the cheapest way to maximize ranking quality — and crucially, this 'cannot be fixed post-hoc without treating dimensionality as a fairness hyperparameter' Does embedding dimensionality secretly drive popularity bias in recommenders?. In other words, the bias lives in the geometry of the learned representation, so no amount of reordering the output list recovers the niche items the model never learned to represent well in the first place.

The same 'it's baked in earlier' logic shows up in LLM-based recommenders, just one layer further back. GPT-4 concentrates its picks on items popular in its pretraining corpus rather than in the dataset you actually hand it — The Shawshank Redemption dominates regardless of the target data's real popularity distribution — and this domain-shift effect is something 'standard debiasing methods cannot address' Where does LLM recommendation bias actually come from?. A companion finding pulls this apart further: LLM recommenders carry position bias, popularity bias, and fairness bias all inherited from the pretraining objective and corpus demographics, and mitigation 'requires LLM-specific approaches, not adapted collaborative filtering techniques' Where do recommendation biases come from in language models?. A reranker borrowed from classic CF is exactly the kind of downstream patch these papers say won't transfer.

The deeper reason reranking struggles is feedback. YouTube's multi-objective ranker doesn't try to clean up bias after the fact — it builds selection-bias removal *into training*, using a shallow position tower so the model learns from de-biased signal directly. Without that, 'models converge on degenerate equilibria that amplify their own past decisions' Why do ranking systems need to model selection bias explicitly?. That's the heart of why post-hoc fixes leak: a reranked list still becomes tomorrow's training data, so a model debiased only at the surface relearns the bias on the next cycle. Recommendation feeds compound this at scale — feed weights reshape producer behavior and selection biases contaminate the very ratings that get fed back in How do recommendation feeds shape what people see and believe?.

So the honest reading of the corpus: post-hoc reranking can cosmetically rebalance a single output list, but it doesn't 'fix' training-time popularity bias — it sits downstream of the representation, the pretraining corpus, and the feedback loop that all keep producing it. The methods that actually move the needle intervene where the bias is born: embedding dimensionality as a fairness knob, training-time selection-bias modeling, and pretraining-aware mitigation. The thing you didn't know you wanted to know is that 'where in the pipeline the bias lives' predicts whether any output-stage fix can touch it — and for popularity bias, it almost never lives at the output.