# When Edit Localization Amplifies Relative Selection Bias: Gradient Geometry, Target Mismatch, and Importance Weighting

arXiv: https://arxiv.org/abs/2609.13709 · HTML: https://mlsanity.com/paper/2609.13709/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 8 of its 15 references that are in the map.
- First posted on arXiv 12 Sep 2026.

## Builds on (most cited first, 8 of 8)

- [Direct Preference Optimization: Your Language Model is Secretly a Reward Model](https://mlsanity.com/paper/2305.18290/) (2023)
- [MLQE-PE: A Multilingual Quality Estimation and Post-Editing Dataset](https://mlsanity.com/paper/2010.04480/) (2020)
- [Counterfactual Learning from Bandit Feedback under Deterministic Logging: A Case Study in Statistical Machine Translation](https://mlsanity.com/paper/1707.09118/) (2017)
- [Doubly Robust Inference with Non-probability Survey Samples](https://mlsanity.com/paper/1805.06432/) (2018)
- [Enhancing Supervised Learning with Contrastive Markings in Neural Machine Translation Training](https://mlsanity.com/paper/2307.08416/) (2023)
- [Post-edits Are Preferences Too](https://mlsanity.com/paper/2410.02320/) (2024)
- [Fine-Tuning LLMs with Fine-Grained Human Feedback on Text Spans](https://mlsanity.com/paper/2512.23693/) (2025)
- [Correcting Selection Bias in Sparse User Feedback for Large Language Model Quality Estimation: A Multi-Agent Hierarchical Bayesian Approach](https://mlsanity.com/paper/2605.12177/) (2026)

## Abstract (arXiv)

Human corrections identify editable spans, but the examples receiving corrections may come from a selective feedback channel. We analyze this interaction at a fixed model checkpoint by decomposing a localized gradient into edited and retained untouched components. Squared relative selection bias is a ratio of quadratics whose derivative has the sign of an explicit quadratic polynomial. Localization can increase, decrease, or nonmonotonically change this diagnostic; its direction depends on component biases and geometry. Oracle importance weighting recovers the population mean for each fixed localization objective, but these objectives have different targets. Against one common full-gradient target, we derive the finite-sample mean-squared error, an analytic optimal retention coefficient, and a fixed-clipping extension. Exact finite-population calculations and 10,000 Monte Carlo repetitions per sample size verify the identities and counterexamples. Public human-post-edit experiments use two translation directions and pretrained models, with declared synthetic selection. An English-German extension differentiates 73.89 million native parameters. In all three declared settings, hard localization has higher relative bias but lower absolute bias than full retention. Untouched-component biases are nonzero and selected component means have negative inner products, so the general criterion applies where the simple unbiased/aligned explanation fails. Output-bias diagnostics show the same endpoint ordering of relative bias across both directions, with one interior maximum. Mechanisms reuse each language's records and include a mixture; they are not independent replications. The evidence separates relative amplification from absolute gradient error and establishes estimation properties, without inferring translation-quality gains or identifying actual complaint propensities.