From Reweighting to Rewriting: Unlocking the Intervention Effects of Influential Samples in Training Data Attribution
Summary
This paper introduces influence-guided response rewriting to intervene on influential training examples, showing that rewriting responses creates stronger and more persistent behavioral shifts in language models than conventional reweighting.
View Cached Full Text
Cached at: 09/10/26, 02:16 PM
Paper page - From Reweighting to Rewriting: Unlocking the Intervention Effects of Influential Samples in Training Data Attribution
Source: https://huggingface.co/papers/2609.02771
Abstract
Influence-guided response rewriting of selected training examples produces stronger and more persistent behavioral shifts in language models than conventional reweighting, highlighting the broader intervention leverage of influential data.
Training data attribution(TDA) aims to identify training examples that shape model behavior, but its intervention value depends on both which examples are selected and how they are modified.Influence functions(IF) estimate behavioral changes under infinitesimalreweighting, yet IF-selected examples often show limited advantages over random selection under conventional weight-based interventions. This raises the question of whether influential examples lack intervention value or whetherreweightingfails to realize their behavioral leverage.We introduce influence-guidedresponse rewriting, which uses IF to identify intervention targets and replaces their responses with behavior-aligned or behavior-opposed supervision while keeping instructions fixed. Across fouropen-weight LLMs, we compare rewriting andreweightingon the same influence-selected examples usingepistemic abstentionas our primary testbed.Response rewritingproduces stronger, more persistent, and bidirectional behavioral shifts, whilereweightingthe same examples yields weak and inconsistent effects. Further analyses show that influence-selected examples provide greater rewriting leverage than alternative selectors, with changes remaining concentrated on target-relevant behaviors. The same qualitative contrast extends to safety refusal. These results distinguish the localreweightingeffects captured by influence estimates from the broader intervention leverage of the examples they identify, motivatingintervention-aware evaluationof TDA methods.
View arXiv pageView PDFAdd to collection
Get this paper in your agent:
hf papers read 2609\.02771
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2609.02771 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2609.02771 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2609.02771 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
Does Fine-Tuning Undo Activation Steering? Behavioural Recovery Without Weight-Edit Reversal
This paper investigates whether fine-tuning undoes activation steering in language models, finding that while behavioral effects like refusal suppression can degrade under optimization pressure, the underlying weight modifications remain mechanistically stable.
Resisting Harmful Instructions through Training-Free Weight Editing
该论文提出一种训练无关的权重编辑方法，通过少量对比提示对直接修改模型权重，降低指令微调模型对有害指令（如说谎、性别歧视）的服从率，并在五个模型上验证了其优于激活引导等方法的鲁棒性。
STARE: Surprisal-Guided Token-Level Advantage Reweighting for Policy Entropy Stability
STARE addresses policy entropy collapse in GRPO-based reinforcement learning for large language models by introducing surprisal-guided token-level advantage reweighting and target-entropy regulation, achieving 4%-8% accuracy gains on AIME benchmarks.
Rethinking Token Reweighting for SFT: Suppress, Reverse, and Extrapolate Learned Features
This paper proposes SCALE (Selective Control of Adaptation via Local Entropy), an entropy-guided method that freezes the pretrained model and SFT delta while learning bounded token- and module-specific gates to suppress, reverse, or extrapolate learned features during supervised fine-tuning. The approach achieves strong math-reasoning and code-generation results on Qwen2.5-Math and Qwen3-4B-Base models, outperforming existing token-reweighting baselines.
Correcting Influence: Unboxing LLM Outputs with Orthogonal Latent Spaces
This paper introduces a framework for token-level influence attribution in large language models by learning orthogonal latent spaces with sparse autoencoders, enabling precise identification of training data tokens that jointly influence predictions, with applications in high-stakes domains like healthcare.