From Reweighting to Rewriting: Unlocking the Intervention Effects of Influential Samples in Training Data Attribution
Yuzhang Luo, Chenpeng Wang, Jianhui Chen, Liangming Pan
cs.CL · cs.AI · cs.LG
Quick verdict
Shows IF-selected SFT examples gain real leverage under response rewriting, not reweighting; well-controlled and convincing, with model-dependent caveats.
Summary
This empirical study asks why influence-function (IF)-selected training examples often show little advantage over random ones under weight-based interventions in LLMs. The authors decouple selection from intervention: EK-FAC influence scores rank SFT examples by their effect on a target behavior (epistemic abstention), and the same selected sets are then treated four ways — deletion, upweighting (α=2), behavior-aligned rewriting, and behavior-opposed rewriting — while OLMo2-1B, Qwen3.5-2B, Gemma3-4B, and OLMo2-7B are retrained from their base checkpoints on a fixed 64,000-example Tülu 3 subset with fixed data order (2.5% intervened, k=1,600). Reweighting produces weak, unstable, sometimes direction-opposed recall changes (Figs. 2-3), whereas rewriting yields strong bidirectional shifts: e.g., OLMo2-1B abstention recall rises from 0.354 to 0.490 with harmful-set aligned rewriting (Table 3). Influence-selected targets beat random and projection-based selectors under rewriting, effects concentrate on target scenarios (answer-unknown, false-premise) without general-capability loss (Table 4), and rankings transfer across models (Fig. 13). Qualitatively similar results hold for safety refusal, but with a marked XSTest over-refusal cost. Limitations include no formal significance testing (95% CIs only on random baselines), model-dependent inconsistencies (Gemma3-4B), and possible surface-form coupling between the abstention-style attribution target and the refusal templates used for rewriting.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Weight-based interventions (upweighting α=2, deletion) on IF-selected SFT examples yield weak, unstable, and sometimes direction-opposed abstention changes across four open-weight LLMs (Figs. 2-3).
- Rewriting the responses of the same influence-selected examples produces strong, persistent, bidirectional behavioral shifts that exceed matched-random rewriting (Fig. 2, Table 3).
- Influence rankings identify rewriting targets with greater leverage than random, projection-based, and other TDA selectors, with effects concentrated on target scenarios and no general-capability loss (Figs. 5, 7; Table 4; App. C.2 as summarized).
- Rewriting shifts recomputed influence of the selected examples toward the target direction, with originally harmful examples reversing sign (Fig. 6, App. B).
- The rewriting-over-reweighting contrast extends qualitatively to safety refusal, but with a substantial over-refusal cost on XSTest (safety section, App. G.2).
Strengths
Clean factorial decoupling of selection from intervention
The same EK-FAC-selected sets receive four interventions (deletion, upweighting α=2, aligned/opposed rewriting) with retraining from identical base checkpoints, fixed data order, and a shared recipe (Sec. 3.2.2, App. H, Table 6), which is exactly the isolation the research question demands.
Bidirectional validation plus dose-response
Opposed rewriting produces opposite-signed shifts, and the k∈{800, 1,600, 3,200} ablation (Fig. 11) shows rewriting scales monotonically with budget while reweighting shows no consistent dose-response (Fig. 3) — strengthening the causal reading of the contrast.
Methodologically careful influence recomputation
Specificity and side-effect checks
Honest negative and trade-off reporting
Weaknesses
Possible stylistic circularity between attribution target and rewrite
Thin statistical granularity for the main contrasts
Consistency claim vs. model-dependent exceptions
Over-refusal cost and limited utility evaluation
Intervention writes boilerplate supervision
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv 2609.02771) addresses a concrete tension in training data attribution (TDA): influence functions (IF) rank SFT examples by predicted effect on a target quantity, yet recent work (Li et al., 2025; Lee et al., 2026a) reports IF-selected examples barely beat random selection under weight-based interventions. The authors split that negative result into two questions — *which* examples to intervene on vs. *how* to intervene — and propose influence-guided response rewriting: EK-FAC influence selects targets, and their responses are replaced with behavior-aligned or behavior-opposed text, instructions fixed. Central claims: (1) reweighting influence-selected examples yields weak, unstable, sometimes reversed abstention effects across four open-weight LLMs (Figs. 2-3); (2) rewriting the same examples produces strong, persistent, bidirectional shifts exceeding matched-random rewriting (Fig. 2, Table 3); (3) influence selection beats projection-based and other TDA selectors under rewriting, with target-concentrated effects and no general-capability loss (Figs. 5, 7; Table 4; App. C.2 as summarized in visible text); (4) recomputed influence shifts after rewriting, with harmful-set examples reversing sign (Fig. 6); (5) the contrast extends qualitatively to safety refusal with a marked over-refusal cost.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does this paper claim about influence functions for LLM training data?
It argues that influence-function rankings do identify training examples with real behavioral leverage, but weight-based operations (upweighting, deletion) fail to realize it; rewriting the responses of the same selected examples produces large, bidirectional behavior changes in abstention and safety refusal across four open-weight LLMs (OLMo2-1B, Qwen3.5-2B, Gemma3-4B, OLMo2-7B).
What is influence-guided response rewriting?
A two-step fine-tuning data intervention: influence functions (EK-FAC) rank SFT examples by their estimated effect on a target behavior, and the responses of the top- and bottom-ranked examples are replaced with behavior-aligned (e.g., abstaining) or behavior-opposed (e.g., compliant) text while the instructions stay unchanged.
Did rewriting work better than reweighting in this study?
Yes, by a wide margin: on OLMo2-1B, aligned rewriting of influence-selected harmful examples raised abstention recall from 0.354 to 0.490, while upweighting and deletion of the same examples produced small, unstable, sometimes opposite-direction changes (Figs. 2-3, Table 3); rewriting effects also scaled monotonically with the number of modified examples while reweighting did not.
What are the main caveats of the paper?
Results vary by model (Gemma3-4B is an outlier with weak separation from random), gains come with precision loss and a clear over-refusal cost on XSTest in the safety setting, no formal significance tests are reported (95% CIs only over random-selection seeds), and both the attribution target and the rewrite templates use a similar generic abstention style, which could partly inflate the selection advantage.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Provided text is a sampled excerpt with three '[...sampling gap...]' breaks: parts of Sec. 3.2.2 (intervention design), the start of the safety section (Sec. 6), parts of App. C.2/G.1, and some reference pages were skipped; judgments on those rely on surrounding visible summaries, not the elided text.
- Figure 4's per-group means (0.985 / 1.131 / 0.534 / 2.301) extracted with ambiguous column alignment, so they are not cited in this review.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.