- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
Summary:
The paper asks why training examples selected by influence functions (IF) often perform no better than random examples when they are upweighted or deleted during fine-tuning. Is IF picking the wrong examples, or does changing their weight fail to bring out what they could teach the model? To separate the two, the authors keep the IF selection fixed and vary only the intervention. They introduce influence-guided response rewriting: the instruction of each selected example is kept, and its response is replaced with a template that either supports the target behavior (aligned) or opposes it (opposed).
Contributions:
-
Influence-guided response rewriting. The paper separates choosing training examples from modifying them: influence functions identify which SFT examples to target, and replacing their responses determines what those examples teach the model. The authors also propose intervention leverage to measure how well a selection method and an intervention work together.
-
Controlled retraining experiments on four models. The authors retrain OLMo2-1B, Qwen3.5-2B, Gemma3-4B and OLMo2-7B under identical settings, varying only the intervention applied to the same IF-selected examples. They report that rewriting produces strong, lasting effects in both directions, while upweighting and deletion do not.
-
Further analyses. The authors compare IF with other ways of selecting examples (a probe for unanswerability, TRAK, gradient and embedding similarity, and others), examine how rewriting changes the influence scores of the selected examples, check whether gains stay on the targeted types of abstention rather than increasing refusal overall, and extend the study to safety refusal on OLMo2-7B.
Validation:
The authors retrain four models (OLMo2-1B, Qwen3.5-2B, Gemma3-4B, OLMo2-7B) on a 64k Tulu 3 SFT subset, applying each intervention separately to the top 2.5% and the bottom 2.5% of IF-ranked examples (1,600 each), and compare against an unmodified baseline and randomly selected examples. They measure abstention recall on AbstentionBench across training checkpoints, and report that rewriting beats matched random selection in both directions on all four models, while upweighting and deletion are weak and inconsistent, sometimes moving the behavior the wrong way. Additional experiments cover other selection methods, scenario-level results, and safety refusal on OLMo2-7B.
Critical strengths
A clear research question. The paper asks whether the limited gains from upweighting or deleting influence-selected examples come from poor selection or from reweighting not being able to change what those examples teach the model. Separating those two is a useful way to think about data attribution, and the question is worth asking even if the motivating claim rests on thin evidence (see Questions). The authors also released their code, which let me check the implementation.
Critical weaknesses
Weakness 1: The upweighting in the code does not match the paper. Section 3.2 (lines 183–184) defines upweighting as multiplying a selected example’s loss by . However, the released code (patches/open-instruct_local_changes.patch, lines 50–52) also divides by the weighted token count, so partly cancels. With the per-device batch sizes in Table 7 (Appendix H), = 2 gives only about 1.6$\times$ weight for OLMo2-1B, and, if the same loss was used for OLMo2-7B (batch size 1), would have no effect at all there. Since the main claim compares rewriting against reweighting, this understates reweighting in the comparison.
Weakness 2: The paper is inconsistent about which templates were used for behavior-opposed rewriting. Table 10 (page 30) says the abstention-opposed rewrite uses the Abstention-answer pool, but Appendix H (lines 1454–1456) says a comply pool was used, and the released code (build/build_interventions.py, lines 27–29) uses the comply templates, which Table 10 lists for safety. The answer templates are included in the release but never used by default. As a result, it is unclear what supervision produced the weaken results in Fig. 2, 5, 9 and Table 9, and whether those templates are actually the opposite of abstention.
Questions for authors
Q1. The released OLMo2-1B training script (train/run_experiments.sh, lines 14 and 38) uses the weighted loss in patches/open-instruct_local_changes.patch (lines 50–52), which divides by the weighted token count. With a per-device batch size of 4, = 2 then corresponds to only about 1.6$\times$ weight. Was the same loss used for Qwen3.5-2B, Gemma3-4B and OLMo2-7B? For OLMo2-7B (per-device batch size 1, Table 7), would cancel completely. Does the conclusion that reweighting is weak still hold with a loss where directly scales each example’s weight?
Q2. Table 10 says abstention-opposed rewriting uses the Abstention-answer pool (answer_templates_20.jsonl), but this file is never used by default. Instead, the code (build/build_interventions.py, lines 27–29) uses comply_templates_20.jsonl, which Table 10 lists as the Safety-comply pool, and Appendix H also mentions a comply pool. Which pool produced the reported abstention-opposed results, and do the results hold with the other pool?
Q3. IF-selected examples are more often abstention-related than the overall training data (Table 3: about 43–46% vs. 28%), so part of IF’s advantage over random selection could come from topic relevance. Apart from Probing (Fig. 5), the alternative selectors in Appendix C.3, including the topic-matched Judge Matched control, are evaluated only on OLMo2-1B (lines 1077–1079). Could the authors repeat these comparisons on the other three models (Qwen3.5-2B, Gemma3-4B, OLMo2-7B) to show that IF’s selection advantage holds across all four?
Q4. The paper’s motivation (lines 043–046) is that upweighting or removing IF-selected examples (such weight-based interventions) sometimes works no better than random selection. The two cited works give limited support for this. Dai et al. (2025) do not upweight or remove examples; they select a small subset to train on, and their main result is that an improved influence-based method (BIDS) beats random selection. Lee et al. (2026a) is a blog post, not a peer-reviewed paper, with no released code. It tests removal (and one content edit), finding removal no better than random for most behaviors but effective for refusal, which is close to the behaviors studied here. Neither work tests upweighting. Can the authors cite work that actually tests upweighting of IF-selected examples?