- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper argues that disagreement between data attribution methods can come from asking different counterfactual questions, rather than simply estimating the same quantity with different levels of accuracy. It describes an influence specification through three components: the behavior being measured, the intervention on a training example, and the counterfactual training process. It also presents a local decomposition that separates the behavior signal, training signal, and parameter response. The experiments first study specification changes and approximation error in controlled settings, then examine noisy-label detection and LLM attribution. The main practical finding is that the best behavior surrogate depends on both the attribution method and the task. The paper supports this finding with experiments across several architectures and LLMs, including removal-and-retraining tests.
Critical strengths
The experiments go beyond showing that different methods produce different rankings. The controlled comparisons separate changes in the target quantity from errors in approximating a fixed target. The practical experiments then show why this distinction matters. For example, on CIFAR-10 at 20% label noise, changing from trusted loss to target logit improves TracIn’s AUPRC from 0.349 to 0.602, but reduces LiSSA’s AUPRC from 0.535 to 0.421. This is a useful result because it challenges the idea that there is one behavior surrogate that should work best across estimators.
The removal-and-retraining results give the practical claims more weight. In Table 17, removing the top 10% of examples selected by TracIn with answer logit reduces Qwen3-8B’s backdoor attack success rate from 0.809 to 0.053, close to 0.050 after oracle poison removal, with little change in clean accuracy. The benign-trigger controls also make this more informative than simply detecting a shared trigger string. I appreciate that the appendix includes additional LLMs and signal-augmented similarity baselines, which help qualify the main comparisons.
Critical weaknesses
The distinction between specifications and approximations needs more consistency. Section 4 describes an “exact” comparison involving a damped inverse-Hessian response, but it is unclear whether this represents a derivative, a constructed local update, or a finite re-optimization effect. This distinction is central to the paper. Proposition 1 also needs a scope restriction: its proof applies to deterministic outcomes, while Equation (1) allows stochastic outcomes whose expected rankings need not survive a nonlinear monotone transformation.
The practical comparisons need clearer implementation details and a tighter control. Appendix A.2 does not fully specify the LLM implementations of TRAK and LESS, including checkpoints, normalization, and query aggregation, making it hard to verify what stays fixed across surrogates. An answer-token loss or log-probability baseline would also help: moving from full-response loss to answer logit changes both the token span and the readout, so the current results do not separate their contributions.
Questions for authors
What exactly is the inverse-Hessian estimand in the training-process comparison? Please specify the counterfactual process, the role of damping, and whether the reported quantity is a derivative or a finite behavior change. For the removal-versus-upweighting comparison, please also give the objective normalization and sign alignment explicitly, since the two interventions act in opposite directions.
Can Proposition 1 be restricted to deterministic counterfactuals, or extended under appropriate assumptions for the expected estimand in Equation (1)? The remark about first-order approximations also needs qualification: at a common factual point, a differentiable increasing transformation with positive derivative simply rescales the behavior gradient for a fixed query, so it does not by itself change the first-order ranking.
How were TRAK and LESS adapted to each behavior surrogate? Please specify the training-side quantities, optimizer treatment, normalization, checkpoints, and aggregation over queries. For answer-token readouts, is the preceding rationale teacher-forced or generated? An answer-only negative log-likelihood baseline would help interpret the response-corruption results.
Could you sharpen the distinction from Bae et al. (2022)? That work already argues that influence estimates can approximate a different counterfactual object, the proximal Bregman response function, rather than leave-one-out retraining. Appendix D acknowledges the paper, but describes its contribution too narrowly as an infinitesimal-upweighting versus removal distinction. I see the behavior-surrogate experiments as the clearest additional contribution here, and would encourage the authors to position the paper around that extension. See Bae et al. (2022).
2 comments
Response to Reviewer csRx (2/2)
Q4: For answer-token readouts, is the preceding rationale teacher-forced or generated? An answer-only negative log-likelihood baseline would help interpret the response-corruption results.
The preceding rationale is teacher-forced. Our answer-token readouts are conditioned on the query prompt and the provided reference rationale. In our experiments, "Answer logit" denotes the target answer token's score after log-softmax normalization over the answer candidates. We additionally compare it with answer-only negative log-likelihood (NLL) and raw logits, reporting AUPRC on Qwen3-8B over three seeds.
| Estimator | Answer-token readout | Answer corruption | Rationale corruption |
|---|---|---|---|
| GradSim | Candidate-normalized Answer logit | 0.785 ± 0.024 | 0.248 ± 0.001 |
| GradSim | Negative answer-only NLL | 0.779 ± 0.024 | 0.248 ± 0.001 |
| GradSim | Raw logit | 0.401 ± 0.010 | 0.269 ± 0.012 |
| TracIn | Candidate-normalized Answer logit | 0.643 ± 0.056 | 0.270 ± 0.010 |
| TracIn | Negative answer-only NLL | 0.636 ± 0.050 | 0.273 ± 0.010 |
| TracIn | Raw logit | 0.281 ± 0.035 | 0.370 ± 0.022 |
| TRAK | Candidate-normalized Answer logit | 0.941 ± 0.006 | 0.169 ± 0.006 |
| TRAK | Negative answer-only NLL | 0.940 ± 0.006 | 0.170 ± 0.006 |
| TRAK | Raw logit | 0.528 ± 0.005 | 0.166 ± 0.015 |
| LESS | Candidate-normalized Answer logit | 0.877 ± 0.020 | 0.198 ± 0.008 |
| LESS | Negative answer-only NLL | 0.875 ± 0.021 | 0.199 ± 0.009 |
| LESS | Raw logit | 0.596 ± 0.051 | 0.222 ± 0.022 |
Candidate-normalized answer logits and negative answer-only NLL yield nearly identical performance, while raw logits substantially reduce answer-corruption detection, further demonstrating the impact of behavior surrogate choice on attribution performance.
Q5: Could you sharpen the distinction from Bae et al. (2022)?
Thank you for highlighting this connection. Bae et al. (2022) distinguish practical influence-function estimates from LOO retraining and introduce the proximal Bregman response function (PBRF) as an alternative response that these estimates approximate. Our work extends this perspective across similarity-based, one-step, trajectory-based, and influence-function estimators through the specification and a decomposition exposing the roles of behavior and training process. We further identify the often-overlooked role of behavior surrogates, showing that surrogate choice can reverse estimator comparisons and leave problematic training data largely undetected. We have revised the appendix to clarify this connection and these distinctions.
Response to Reviewer csRx (1/2)
Thank you for your thoughtful review and constructive suggestions. We appreciate your recognition of our controlled experiments separating specification mismatch from approximation error, and we are glad you found the interactions between estimators and behavior surrogates informative and the removal-and-retraining results supportive of our practical claims. Below are our responses to the issues raised:
Thank you for pointing out this ambiguity. The inverse-Hessian quantity is the parameter response to infinitesimal upweighting at the converged checkpoint, paired with the behavior gradient to obtain the local behavior derivative. It is not a finite behavior change. Without damping, the parameter response is
−HD−1∇θℓ(zk;θD).
The training process comparison uses damping 0.01 for numerical stability, replacing HD with HD+0.01I. Accordingly, exact computation here refers to the specified damped response. The factual solution is refined in double precision until its gradient norm falls below 10−12.
For objective normalization, we use
LD=n1∑iℓi+2λ∥W∥F2,Lup,k=LD+αℓk,L−k=n−11∑i=kℓi+2λ∥W∥F2.
with unchanged regularization and unregularized biases. To align signs, we use negative cross-entropy for the loss behavior, and compute upweighting scores as Bup−Bfactual and removal scores as Bfactual−Bremoved. Positive scores therefore indicate beneficial contributions to the selected behavior.
Thank you for identifying this scope issue. We have restricted Proposition 1 to deterministic counterfactuals and behavior evaluations, and clarified that nonlinear monotone transformations need not preserve rankings after taking expectations. We also agree that, for a fixed query at a common factual model, a transformation with positive derivative rescales all first-order scores by the same positive factor and preserves their ranking. We have corrected the accompanying remark accordingly.
Thank you for requesting these implementation details. We adapt TRAK and LESS by replacing the query-side gradient with the gradient of each behavior surrogate, while keeping the training-side quantity fixed as the gradient of the mean full-response negative log-likelihood. LESS additionally uses the saved Adam second-moment estimates to precondition both training and query gradients. For consistent comparison across behavior surrogates, we do not apply additional unit-norm normalization to either side. Both methods use the final training checkpoint, and attribution scores are averaged equally over all the queries.