acceptodds
Withdrawn from ICLR 2027

Data-Weight Directions Rotate Along Training: Horizon-Matched Attribution for GUI Trajectories

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

When attribution scores determine training-data weights, their effects depend on how long training continues. We ask whether weights selected after \(K\) updates remain beneficial after \(T>K\) updates. We formulate attribution as a derivative through a finite training run. Our method, horizon-matched attribution, evaluates candidate weights at the intended training endpoint. We compare short- and terminal-horizon choices under the same training recipe, using paired weight changes that keep the total weight unchanged. Across six Qwen3-VL Mind2Web training subsets, cosine similarity between the short- and terminal-horizon directions is always negative. Weights selected at 88 updates lower loss on the dataset used for selection on all six subsets but raise it at 701 updates. Terminal selection lowers held-out offline expert-action cross entropy on the task, website, and domain panels for every subset. On three subsets, the failure persists across several early selection horizons, while later selection recovers positive mean terminal gains. Fixed choices reproduce the reversal under two additional optimizer seeds on those subsets. InternVL3.5-2B also favors terminal selection, while controlled Android experiments separate sign correction from combining directions. Matching the measurement horizon to the training endpoint corrects short-horizon weighting failures in these GUI fine-tuning settings. Code and results are available at https://anonymous.4open.science/r/horizon-matched-attribution-3C50/.

Primary area:
optimization
Keywords:
Data reweighting, training dynamics, GUI agents, vision-language models.

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_uzTAOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies whether data weights that help early in training still help at the end. The authors keep the training setup fixed and compare weights selected at early and final checkpoints. On Mind2Web, the early choices can actually make the final loss worse, while terminal selection improves held-out loss. They also check the result with new training seeds, another model, and some Android experiments.

Critical strengths

The main strength is the controlled design. Only the data weights change, and the comparison uses actual training outcomes, so the reversal is unlikely to be an artifact of a bad attribution estimate. The reversal also survives the new-seed check with the same weights, which suggests it is not tied to the seed used for selection.

Critical weaknesses
  1. I am not convinced that the cosine results establish local direction rotation. Eq. (5) is about very small perturbations, but every experiment uses h = 0.1. Adding more directions at the same radius doesn’t address this.

  2. The sign of the Android “Raw local” baseline is unclear. IIn App. A.2, positive utility means a trajectory helps, but a = −ZᵀDqv seems to point the other way. If Raw local uses +ρu₁, the sign correction may just be undoing a flipped baseline. This only matters for Table 4.

  3. The novelty of this paper is low and i did not learn somethong new from this paper.

Questions for authors
  1. Could you rerun one existing Mind2Web support with h = 0.05 and check whether the negative cosine and the early-selection reversal still hold?

  2. Is Raw local +ρu₁ or −ρu₁, and how does that sign fit the utility definition in App. A.2?

  3. Keep consistent wordings throughout the whole paper, and some terms and sentence expressions are used in the paper inconsistently.

If the author could solve these questions, i will improve the score.

Reviewer_gpYdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies horizon-matched attribution: choosing data weights based on their effect at the intended training endpoint. It compares short selection and terminal selection using the same candidate weights and training settings. On six Qwen3-VL Mind2Web training subsets, weights chosen at 88 updates reduce early loss but increase loss at 701 updates. Terminal selection reduces loss on held-out data. The paper also tests other selection horizons, new optimizer seeds, InternVL3.5-2B, and AndroidControl.

Critical strengths
  1. The experiment tests the effect of the selection horizon directly. The choices use actual measured losses, rather than an estimate of early loss from an attribution method. Table 10 shows early gains followed by final losses on all six Qwen training subsets. Table 11 shows lower held-out loss after terminal selection. Uniform weighting is not one of the candidates, so beating it is not guaranteed by the selection rule.
  2. The extra tests address useful questions. The failure occurs at several early horizons, not only at 88 updates. Extending from four to six directions keeps the same selected weights. The selected weights also show the reversal with two new optimizer seeds, without choosing them again. The paper reports mixed results as well, including confidence intervals that include zero and an Android case where short projection remains competitive.
Critical weaknesses
  1. The new idea is limited. Earlier work already studies short-horizon bias and changes in data influence during training. Equation (5) is a first-order expansion. Terminal selection trains every candidate to the endpoint and chooses the one with the lowest selection-query loss. Its main contribution is the measured reversal and the result on held-out data, rather than a new efficient selection method. Appendix E reports 5,608 updates for terminal selection and 1,317 for short selection.
  2. The experiments cover a narrow setting. The main training supports contain 25 trajectories and use rank-2 LoRA, one measurement radius, and one direction construction. Some supports overlap and share evaluation data. InternVL also uses a Qwen language backbone. Testing six directions is useful, but it does not test other radii or feature sets. The gains are in offline expert-action cross entropy, which measures loss on recorded expert actions. The paper does not yet show better generated actions or higher task success.
Questions for authors
  1. Clarification: Appendix A.2 defines positive checkpoint utility as helpful but sets the first axis to a=−Z⊤Dqva=-Z^\top D_qv. Which sign does “Raw local” use in Table 4? Both signs are tested in the main candidate set, but this detail matters for the sign-correction result.
  2. Clarification or additional validation: please give the absolute losses under uniform weighting for the main Mind2Web tests. Is any action-accuracy result available for the selected models? This would help explain what the loss changes mean in practice.
Reviewer_c1yEOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies whether data weights selected to improve an early training checkpoint remain beneficial after longer training. Using controlled GUI fine-tuning experiments, it finds that fixed trajectory weights selected from early query losses can worsen terminal performance. The proposed “horizon-matched attribution” evaluates candidate weightings at the intended training endpoint and selects the best candidate using query loss. Experiments on Mind2Web show consistent improvements in held-out expert-action cross entropy, with supporting checks across training seeds, selection horizons, and a second multimodal model.

Critical strengths
  1. The paper shows that weights with genuinely beneficial early effects can become harmful after continued training. Because selection uses actual measured candidate losses, this finding separates horizon mismatch from errors in estimating influence.
  2. Carefully controlled experimental comparisons. Early and terminal selection use the same candidate weights, initialization, training schedule, token budget, and selection query. These controls make the observed reversal interpretable and provide a sound protocol for evaluating whether weighting decisions transfer across training horizons.
Critical weaknesses
  1. Methods such as DVEmb (Data value embedding, Wang et al.) already account for the dependence of influence on the optimization trajectory and the iteration at which a sample contributes. This paper’s distinct question is whether a persistent weighting selected using early performance remains useful at a later endpoint. That is a worthwhile empirical distinction, but the methodological advance remains unclear: the main contribution appears to be a controlled demonstration of weighting reversal. The paper should more clearly establish what it adds beyond existing trajectory-aware attribution and short-horizon bias analyses, including comparisons with appropriately adapted trajectory-based baselines.
  2. All experiments concern supervised GUI trajectory fine-tuning on Mind2Web and AndroidControl. The main experiments further use very small training subsets and rank-2 LoRA, leaving unclear whether the observed horizon-dependent reversals are common in data reweighting or are particularly pronounced in this restricted regime. Given that the paper’s main contribution is an empirical observation, evaluation on non-GUI tasks—such as image classification, language modeling, or general instruction tuning—and across training scales would substantially strengthen its significance. The current evidence establishes a failure mode in a narrow setting but provides limited guidance about when it should be expected more broadly.
Questions for authors

See weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.