TL;DR
Researchers validated "MAGIC," a trajectory-based data attribution method, by comparing its predictions against the expensive gold standard of actual model retraining. They discovered that while trying to attribute a model's entire training history leads to numerical collapse, focusing on short "checkpoint windows" (e.g., 100 steps) yields remarkably accurate predictions (LDS 0.86–0.99) of how specific data affects learning.
What Problem Does This Paper Address?
In the world of Large Language Models (LLMs), we often ask: Which specific training documents taught the model this skill? This is "Data Attribution." If we can answer this, we can debug models, remove toxic content more effectively, and understand the "curriculum" of machine learning.
The current state-of-the-art method, MAGIC, uses "metagradients" to trace influence back through the optimizer's path. However, there’s a catch: nobody knew if these mathematical scores actually matched reality. The only way to be sure is to remove the data and retrain the model from scratch—a process so computationally expensive it's rarely done at scale. This paper tackles the "validation gap"—testing if trajectory attribution actually predicts the outcome of data removal during the critical moments when a model acquires a new capability (like multi-step arithmetic).
Methods
The authors trained GPT-2 style models (86M and 152M parameters) and applied the following approach:
- MAGIC Attribution: They differentiated the loss of held-out test queries backward through the actual training steps.
- Retraining Validation: To verify the scores, they performed 48 separate retrains, removing 1% of the data at a time to see if the actual change in model performance matched MAGIC’s predictions.
- Windowed Analysis: They compared "full-run" attribution (tracking influence over the entire training history) against "short-window" attribution (tracking influence over just 100 steps from a specific checkpoint).
- Linear Datamodeling Score (LDS): This was their primary metric—a correlation score where 1.0 means the attribution perfectly predicts the retraining outcome.
Datasets
The models were trained on a 50/50 mix of:
- Filtered Web Text: FineWeb-Edu.
- Synthetic Math Families: Four distinct formats of arithmetic (Worked examples, Primitive facts, Verbal procedures, and Code traces).
- Evaluation Task: The models were trained on
MULTIPLY-then-ADDproblems but tested on held-outADD-then-MULTIPLYcompositions to ensure they were learning logic, not just memorizing strings.
Key Insights and Ideas
- The "Horizon" Problem: You cannot look too far back. Attempting to attribute a full training run (thousands of steps) causes numerical overflow. Even when "clipped" to stay finite, the scores become useless (LDS ~0.20), essentially guessing at chance.
- Short Windows are the Sweet Spot: Attribution is highly reliable when confined to short windows (100 steps). In these local regimes, the math holds up, achieving near-perfect correlation with actual retraining.
- Individual Blocks are Noisy, Families are Stable: If you rank individual 512-token blocks of data by their influence, that ranking fails to generalize across different random seeds or hyperparameters (correlation $\le 0.09$). However, data families (like "worked examples") show consistent patterns across every run.
- Bidirectional Sensitivity: "Worked examples" and "primitive facts" are high-variance data. They are 2.1–2.6x more likely to be in the top 5% of both the most helpful and most harmful data. This suggests these formats are "loud" signals that the model is highly sensitive to, for better or worse, depending on the specific training state.
Key Results
- High Fidelity: Short-window attribution achieved an LDS of 0.86–0.99, proving that MAGIC is a trustworthy tool for local analysis.
- Baseline Failure: A simpler "gradient-dot" baseline (which doesn't account for the optimizer's trajectory) showed near-zero correlation with retraining, highlighting the necessity of trajectory-based methods.
- Capability Acquisition: The 152M model successfully solved 200/200 held-out math compositions. Attribution revealed that "worked examples" were most influential before the skill emerged, while "primitive facts" became most influential during the emergence phase.
Limitations and Future Work
- Synthetic Constraints: The study focuses on synthetic arithmetic. While this provides "ground truth" provenance, it remains to be seen if these insights translate to more abstract reasoning tasks in natural language.
- Scale: While 152M parameters is significant for a validation study involving dozens of retrains, it is small compared to frontier models (7B+ parameters).
- The "Web Text" Confound: Deleting web text changed the ratio of math-to-text in the corpus, making it hard to isolate whether the effect was due to the text's content or simply the resulting increase in "math density."
- Future Direction: The authors suggest that because family-level insights generalize while block-level rankings don't, future research should focus on "family-level" data curation rather than trying to prune individual "bad" documents.
Summary updated