Daily AI Paper

Data attribution for language models is the problem of figuring out which training examples caused a model to produce a particular answer. That sounds straightforward, but in practice it is very hard to verify. If you want to know whether a method is correct, the most direct test is to retrain the model without the supposedly important documents and see whether the behavior disappears. For large language models, that kind of counterfactual retraining is expensive, so the field has relied on indirect benchmarks. The issue this paper starts from is that those benchmarks often change many things at once, so when an attribution method performs poorly, you still do not know why. Is it because the method is bad? Because the dataset is too redundant? Because there are distractors that look like the query? Or because the evidence is spread across a long document and gets washed out? To answer that, the paper introduces a controlled benchmark called CATT, short for Controlled Attribution Benchmark. The key idea is to create synthetic facts that are completely fictitious, so the training documents are the only source of the knowledge. The model is fine-tuned on these made-up facts, and then tested with held-out queries in a different format. Since the facts are synthetic, the ground truth for attribution is unambiguous: the documents that describe a given fact are the only correct supporting documents. That gives the authors a way to evaluate attribution methods with real diagnostic power. This matters because attribution is not just an academic exercise. It is closely tied to questions of interpretability, debugging, dataset auditing, copyright, and trust. If a model gives a correct answer, we may want to know whether it came from a relevant training example or from some spurious association. Attribution methods are supposed to help with that by assigning a score to each training document, estimating how much it influenced the prediction. But existing fact-tracing benchmarks often mix together several dataset properties, like how many documents support each fact, how many total facts exist, and whether there are similar-looking distractors. That makes results hard to interpret. One method may look good on one benchmark and bad on another, not because the method fundamentally changed, but because the benchmark design changed. The paper’s broader contribution is to explore the design space of attribution evaluation, rather than just proposing yet another estimator. The authors build configurable synthetic datasets and vary one property at a time. They focus on three main factors: document redundancy, meaning how many different documents express the same fact; lexical distractors, meaning documents that look similar to the query but are irrelevant; and the number of facts in the dataset. By controlling these variables, they can isolate what breaks attribution. They test two synthetic tasks. The first is function bindings, where a new invented function token is mapped to a constant answer token. For example, a fake function might always return a particular value. Because the tokens are invented, the model cannot have learned them during pretraining. Any correct answer must come from the fine-tuning documents. The second task is routing rules, where a fictitious rule maps objects with certain attributes to a queue or category. This tests more compositional reasoning: the model has to learn a rule and apply it to a new object described in a different way at test time. These tasks are useful because they are synthetic but still resemble the kinds of structured knowledge and rule learning that attribution methods may encounter in real models. The benchmark is built by fine-tuning a pretrained language model, specifically Qwen3.5-4B with LoRA, on each synthetic dataset configuration. For each configuration, they train multiple seeds to measure variance. Then they ask a held-out question about one of the synthetic facts and see which training documents the attribution method ranks highest. A good method should place the true supporting documents above unrelated ones. The first major finding is that redundancy and distractors consistently hurt attribution. In other words, when the same fact is repeated across more documents, or when there are many irrelevant documents that resemble the query, the attribution methods become less reliable. This is a dilution effect: the evidence is there, but it gets spread out and blurred. The paper shows that this is not just a small nuisance; it is a systematic failure mode across multiple baselines and settings. To understand why, the authors inspect attribution at the token level. This is an important move. Most attribution methods are document-level by default, but a document is made of tokens, and the evidence for a fact may be concentrated in only a few of them. The paper shows that the causal evidence inside a training document is often highly localized. Standard methods, however, average or aggregate scores across the whole document, which mixes the important tokens with lots of irrelevant context. That can bury the signal in noise. This leads to two design choices the paper studies in detail. The first is the loss used to compute gradients. Many gradient-based attribution methods rely on cross-entropy loss. But cross-entropy can suppress gradients when the model is already confident about a token. The authors instead use a negative log-odds loss, which compares the target token against the rest of the probability mass. In plain terms, this makes the important tokens stand out more clearly, rather than being muted by model confidence. The paper’s claim is not that this is a new attribution method from scratch, but that this loss choice exposes the relevant evidence better when computing token-level gradients. The second design choice is aggregation. Once you have token-level attribution scores, you still need to collapse them into a single document score. The standard approach is to average over all tokens, but that is exactly what causes dilution when only a few tokens matter. The authors instead aggregate only the top token scores, keeping the most informative parts and ignoring the rest. This simple change is meant to preserve the localized evidence instead of smearing it out across the whole document. Experimentally, these two adjustments make a large difference. Across multiple baselines and dataset settings, using token-level negative log-odds and top-token aggregation improves fact-tracing recall by 35 to 72 percentage points. That is a substantial gain, and it supports the paper’s central diagnosis: the main problem is not always that attribution methods are fundamentally unable to find the right documents, but that their default design choices can wash out the signal they are trying to measure. The paper also checks whether these ideas generalize beyond the synthetic benchmark. They evaluate on FTRACE-TREX, an external fact-tracing benchmark, and find that token-level negative log-odds improves counterfactual retraining performance for every method they test, including TRAK, EK-FAC, and Grad-Dot, by up to 4.9 times. That result is important because it suggests the loss correction is not just exploiting the synthetic setup; it can also help on a more realistic benchmark. What should we take away from the results? First, benchmark design matters a lot. If evaluation datasets vary several properties at once, attribution methods can look better or worse for reasons that are impossible to disentangle. CATT shows that redundancy and distractors are especially damaging. Second, document-level attribution can hide token-level evidence. If the real support is concentrated in a few words or phrases, averaging over the whole document is a bad idea. Third, some of the most effective fixes are surprisingly simple: change the loss to negative log-odds, and aggregate only the strongest token scores. There are also limitations to keep in mind. The benchmark is synthetic, so although it gives clean ground truth, it does not capture every messiness of real pretraining data. Real-world facts are often spread across paraphrases, noisy web text, and overlapping sources in ways that are harder to simulate. The paper does try to bridge that gap by varying the benchmark along realistic axes like redundancy and distractors, and by testing on an external benchmark, but synthetic control is still not the same as open-world pretraining. Also, the token-level analysis is described as a localized gradient signal, not a perfect causal deletion test, so it remains an approximation to true influence. And the improvements are shown for gradient-based methods and specific baselines; other attribution families may behave differently. Even so, the paper makes a strong case that evaluation itself is part of the research problem. If we want better data attribution, we need benchmarks that tell us not just whether a method fails, but why. CATT provides that diagnostic lens, and the main lesson is clear: when attribution seems weak, the issue may be dilution, not absence of evidence.