Who's Behind the Text? Robust Model Authorship Attribution
As large language models proliferate, identifying which model family generated a given text must remain reliable even when the domain or model version changes at test time. This paper proposes ATLAS, a generative framework that learns source-conditioned text distributions through lightweight adapter tuning and attributes text by comparing clipped token-level likelihoods. A new benchmark, MAABench, demonstrates that this likelihood-based approach outperforms direct discriminative classifiers under distribution shifts, while also enabling identification of distillation teachers.
ATLAS presents a principled and empirically strong advance in model authorship attribution. The likelihood-based formulation yields large and consistent gains over discriminative baselines under out-of-distribution shifts, and the distillation-teacher identification is a compelling extension. Nevertheless, absolute accuracy on the hardest joint-shift setting remains 74.27%, which is well below the threshold needed for high-stakes deployment, and the benchmark covers only five model families and relatively short texts.
The core generative design is well motivated and clearly implemented. Modeling each source family with its own prefix embedding and LoRA up-projection, then aggregating token-level evidence rather than predicting source labels directly, is an elegant way to separate persistent stylistic signals from spurious domain cues. The ablation supports this: clipping the relative likelihoods is the single largest driver of version-shift robustness.
The benchmark scope is narrow—only five families, constrained 90–220 word lengths, and no evaluation of mixed-authorship or adversarial rewriting. The paper itself acknowledges this limitation: 'This work focuses on attribution among a fixed set of candidate source families under the domains and generation conditions covered by MAABench.' Moreover, the inference-time scoring relies on a heuristic clipping threshold and a fixed temperature grid {1,2,4,8} that receive limited theoretical justification. In the distillation analysis, the authors switch to a modified inference protocol—standardizing against unrelated families—with small validation sets, which weakens the claim that the learned conditions alone identify teachers.
The evidence robustly supports the central claim that source-conditioned likelihoods outperform direct discriminative prediction under distribution shift. ATLAS leads the best baseline by 9.00 percentage points on MAABench joint shift and by 33.52pp on OpenTuringBench OOD. However, the baseline pool is heterogeneous, mixing zero-shot scalar statistics with heavily fine-tuned encoders, which complicates attribution of the gains. More importantly, the results also show that version shift remains unsolved in absolute terms: 'relative to ID, accuracy decreases by only 1.75pp on Domain shift, but with 15.42pp on Version shift.'
The authors release code, a HuggingFace demo, and detailed hyperparameters, including LoRA rank 16, learning rate 1e-4, and training durations. Appendix B.1 records the exact optimization settings and hardware footprint: 'Optimization is done with hyperparameters and settings in Table 12 and runs for five epochs on MAABench and three on OpenTuringBench, on four H100-80GB GPUs.' Reproducing MAABench is more problematic because generation depends on proprietary black-box APIs: 'Black-box models are accessed through their respective APIs, while openly accessible white-box models are locally served.' Independent replication therefore requires sustained access to specific, potentially deprecated, API versions.
As large language models (LLMs) are increasingly used to generate text, their misuse has become a growing societal problem, with implications for academic misconduct, model distillation, and cyberattacks. For these reasons, Model Authorship Attribution (MAA), which identifies the source model of a given text, is important for promoting the responsible use of LLMs and improving AI safety. In practice, attribution systems must remain reliable beyond the distributions observed during training. They may encounter texts from unseen domains and writing settings, as well as generations from future versions of model families that were not available when the system was trained. However, existing MAA methods often exhibit substantial performance degradation under such distribution shifts. To address this challenge, we propose Authorship Tracing via Likelihood-based Attribution of Sources (ATLAS), which formulates MAA through source-conditioned text modeling rather than direct source prediction. ATLAS learns how text from each source family is generated and attributes a given text by comparing its likelihood under each learned source condition. We aggregate this likelihood evidence at the token level with clipping so that no single token dominates the decision. We further introduce MAABench, a broad MAA benchmark for evaluating practical OOD settings such as domain and version. Our experiments show that ATLAS consistently outperforms existing baselines across MAABench and OpenTuringBench. In particular, when both the generation domain and the model version change on MAABench, ATLAS achieves 74.27% accuracy, 9.00pp higher than the strongest baseline. Beyond MAA, ATLAS can also help identify distillation source models by measuring which source model a target LLM most closely resembles. Our extended experiments show that ATLAS recovers the known teacher of every distilled model we evaluate.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.