Skip to content
conftrace_
2026 EMNLP EMNLP 2026

ReProF: Residual Prototype Fusion for Continual LLM Attribution

AbstractContinual LLM attribution aims to identify the source of generated text amid the ongoing emergence of new models. While existing training-free, prototype-based methods excel at this task, they rely solely on the final-layer features of a frozen extractor, overlooking rich hierarchical stylistic cues. Moreover, naive multi-layer fusion yields suboptimal results. To address these limitations, we propose ReProF (Residual Prototype Fusion), a training-free framework that partitions the feature extractor’s hierarchy into layer blocks and derives depth-aware analytic prototypes via shallow-to-deep residual fitting: deeper prototypes model the residuals of shallower ones. By fusing predictions across these blocks, ReProF captures multi-granular stylistic cues without requiring parameter tuning or data rehearsal, thereby intrinsically eliminating catastrophic forgetting. Extensive experiments across 19 LLMs and 8 domains demonstrate that ReProF establishes a new state-of-the-art, consistently outperforming existing approaches by a wide margin. Code is available at: https://github.com/wyj-hb/ReProF.

More from EMNLP 2026

All EMNLP 2026 papers →