Mitigating the Influence of Distractor Tasks in LMs with Prior-Aware Decoding
The broad capabilities of Language Models (LMs) can be limited by their sensitivity to distractor tasks: LMs can infer secondary tasks from the prompt in addition to the intended one, leading to unwanted outputs. For example, prompt injection attacks can cause models to deviate from explicit directives. In some 'inverse scaling' cases, this unwanted behaviour actually worsens as models scale up to at least 540B parameters. We present a theoretical framework that interprets LMs as a product of experts that combine multiple data generation processes. Based on this framework, we demonstrate prior-aware decoding (PAD) - a simple contrastive inference method to reduce the influence of distractor tasks. We apply PAD to eleven models, across four datasets, and find improvements in 41 out of 44 task-model combinations, with a median increase in task completion proportion of 40%. The results suggest a promising direction for further development towards more reliable language models.
Code (0)
등록된 구현이 없습니다.
Tasks
Instruction FollowingMethods 이 논문이 사용한 방법론
Similar Papers 제목 키워드 기반
PruNeRF: Segment-Centric Dataset Pruning via 3D Spatial Consistency
Neural Radiance Fields (NeRF) have shown remarkable performance in learning 3D scenes. However, NeRF exhibits vulnerability when confronted with distractors in the training images -- unexpected objects are present only w…
NeRFSparse View Distractor-Free Gaussian Splatting
3D Gaussian Splatting (3DGS) enables efficient training and fast novel view synthesis in static environments. To address challenges posed by transient objects, distractor-free 3DGS methods have emerged and shown promisin…
Novel View SynthesisOn the Importance of Distractors for Few-Shot Classification
Few-shot classification aims at classifying categories of a novel task by learning from just a few (typically, 1 to 5) labelled examples. An effective approach to few-shot classification involves a prior model trained on…
ClassificationContrastive LearningWhen words collide: Bayesian meta-analyses of distractor and target properties in the picture-word interference paradigm
In the picture-word interference paradigm, participants name pictures while ignoring a written or spoken distractor word. Naming times to the pictures are slowed down by the presence of the distractor word. Various prope…
Distractor-Aware Memory-Based Visual Object Tracking
Recent emergence of memory-based video segmentation methods such as SAM2 has led to models with excellent performance in segmentation tasks, achieving leading results on numerous benchmarks. However, these modes are not …
Visual Object TrackingVideo Segmentation