Unifying Corroborative and Contributive Attributions in Large Language Models

Published: April 9, 2024

As businesses, products, and services spring up around large language models, the trustworthiness of these models hinges on verifiability their outputs. However, methods for explaining model outputs largely fall across two distinct fields study which both use term "attribution" to refer entirely separate techniques: citation generation training data attribution. In many modern applications, such as legal document medical question answering, types attributions are important. this work, we argue present a unified framework attributions. We show how existing different attribution under framework. also discuss real-world cases where one or required. believe that will guide case driven development systems leverage attribution, well standardization evaluation.

Language: Английский

---

Language-Agnostic Bias Detection in Language Models with Bias Probing

Published: Jan. 1, 2023

Pretrained language models (PLMs) are key components in NLP, but they contain strong social biases. Quantifying these biases is challenging because current methods focusing on fill-the-mask objectives sensitive to slight changes input. To address this, we propose a bias probing technique called LABDet, for evaluating PLMs with robust and language-agnostic method. For nationality as case study, show that LABDet “surfaces” by training classifier top of frozen PLM non-nationality sentiment detection. We find consistent patterns across monolingual six languages align historical political context. also English BERT surfaced correlates well the pretraining data; thus, our work one few studies directly links data behavior. Finally, verify LABDet’s reliability applicability different templates through an extensive set robustness checks. publicly share code dataset https://github.com/akoksal/LABDet.

Language: Английский

---

Spacerini: Plug-and-play Search Engines with Pyserini and Hugging Face

Published: Jan. 1, 2023

Christopher Akiki, Odunayo Ogundepo, Aleksandra Piktus, Xinyu Zhang, Akintunde Oladipo, Jimmy Lin, Martin Potthast. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations. 2023.

Language: Английский

---

Mitigating Hallucination Issues in Small-Parameter LLMs through Inter-Layer Contrastive Decoding

2022 International Joint Conference on Neural Networks (IJCNN), Journal Year: 2024, Volume and Issue: 33, P. 1 - 8

Published: June 30, 2024

Language: Английский

---

The Mirrored Influence Hypothesis: Efficient Data Influence Estimation by Harnessing Forward Passes

2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Journal Year: 2024, Volume and Issue: 18, P. 26276 - 26285

Published: June 16, 2024

Language: Английский

---

SLANGO - The Initial Blueprint of Privacy-Oriented Legal Query Assistance: Exploring the Potential of Retrieval-Augmented Generation for German Law Using SPR

Lecture notes in computer science, Journal Year: 2024, Volume and Issue: unknown, P. 208 - 221

Published: Nov. 28, 2024

Language: Английский

---

Enhancing Long Context Performance in LLMs Through Inner Loop Query Mechanism

Published: Dec. 17, 2024

Transformers have a quadratic scaling of computational complexity with input size, which limits the context window size large language models (LLMs) in both training and inference. Meanwhile, retrieval-augmented generation (RAG) besed can better handle longer contexts by using retrieval system to filter out unnecessary information. However, most RAG methods only perform based on initial query, may not work well complex questions that require deeper reasoning. We introduce novel approach, Inner Loop Memory Augmented Tree Retrieval (ILM-TR), involving inner-loop queries, query question itself but also intermediate findings. At inference time, our model retrieves information from system, integrating data lengthy documents at various levels abstraction. Based retrieved, LLM generates texts stored an area named Short-Term (STM) is then used formulate next query. This process repeated until text STM converged. Our experiments demonstrate offers improvements over traditional LLMs, particularly long tests such as Multi-Needle In A Haystack (M-NIAH) BABILong.

Language: Английский

---

Disentangling Fine-Tuning from Pre-Training in Visual Captioning with Hybrid Markov Logic

2021 IEEE International Conference on Big Data (Big Data), Journal Year: 2024, Volume and Issue: unknown, P. 5422 - 5431

Published: Dec. 15, 2024

Language: Английский

---

GEM-RAG: Graphical Eigen Memories for Retrieval Augmented Generation

Published: Dec. 18, 2024

Language: Английский

---

A Dynamic Grid Index for CkNN Queries on Large-Scale Road Networks with Moving Objects

Applied Sciences, Journal Year: 2023, Volume and Issue: 13(8), P. 4946 - 4946

Published: April 14, 2023

As the Internet of Things devices are deployed on a large scale, location-based services being increasingly utilized. Among these services, kNN (k-nearest neighbor) queries based road network constraints have gained importance. This study focuses CkNN (continuous k-nearest for non-uniformly distributed moving objects with large-scale dynamic constraints, where continuously and periodically queried their motion evolution. The present high-concurrency query under super-large faces problems, such as high computational cost low efficiency. aim this is to ensure concurrency nearest neighbor requests while shortening response time reducing global computation costs. To address issue, we propose DVTG-Index (Dynamic V-Tree Double-Layer Grid Index), which intelligently adjusts index granularity by merging splitting subgraphs move, thereby filtering unnecessary vertices. Based DVTG-Index, further DVTG-CkNN algorithm calculate initial utilize existing results speed up query. Finally, extensive experiments real networks confirm superior performance our proposed method, has significant practical applications in objects.

Language: Английский

---