Private Benchmarking to Prevent Contamination and Improve Comparative Evaluation of LLMs

Benchmarking is the de-facto standard for evaluating LLMs, due to its speed, replicability and low cost. However, recent work has pointed out that the majority of the open source benchmarks available today have been contaminated or leaked into LLMs, meaning that LLMs have access to test data during pretraining and/or fine-tuning. This raises serious concerns about the validity of benchmarking studies conducted so far and the future of evaluation using benchmarks. To solve this problem, we propose Private Benchmarking, a solution where test datasets are kept private and models are evaluated without revealing the test data to the model. We describe various scenarios (depending on the trust placed on model owners or dataset owners), and present solutions to avoid data contamination using private benchmarking. For scenarios where the model weights need to be kept private, we describe solutions from confidential computing and cryptography that can aid in private benchmarking. Finally, we present solutions the problem of benchmark dataset auditing, to ensure that private benchmarks are of sufficiently high quality.

Group Testing for Accurate and Efficient Range-Based Near Neighbor Search : An Adaptive Binary Splitting Approach

This work presents an adaptive group testing framework for the range-based high dimensional near neighbor search problem. The proposed method detects high-similarity vectors from an extensive collection of high dimensional vectors, where each vector represents an image descriptor. Our method efficiently marks each item in the collection as neighbor or non-neighbor on the basis of a cosine distance threshold without exhaustive search. Like other methods in the domain of large scale retrieval, our approach exploits the assumption that most of the items in the collection are unrelated to the query. Unlike other methods, it does not assume a large difference between the cosine similarity of the query vector with the least related neighbor and that with the least unrelated non-neighbor. Following the procedure of binary splitting, a multi-stage adaptive group testing algorithm, we split the set of items to be searched into half at each step, and perform dot product tests on smaller and smaller subsets, many of which we are able to prune away. We experimentally show that our method achieves a speed-up over exhaustive search by a factor of more than ten with an accuracy same as that of exhaustive search, on a variety of large datasets. We present a theoretical analysis of the expected number of distance computations per query and the probability that a pool with a certain number of members will be pruned. In this way, our method exploits very useful and practical distributional properties unlike other methods. In our method, all required data structures are created purely offline. Moreover, our method does not impose any strong assumptions on the number of true near neighbors, is adaptible to streaming settings where new vectors are dynamically added to the database, and does not require any parameter tuning.

ABScribe: Rapid Exploration of Multiple Writing Variations in Human-AI Co-Writing Tasks using Large Language Models

Exploring alternative ideas by rewriting text is integral to the writing process. State-of-the-art large language models (LLMs) can simplify writing variation generation. However, current interfaces pose challenges for simultaneous consideration of multiple variations: creating new versions without overwriting text can be difficult, and pasting them sequentially can clutter documents, increasing workload and disrupting writers' flow. To tackle this, we present ABScribe, an interface that supports rapid, yet visually structured, exploration of writing variations in human-AI co-writing tasks. With ABScribe, users can swiftly produce multiple variations using LLM prompts, which are auto-converted into reusable buttons. Variations are stored adjacently within text segments for rapid in-place comparisons using mouse-over interactions on a context toolbar. Our user study with 12 writers shows that ABScribe significantly reduces task workload (d = 1.20, p < 0.001), enhances user perceptions of the revision process (d = 2.41, p < 0.001) compared to a popular baseline workflow, and provides insights into how writers explore variations using LLMs.