Great Models Think Alike and this Undermines AI Oversight

• URL: http://arxiv.org/abs/2502.04313v1
• Date: Thu, 06 Feb 2025 18:56:01 GMT
• Title: Great Models Think Alike and this Undermines AI Oversight
• Authors: Shashwat Goel, Joschka Struber, Ilze Amanda Auzina, Karuna K Chandra, Ponnurangam Kumaraguru, Douwe Kiela, Ameya Prabhu, Matthias Bethge, Jonas Geiping,
• Abstract summary: We study how model similarity affects both aspects of AI oversight. We propose a probabilistic metric for LM similarity based on overlap in model mistakes. Our work underscores the importance of reporting and correcting for model similarity.
• Score: 47.7725284401918
• License:
• Abstract: As Language Model (LM) capabilities advance, evaluating and supervising them at scale is getting harder for humans. There is hope that other language models can automate both these tasks, which we refer to as "AI Oversight". We study how model similarity affects both aspects of AI oversight by proposing a probabilistic metric for LM similarity based on overlap in model mistakes. Using this metric, we first show that LLM-as-a-judge scores favor models similar to the judge, generalizing recent self-preference results. Then, we study training on LM annotations, and find complementary knowledge between the weak supervisor and strong student model plays a crucial role in gains from "weak-to-strong generalization". As model capabilities increase, it becomes harder to find their mistakes, and we might defer more to AI oversight. However, we observe a concerning trend -- model mistakes are becoming more similar with increasing capabilities, pointing to risks from correlated failures. Our work underscores the importance of reporting and correcting for model similarity, especially in the emerging paradigm of AI oversight.

Related papers

• Self-supervised Analogical Learning using Language Models [59.64260218737556]
  We propose SAL, a self-supervised analogical learning framework. SAL mimics the human analogy process and trains models to explicitly transfer high-quality symbolic solutions. We show that the resulting models outperform base language models on a wide range of reasoning benchmarks.
  arXiv  Detail & Related papers  (2025-02-03T02:31:26Z)
• Debate Helps Weak-to-Strong Generalization [68.70065254564642]
  We investigate ways of improving human supervision with a strong pretrained model and then supervise the strong model with enhanced weak human supervision. We find that debate can assist a weak model in extracting trustworthy information from an untrustworthy strong model. Experiments on the OpenAI weak-to-strong NLP benchmarks show that the combination approach leads to better alignment.
  arXiv  Detail & Related papers  (2025-01-21T05:36:13Z)
• Unveiling AI's Blind Spots: An Oracle for In-Domain, Out-of-Domain, and Adversarial Errors [4.525077884001726]
  We conduct empirical evaluations using a "mentor" model-a deep neural network designed to predict another "mentee" model's errors. We develop an "oracle" mentor model, dubbed SuperMentor, that can outperform baseline mentors in predicting errors across different error types from the ImageNet-1K dataset.
  arXiv  Detail & Related papers  (2024-10-03T11:02:39Z)
• Scaling Laws Do Not Scale [54.72120385955072]
  Recent work has argued that as the size of a dataset increases, the performance of a model trained on that dataset will increase. We argue that this scaling law relationship depends on metrics used to measure performance that may not correspond with how different groups of people perceive the quality of models' output. Different communities may also have values in tension with each other, leading to difficult, potentially irreconcilable choices about metrics used for model evaluations.
  arXiv  Detail & Related papers  (2023-07-05T15:32:21Z)
• A Review of the Role of Causality in Developing Trustworthy AI Systems [16.267806768096026]
  State-of-the-art AI models largely lack an understanding of the cause-effect relationship that governs human understanding of the real world. Recently, causal modeling and inference methods have emerged as powerful tools to improve the trustworthiness aspects of AI models.
  arXiv  Detail & Related papers  (2023-02-14T11:08:26Z)
• Large Language Models with Controllable Working Memory [64.71038763708161]
  Large language models (LLMs) have led to a series of breakthroughs in natural language processing (NLP) What further sets these models apart is the massive amounts of world knowledge they internalize during pretraining. How the model's world knowledge interacts with the factual information presented in the context remains under explored.
  arXiv  Detail & Related papers  (2022-11-09T18:58:29Z)
• Distill on the Go: Online knowledge distillation in self-supervised learning [1.1470070927586016]
  Recent works have shown that wider and deeper models benefit more from self-supervised learning than smaller models. We propose Distill-on-the-Go (DoGo), a self-supervised learning paradigm using single-stage online knowledge distillation. Our results show significant performance gain in the presence of noisy and limited labels.
  arXiv  Detail & Related papers  (2021-04-20T09:59:23Z)
• What do we expect from Multiple-choice QA Systems? [70.86513724662302]
  We consider a top performing model on several Multiple Choice Question Answering (MCQA) datasets. We evaluate it against a set of expectations one might have from such a model, using a series of zero-information perturbations of the model's inputs.
  arXiv  Detail & Related papers  (2020-11-20T21:27:10Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.