A key source of complexity in next-generation AI models is the size of model outputs, making it time-consuming to parse and provide reliable feedback on. To ensure such models are aligned, we will need to bolster our understanding of scalable oversight and how to scale up human feedback. To this end, we study the challenges of scalable oversight in the context of goal-conditioned hierarchical reinforcement learning. Hierarchical structure is a promising entrypoint into studying how to scale up human feedback, which in this work we assume can only be provided for model outputs below a threshold size. In the cardinal feedback setting, we develop an apt sub-MDP reward and algorithm that allows us to acquire and scale up low-level feedback for learning with sublinear regret. In the ordinal feedback setting, we show the necessity of both high-and low-level feedback, and develop a hierarchical experimental design algorithm that efficiently acquires both types of feedback for learning. Altogether, our work aims to consolidate the foundations of scalable oversight, formalizing and studying the various challenges thereof.

Accurate detection of errors in large language models (LLM) responses is central to the success of scalable oversight, or providing effective supervision to superhuman intelligence. Yet, self-diagnosis is often unreliable on complex tasks unless aided by reliable external feedback. Multi-agent debate (MAD) seems to be a natural alternative to external feedback: multiple LLMs provide complementary perspectives and cross-checks for error detection. However, prior MAD protocols frame debate as a zero-sum game, where the debaters compete to win the game instead of seeking the truth. Consequently, it leads to debate hacking: debaters tend to mislead the judge by misinterpreting the task or presenting overconfident claims, which introduce more mistakes and underperform single-agent methods. To mitigate the issue, we introduce a new collaborative MAD protocol, termed ColMAD, that reframes MAD as a non-zero sum game. Specifically, ColMAD encourages multiple agents to criticize each other in a supportive way, such that they can complement the missing points of each other. Therefore, the judge agent can make a more informative conclusion based on more comprehensive evidence. Empirically, we show that ColMAD significantly outperforms previous competitive MAD by 19% and brings non-trivial improvements over single-agent methods in error detection.

To this end, we study the challenges of scalable oversight in the context of goal-conditioned hierarchical reinforcement learning.

Scalable oversight protocols aim to enable humans to accurately supervise superhuman AI. In this paper we study debate, where two AI's compete to convince a judge; consultancy, where a single AI tries to convince a judge that asks questions;and compare to a baseline of direct question-answering, where the judge just answers outright without the AI.We use large language models (LLMs) as both AI agents and as stand-ins for human judges, taking the judge models to be weaker than agent models. We benchmark on a diverse range of asymmetries between judges and agents, extending previous work on a single extractive QA task with information asymmetry, to also include mathematics, coding, logic and multimodal reasoning asymmetries. We find that debate outperforms consultancy across all tasks when the consultant is randomly assigned to argue for the correct/incorrect answer. Comparing debate to direct question answering, the results depend on the type of task: in extractive QA tasks with information asymmetry debate outperforms direct question answering, but in other tasks without information asymmetry the results are mixed.Previous work assigned debaters/consultants an answer to argue for.

A key source of complexity in next-generation AI models is the size of model outputs, making it time-consuming to parse and provide reliable feedback on. To ensure such models are aligned, we will need to bolster our understanding of scalable oversight and how to scale up human feedback. To this end, we study the challenges of scalable oversight in the context of goal-conditioned hierarchical reinforcement learning. Hierarchical structure is a promising entrypoint into studying how to scale up human feedback, which in this work we assume can only be provided for model outputs below a threshold size. In the cardinal feedback setting, we develop an apt sub-MDP reward and algorithm that allows us to acquire and scale up low-level feedback for learning with sublinear regret.

As Large Language Models (LLMs) gain expertise across diverse domains and modalities, scalable oversight becomes increasingly challenging, particularly when their capabilities may surpass human evaluators. Debate has emerged as a promising mechanism for enabling such oversight. In this work, we extend the debate paradigm to a multimodal setting, exploring its potential for weaker models to supervise and enhance the performance of stronger models. We focus on visual question answering (VQA), where two "sighted" expert vision-language models debate an answer, while a "blind" (text-only) judge adjudicates based solely on the quality of the arguments. In our framework, the experts defend only answers aligned with their beliefs, thereby obviating the need for explicit role-playing and concentrating the debate on instances of expert disagreement. Experiments on several multimodal tasks demonstrate that the debate framework consistently outperforms individual expert models. Moreover, judgments from weaker LLMs can help instill reasoning capabilities in vision-language models through finetuning.

As AI capabilities increasingly surpass human proficiency in complex tasks, current alignment techniques including SFT and RLHF face fundamental challenges in ensuring reliable oversight. These methods rely on direct human assessment and become untenable when AI outputs exceed human cognitive thresholds. In response to this challenge, we explore two hypotheses: (1) critique of critique can be easier than critique itself, extending the widely-accepted observation that verification is easier than generation to the critique domain, as critique itself is a specialized form of generation; (2) this difficulty relationship is recursively held, suggesting that when direct evaluation is infeasible, performing high-order critiques (e.g., critique of critique of critique) offers a more tractable supervision pathway. To examine these hypotheses, we perform Human-Human, Human-AI, and AI-AI experiments across multiple tasks. Our results demonstrate encouraging evidence supporting these hypotheses and suggest that recursive self-critiquing is a promising direction for scalable oversight.

The emergence of large language models (LLMs) has sparkedthe discussion on Artificial Superintelligence (ASI), a hypothetical AI system surpassing human intelligence. Though ASI is still hypothetical and far from current AI capabilities, existing alignment methods struggle to guide such advanced AI ensure its safety in the future. It is essential to discuss the alignment of such AI now. Superalignment, the alignment of AI at superhuman levels of capability systems with human values and safety requirements, aims to address two primary goals: scalability in supervision to provide high-quality guidance signals and robust governance to ensure alignment with human values. In this survey, we review the original scalable oversight problem and corresponding methods and potential solutions for superalignment. Specifically, we introduce the Figure 1: Challenges from the perspectives of supervision challenges and limitations of current alignment and governance. While supervision perspective paradigms in addressing the superalignment focuses on providing high-quality guidance signals for problem. Then we review scalable oversight enhancing system competence, governance perspective methods for superalignment. Finally, we discuss emphasizes aligning the behavior of advanced aI with the key challenges and propose pathways human values to prevent harmful outcomes.

We have witnessed superhuman intelligence thanks to the fast development of large language models and multimodal language models. As the application of such superhuman models becomes more and more popular, a critical question arises here: how can we ensure superhuman models are still safe, reliable and aligned well to human values? In this position paper, we discuss the concept of superalignment from the learning perspective to answer this question by outlining the learning paradigm shift from large-scale pretraining, supervised fine-tuning, to alignment training. We define superalignment as designing effective and efficient alignment algorithms to learn from noisy-labeled data (point-wise samples or pair-wise preference data) in a scalable way when the task becomes very complex for human experts to annotate and the model is stronger than human experts. We highlight some key research problems in superalignment, namely, weak-to-strong generalization, scalable oversight, and evaluation. We then present a conceptual framework for superalignment, which consists of three modules: an attacker which generates adversary queries trying to expose the weaknesses of a learner model; a learner which will refine itself by learning from scalable feedbacks generated by a critic model along with minimal human experts; and a critic which generates critics or explanations for a given query-response pair, with a target of improving the learner by criticizing. We discuss some important research problems in each component of this framework and highlight some interesting research ideas that are closely related to our proposed framework, for instance, self-alignment, self-play, self-refinement, and more. Last, we highlight some future research directions for superalignment, including identification of new emergent risks and multi-dimensional alignment.