Tong Mu

Sandhini Agarwal, Lama Ahmad, Jason Ai et al. · openai

We present gpt-oss-120b and gpt-oss-20b, two open-weight reasoning models that push the frontier of accuracy and inference cost. The models use an efficient mixture-of-expert transformer architecture and are trained using large-scale distillation and reinforcement learning. We optimize the models to have strong agentic capabilities (deep research browsing, python tool use, and support for developer-provided functions), all while using a rendered chat format that enables clear instruction following and role delineation. Both models achieve strong results on benchmarks ranging from mathematics, coding, and safety. We release the model weights, inference implementations, tool environments, and tokenizers under an Apache 2.0 license to enable broad use and further research.

Josh Achiam, Steven Adler, Sandhini Agarwal et al. · berkeley, deepmind

We report the development of GPT-4, a large-scale, multimodal model which can accept image and text inputs and produce text outputs. While less capable than humans in many real-world scenarios, GPT-4 exhibits human-level performance on various professional and academic benchmarks, including passing a simulated bar exam with a score around the top 10% of test takers. GPT-4 is a Transformer-based model pre-trained to predict the next token in a document. The post-training alignment process results in improved performance on measures of factuality and adherence to desired behavior. A core component of this project was developing infrastructure and optimization methods that behave predictably across a wide range of scales. This allowed us to accurately predict some aspects of GPT-4's performance based on models trained with no more than 1/1,000th the compute of GPT-4.

Evan Zheran Liu, Sahaana Suri, Tong Mu et al.

Whereas machine learning models typically learn language by directly training on language tasks (e.g., next-word prediction), language emerges in human children as a byproduct of solving non-language tasks (e.g., acquiring food). Motivated by this observation, we ask: can embodied reinforcement learning (RL) agents also indirectly learn language from non-language tasks? Learning to associate language with its meaning requires a dynamic environment with varied language. Therefore, we investigate this question in a multi-task environment with language that varies across the different tasks. Specifically, we design an office navigation environment, where the agent's goal is to find a particular office, and office locations differ in different buildings (i.e., tasks). Each building includes a floor plan with a simple language description of the goal office's location, which can be visually read as an RGB image when visited. We find RL agents indeed are able to indirectly learn language. Agents trained with current meta-RL algorithms successfully generalize to reading floor plans with held-out layouts and language phrases, and quickly navigate to the correct office, despite receiving no direct language supervision.

Zhimin Zhu, Jianguo Zhao, Tong Mu et al.

In deep learning, Multi-Layer Perceptrons (MLPs) have once again garnered attention from researchers. This paper introduces MC-MLP, a general MLP-like backbone for computer vision that is composed of a series of fully-connected (FC) layers. In MC-MLP, we propose that the same semantic information has varying levels of difficulty in learning, depending on the coordinate frame of features. To address this, we perform an orthogonal transform on the feature information, equivalent to changing the coordinate frame of features. Through this design, MC-MLP is equipped with multi-coordinate frame receptive fields and the ability to learn information across different coordinate frames. Experiments demonstrate that MC-MLP outperforms most MLPs in image classification tasks, achieving better performance at the same parameter level. The code will be available at: https://github.com/ZZM11/MC-MLP.

Tong Mu, Alec Helyar, Johannes Heidecke et al.

Reinforcement learning based fine-tuning of large language models (LLMs) on human preferences has been shown to enhance both their capabilities and safety behavior. However, in cases related to safety, without precise instructions to human annotators, the data collected may cause the model to become overly cautious, or to respond in an undesirable style, such as being judgmental. Additionally, as model capabilities and usage patterns evolve, there may be a costly need to add or relabel data to modify safety behavior. We propose a novel preference modeling approach that utilizes AI feedback and only requires a small amount of human data. Our method, Rule Based Rewards (RBR), uses a collection of rules for desired or undesired behaviors (e.g. refusals should not be judgmental) along with a LLM grader. In contrast to prior methods using AI feedback, our method uses fine-grained, composable, LLM-graded few-shot prompts as reward directly in RL training, resulting in greater control, accuracy and ease of updating. We show that RBRs are an effective training method, achieving an F1 score of 97.1, compared to a human-feedback baseline of 91.7, resulting in much higher safety-behavior accuracy through better balancing usefulness and safety.

Tong Mu, Yihao Liu, Mehran Armand

Imitation learning frameworks for robotic manipulation have drawn attention in the recent development of language model grounded robotics. However, the success of the frameworks largely depends on the coverage of the demonstration cases: When the demonstration set does not include examples of how to act in all possible situations, the action may fail and can result in cascading errors. To solve this problem, we propose a framework that uses serialized Finite State Machine (FSM) to generate demonstrations and improve the success rate in manipulation tasks requiring a long sequence of precise interactions. To validate its effectiveness, we use environmentally evolving and long-horizon puzzles that require long sequential actions. Experimental results show that our approach achieves a success rate of up to 98 in these tasks, compared to the controlled condition using existing approaches, which only had a success rate of up to 60, and, in some tasks, almost failed completely.

Tong Mu, Stephan Zheng, Alexander Trott

Principal-Agent (PA) problems describe a broad class of economic relationships characterized by misaligned incentives and asymmetric information. The Principal's problem is to find optimal incentives given the available information, e.g., a manager setting optimal wages for its employees. Whereas the Principal is often assumed rational, comparatively little is known about solutions when the Principal is boundedly rational, especially in the sequential setting, with multiple Agents, and with multiple information channels. Here, we develop RIRL, a deep reinforcement learning framework that solves such complex PA problems with a rationally inattentive Principal. Such a Principal incurs a cost for paying attention to information, which can model forms of bounded rationality. We use RIRL to analyze rich economic phenomena in manager-employee relationships. In the single-step setting, 1) RIRL yields wages that are consistent with theoretical predictions; and 2) non-zero attention costs lead to simpler but less profitable wage structures, and increased Agent welfare. In a sequential setting with multiple Agents, RIRL shows opposing consequences of the Principal's inattention to different information channels: 1) inattention to Agents' outputs closes wage gaps based on ability differences; and 2) inattention to Agents' efforts induces a social dilemma dynamic in which Agents work harder, but essentially for free. Moreover, RIRL reveals non-trivial relationships between the Principal's inattention and Agent types, e.g., if Agents are prone to sub-optimal effort choices, payment schedules are more sensitive to the Principal's attention cost. As such, RIRL can reveal novel economic relationships and enables progress towards understanding the effects of bounded rationality in dynamic settings.

Tong Mu, Georgios Theocharous, David Arbour et al.

Online reinforcement learning (RL) algorithms are often difficult to deploy in complex human-facing applications as they may learn slowly and have poor early performance. To address this, we introduce a practical algorithm for incorporating human insight to speed learning. Our algorithm, Constraint Sampling Reinforcement Learning (CSRL), incorporates prior domain knowledge as constraints/restrictions on the RL policy. It takes in multiple potential policy constraints to maintain robustness to misspecification of individual constraints while leveraging helpful ones to learn quickly. Given a base RL learning algorithm (ex. UCRL, DQN, Rainbow) we propose an upper confidence with elimination scheme that leverages the relationship between the constraints, and their observed performance, to adaptively switch among them. We instantiate our algorithm with DQN-type algorithms and UCRL as base algorithms, and evaluate our algorithm in four environments, including three simulators based on real data: recommendations, educational activity sequencing, and HIV treatment sequencing. In all cases, CSRL learns a good policy faster than baselines.

Tong Mu, Karan Goel, Emma Brunskill

In many cases an intelligent agent may want to learn how to mimic a single observed demonstrated trajectory. In this work we consider how to perform such procedural learning from observation, which could help to enable agents to better use the enormous set of video data on observation sequences. Our approach exploits the properties of this setting to incrementally build an open loop action plan that can yield the desired subsequence, and can be used in both Markov and partially observable Markov domains. In addition, procedures commonly involve repeated extended temporal action subsequences. Our method optimistically explores actions to leverage potential repeated structure in the procedure. In comparing to some state-of-the-art approaches we find that our explicit procedural learning from observation method is about 100 times faster than policy-gradient based approaches that learn a stochastic policy and is faster than model based approaches as well. We also find that performing optimistic action selection yields substantial speed ups when latent dynamical structure is present.