Avi Singh

Aviral Kumar, Vincent Zhuang, Rishabh Agarwal et al. · deepmind

Self-correction is a highly desirable capability of large language models (LLMs), yet it has consistently been found to be largely ineffective in modern LLMs. Current methods for training self-correction typically depend on either multiple models, a more advanced model, or additional forms of supervision. To address these shortcomings, we develop a multi-turn online reinforcement learning (RL) approach, SCoRe, that significantly improves an LLM's self-correction ability using entirely self-generated data. To build SCoRe, we first show that variants of supervised fine-tuning (SFT) on offline model-generated correction traces are often insufficient for instilling self-correction behavior. In particular, we observe that training via SFT falls prey to either a distribution mismatch between mistakes made by the data-collection policy and the model's own responses, or to behavior collapse, where learning implicitly prefers only a certain mode of correction behavior that is often not effective at self-correction on test problems. SCoRe addresses these challenges by training under the model's own distribution of self-generated correction traces and using appropriate regularization to steer the learning process into learning a self-correction behavior that is effective at test time as opposed to fitting high-reward responses for a given prompt. This regularization process includes an initial phase of multi-turn RL on a base model to generate a policy initialization that is less susceptible to collapse, followed by using a reward bonus to amplify self-correction. With Gemini 1.0 Pro and 1.5 Flash models, we find that SCoRe achieves state-of-the-art self-correction performance, improving the base models' self-correction by 15.6% and 9.1% respectively on MATH and HumanEval.

Homer Walke, Jonathan Yang, Albert Yu et al.

Reinforcement learning (RL) algorithms hold the promise of enabling autonomous skill acquisition for robotic systems. However, in practice, real-world robotic RL typically requires time consuming data collection and frequent human intervention to reset the environment. Moreover, robotic policies learned with RL often fail when deployed beyond the carefully controlled setting in which they were learned. In this work, we study how these challenges can all be tackled by effective utilization of diverse offline datasets collected from previously seen tasks. When faced with a new task, our system adapts previously learned skills to quickly learn to both perform the new task and return the environment to an initial state, effectively performing its own environment reset. Our empirical results demonstrate that incorporating prior data into robotic reinforcement learning enables autonomous learning, substantially improves sample-efficiency of learning, and enables better generalization. Project website: https://sites.google.com/view/ariel-berkeley/

David B. D'Ambrosio, Jonathan Abelian, Saminda Abeyruwan et al.

We present a deep-dive into a real-world robotic learning system that, in previous work, was shown to be capable of hundreds of table tennis rallies with a human and has the ability to precisely return the ball to desired targets. This system puts together a highly optimized perception subsystem, a high-speed low-latency robot controller, a simulation paradigm that can prevent damage in the real world and also train policies for zero-shot transfer, and automated real world environment resets that enable autonomous training and evaluation on physical robots. We complement a complete system description, including numerous design decisions that are typically not widely disseminated, with a collection of studies that clarify the importance of mitigating various sources of latency, accounting for training and deployment distribution shifts, robustness of the perception system, sensitivity to policy hyper-parameters, and choice of action space. A video demonstrating the components of the system and details of experimental results can be found at https://youtu.be/uFcnWjB42I0.

David Brandfonbrener, Stephen Tu, Avi Singh et al.

We consider how to most efficiently leverage teleoperator time to collect data for learning robust image-based value functions and policies for sparse reward robotic tasks. To accomplish this goal, we modify the process of data collection to include more than just successful demonstrations of the desired task. Instead we develop a novel protocol that we call Visual Backtracking Teleoperation (VBT), which deliberately collects a dataset of visually similar failures, recoveries, and successes. VBT data collection is particularly useful for efficiently learning accurate value functions from small datasets of image-based observations. We demonstrate VBT on a real robot to perform continuous control from image observations for the deformable manipulation task of T-shirt grasping. We find that by adjusting the data collection process we improve the quality of both the learned value functions and policies over a variety of baseline methods for data collection. Specifically, we find that offline reinforcement learning on VBT data outperforms standard behavior cloning on successful demonstration data by 13% when both methods are given equal-sized datasets of 60 minutes of data from the real robot.

Yixin Liu, Avi Singh, C. Daniel Freeman et al.

Despite their success in many natural language tasks, solving math problems remains a significant challenge for large language models (LLMs). A large gap exists between LLMs' pass-at-one and pass-at-N performance in solving math problems, suggesting LLMs might be close to finding correct solutions, motivating our exploration of fine-tuning methods to unlock LLMs' performance. Using the challenging MATH dataset, we investigate three fine-tuning strategies: (1) solution fine-tuning, where we fine-tune to generate a detailed solution for a given math problem; (2) solution-cluster re-ranking, where the LLM is fine-tuned as a solution verifier/evaluator to choose among generated candidate solution clusters; (3) multi-task sequential fine-tuning, which integrates both solution generation and evaluation tasks together efficiently to enhance the LLM performance. With these methods, we present a thorough empirical study on a series of PaLM 2 models and find: (1) The quality and style of the step-by-step solutions used for fine-tuning can make a significant impact on the model performance; (2) While solution re-ranking and majority voting are both effective for improving the model performance when used separately, they can also be used together for an even greater performance boost; (3) Multi-task fine-tuning that sequentially separates the solution generation and evaluation tasks can offer improved performance compared with the solution fine-tuning baseline. Guided by these insights, we design a fine-tuning recipe that yields approximately 58.8% accuracy on the MATH dataset with fine-tuned PaLM 2-L models, an 11.2% accuracy improvement over the few-shot performance of pre-trained PaLM 2-L model with majority voting.

Gemini Team, Petko Georgiev, Ving Ian Lei et al. · deepmind, mila

In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75% time savings across 10 different job categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Avi Singh, John D. Co-Reyes, Rishabh Agarwal et al. · anthropic, deepmind

Fine-tuning language models~(LMs) on human-generated data remains a prevalent practice. However, the performance of such models is often limited by the quantity and diversity of high-quality human data. In this paper, we explore whether we can go beyond human data on tasks where we have access to scalar feedback, for example, on math problems where one can verify correctness. To do so, we investigate a simple self-training method based on expectation-maximization, which we call ReST$^{EM}$, where we (1) generate samples from the model and filter them using binary feedback, (2) fine-tune the model on these samples, and (3) repeat this process a few times. Testing on advanced MATH reasoning and APPS coding benchmarks using PaLM-2 models, we find that ReST$^{EM}$ scales favorably with model size and significantly surpasses fine-tuning only on human data. Overall, our findings suggest self-training with feedback can substantially reduce dependence on human-generated data.

Rishabh Agarwal, Avi Singh, Lei M. Zhang et al. · mila

Large language models (LLMs) excel at few-shot in-context learning (ICL) -- learning from a few examples provided in context at inference, without any weight updates. Newly expanded context windows allow us to investigate ICL with hundreds or thousands of examples -- the many-shot regime. Going from few-shot to many-shot, we observe significant performance gains across a wide variety of generative and discriminative tasks. While promising, many-shot ICL can be bottlenecked by the available amount of human-generated examples. To mitigate this limitation, we explore two new settings: Reinforced and Unsupervised ICL. Reinforced ICL uses model-generated chain-of-thought rationales in place of human examples. Unsupervised ICL removes rationales from the prompt altogether, and prompts the model only with domain-specific questions. We find that both Reinforced and Unsupervised ICL can be quite effective in the many-shot regime, particularly on complex reasoning tasks. Finally, we demonstrate that, unlike few-shot learning, many-shot learning is effective at overriding pretraining biases, can learn high-dimensional functions with numerical inputs, and performs comparably to fine-tuning. We also find that inference cost increases linearly in the many-shot regime, and frontier LLMs benefit from many-shot ICL to varying degrees. Our analysis also reveals the limitations of next-token prediction loss as an indicator of downstream ICL performance.

Avi Singh, Huihan Liu, Gaoyue Zhou et al.

Reinforcement learning provides a general framework for flexible decision making and control, but requires extensive data collection for each new task that an agent needs to learn. In other machine learning fields, such as natural language processing or computer vision, pre-training on large, previously collected datasets to bootstrap learning for new tasks has emerged as a powerful paradigm to reduce data requirements when learning a new task. In this paper, we ask the following question: how can we enable similarly useful pre-training for RL agents? We propose a method for pre-training behavioral priors that can capture complex input-output relationships observed in successful trials from a wide range of previously seen tasks, and we show how this learned prior can be used for rapidly learning new tasks without impeding the RL agent's ability to try out novel behaviors. We demonstrate the effectiveness of our approach in challenging robotic manipulation domains involving image observations and sparse reward functions, where our method outperforms prior works by a substantial margin.

Avi Singh, Albert Yu, Jonathan Yang et al.

Reinforcement learning has been applied to a wide variety of robotics problems, but most of such applications involve collecting data from scratch for each new task. Since the amount of robot data we can collect for any single task is limited by time and cost considerations, the learned behavior is typically narrow: the policy can only execute the task in a handful of scenarios that it was trained on. What if there was a way to incorporate a large amount of prior data, either from previously solved tasks or from unsupervised or undirected environment interaction, to extend and generalize learned behaviors? While most prior work on extending robotic skills using pre-collected data focuses on building explicit hierarchies or skill decompositions, we show in this paper that we can reuse prior data to extend new skills simply through dynamic programming. We show that even when the prior data does not actually succeed at solving the new task, it can still be utilized for learning a better policy, by providing the agent with a broader understanding of the mechanics of its environment. We demonstrate the effectiveness of our approach by chaining together several behaviors seen in prior datasets for solving a new task, with our hardest experimental setting involving composing four robotic skills in a row: picking, placing, drawer opening, and grasping, where a +1/0 sparse reward is provided only on task completion. We train our policies in an end-to-end fashion, mapping high-dimensional image observations to low-level robot control commands, and present results in both simulated and real world domains. Additional materials and source code can be found on our project website: https://sites.google.com/view/cog-rl

Henry Zhu, Justin Yu, Abhishek Gupta et al.

The success of reinforcement learning for real world robotics has been, in many cases limited to instrumented laboratory scenarios, often requiring arduous human effort and oversight to enable continuous learning. In this work, we discuss the elements that are needed for a robotic learning system that can continually and autonomously improve with data collected in the real world. We propose a particular instantiation of such a system, using dexterous manipulation as our case study. Subsequently, we investigate a number of challenges that come up when learning without instrumentation. In such settings, learning must be feasible without manually designed resets, using only on-board perception, and without hand-engineered reward functions. We propose simple and scalable solutions to these challenges, and then demonstrate the efficacy of our proposed system on a set of dexterous robotic manipulation tasks, providing an in-depth analysis of the challenges associated with this learning paradigm. We demonstrate that our complete system can learn without any human intervention, acquiring a variety of vision-based skills with a real-world three-fingered hand. Results and videos can be found at https://sites.google.com/view/realworld-rl/

Avi Singh, Eric Jang, Alexander Irpan et al.

While robot learning has demonstrated promising results for enabling robots to automatically acquire new skills, a critical challenge in deploying learning-based systems is scale: acquiring enough data for the robot to effectively generalize broadly. Imitation learning, in particular, has remained a stable and powerful approach for robot learning, but critically relies on expert operators for data collection. In this work, we target this challenge, aiming to build an imitation learning system that can continuously improve through autonomous data collection, while simultaneously avoiding the explicit use of reinforcement learning, to maintain the stability, simplicity, and scalability of supervised imitation. To accomplish this, we cast the problem of imitation with autonomous improvement into a multi-task setting. We utilize the insight that, in a multi-task setting, a failed attempt at one task might represent a successful attempt at another task. This allows us to leverage the robot's own trials as demonstrations for tasks other than the one that the robot actually attempted. Using an initial dataset of multi-task demonstration data, the robot autonomously collects trials which are only sparsely labeled with a binary indication of whether the trial accomplished any useful task or not. We then embed the trials into a learned latent space of tasks, trained using only the initial demonstration dataset, to draw similarities between various trials, enabling the robot to achieve one-shot generalization to new tasks. In contrast to prior imitation learning approaches, our method can autonomously collect data with sparse supervision for continuous improvement, and in contrast to reinforcement learning algorithms, our method can effectively improve from sparse, task-agnostic reward signals.

Avi Singh, Larry Yang, Kristian Hartikainen et al.

The combination of deep neural network models and reinforcement learning algorithms can make it possible to learn policies for robotic behaviors that directly read in raw sensory inputs, such as camera images, effectively subsuming both estimation and control into one model. However, real-world applications of reinforcement learning must specify the goal of the task by means of a manually programmed reward function, which in practice requires either designing the very same perception pipeline that end-to-end reinforcement learning promises to avoid, or else instrumenting the environment with additional sensors to determine if the task has been performed successfully. In this paper, we propose an approach for removing the need for manual engineering of reward specifications by enabling a robot to learn from a modest number of examples of successful outcomes, followed by actively solicited queries, where the robot shows the user a state and asks for a label to determine whether that state represents successful completion of the task. While requesting labels for every single state would amount to asking the user to manually provide the reward signal, our method requires labels for only a tiny fraction of the states seen during training, making it an efficient and practical approach for learning skills without manually engineered rewards. We evaluate our method on real-world robotic manipulation tasks where the observations consist of images viewed by the robot's camera. In our experiments, our method effectively learns to arrange objects, place books, and drape cloth, directly from images and without any manually specified reward functions, and with only 1-4 hours of interaction with the real world.

Annie Xie, Avi Singh, Sergey Levine et al.

Reinforcement learning and planning methods require an objective or reward function that encodes the desired behavior. Yet, in practice, there is a wide range of scenarios where an objective is difficult to provide programmatically, such as tasks with visual observations involving unknown object positions or deformable objects. In these cases, prior methods use engineered problem-specific solutions, e.g., by instrumenting the environment with additional sensors to measure a proxy for the objective. Such solutions require a significant engineering effort on a per-task basis, and make it impractical for robots to continuously learn complex skills outside of laboratory settings. We aim to find a more general and scalable solution for specifying goals for robot learning in unconstrained environments. To that end, we formulate the few-shot objective learning problem, where the goal is to learn a task objective from only a few example images of successful end states for that task. We propose a simple solution to this problem: meta-learn a classifier that can recognize new goals from a few examples. We show how this approach can be used with both model-free reinforcement learning and visual model-based planning and show results in three domains: rope manipulation from images in simulation, visual navigation in a simulated 3D environment, and object arrangement into user-specified configurations on a real robot.

Justin Fu, Avi Singh, Dibya Ghosh et al.

The design of a reward function often poses a major practical challenge to real-world applications of reinforcement learning. Approaches such as inverse reinforcement learning attempt to overcome this challenge, but require expert demonstrations, which can be difficult or expensive to obtain in practice. We propose variational inverse control with events (VICE), which generalizes inverse reinforcement learning methods to cases where full demonstrations are not needed, such as when only samples of desired goal states are available. Our method is grounded in an alternative perspective on control and reinforcement learning, where an agent's goal is to maximize the probability that one or more events will happen at some point in the future, rather than maximizing cumulative rewards. We demonstrate the effectiveness of our methods on continuous control tasks, with a focus on high-dimensional observations like images where rewards are hard or even impossible to specify.

Dibya Ghosh, Avi Singh, Aravind Rajeswaran et al.

Standard model-free deep reinforcement learning (RL) algorithms sample a new initial state for each trial, allowing them to optimize policies that can perform well even in highly stochastic environments. However, problems that exhibit considerable initial state variation typically produce high-variance gradient estimates for model-free RL, making direct policy or value function optimization challenging. In this paper, we develop a novel algorithm that instead partitions the initial state space into "slices", and optimizes an ensemble of policies, each on a different slice. The ensemble is gradually unified into a single policy that can succeed on the whole state space. This approach, which we term divide-and-conquer RL, is able to solve complex tasks where conventional deep RL methods are ineffective. Our results show that divide-and-conquer RL greatly outperforms conventional policy gradient methods on challenging grasping, manipulation, and locomotion tasks, and exceeds the performance of a variety of prior methods. Videos of policies learned by our algorithm can be viewed at http://bit.ly/dnc-rl

Avi Singh, Larry Yang, Sergey Levine

We tackle the problem of learning robotic sensorimotor control policies that can generalize to visually diverse and unseen environments. Achieving broad generalization typically requires large datasets, which are difficult to obtain for task-specific interactive processes such as reinforcement learning or learning from demonstration. However, much of the visual diversity in the world can be captured through passively collected datasets of images or videos. In our method, which we refer to as GPLAC (Generalized Policy Learning with Attentional Classifier), we use both interaction data and weakly labeled image data to augment the generalization capacity of sensorimotor policies. Our method combines multitask learning on action selection and an auxiliary binary classification objective, together with a convolutional neural network architecture that uses an attentional mechanism to avoid distractors. We show that pairing interaction data from just a single environment with a diverse dataset of weakly labeled data results in greatly improved generalization to unseen environments, and show that this generalization depends on both the auxiliary objective and the attentional architecture that we propose. We demonstrate our results in both simulation and on a real robotic manipulator, and demonstrate substantial improvement over standard convolutional architectures and domain adaptation methods.

Abhishek Das, Satwik Kottur, Khushi Gupta et al.

We introduce the task of Visual Dialog, which requires an AI agent to hold a meaningful dialog with humans in natural, conversational language about visual content. Specifically, given an image, a dialog history, and a question about the image, the agent has to ground the question in image, infer context from history, and answer the question accurately. Visual Dialog is disentangled enough from a specific downstream task so as to serve as a general test of machine intelligence, while being grounded in vision enough to allow objective evaluation of individual responses and benchmark progress. We develop a novel two-person chat data-collection protocol to curate a large-scale Visual Dialog dataset (VisDial). VisDial v0.9 has been released and contains 1 dialog with 10 question-answer pairs on ~120k images from COCO, with a total of ~1.2M dialog question-answer pairs. We introduce a family of neural encoder-decoder models for Visual Dialog with 3 encoders -- Late Fusion, Hierarchical Recurrent Encoder and Memory Network -- and 2 decoders (generative and discriminative), which outperform a number of sophisticated baselines. We propose a retrieval-based evaluation protocol for Visual Dialog where the AI agent is asked to sort a set of candidate answers and evaluated on metrics such as mean-reciprocal-rank of human response. We quantify gap between machine and human performance on the Visual Dialog task via human studies. Putting it all together, we demonstrate the first 'visual chatbot'! Our dataset, code, trained models and visual chatbot are available on https://visualdialog.org

Ashesh Jain, Hema S Koppula, Shane Soh et al.

Advanced Driver Assistance Systems (ADAS) have made driving safer over the last decade. They prepare vehicles for unsafe road conditions and alert drivers if they perform a dangerous maneuver. However, many accidents are unavoidable because by the time drivers are alerted, it is already too late. Anticipating maneuvers beforehand can alert drivers before they perform the maneuver and also give ADAS more time to avoid or prepare for the danger. In this work we propose a vehicular sensor-rich platform and learning algorithms for maneuver anticipation. For this purpose we equip a car with cameras, Global Positioning System (GPS), and a computing device to capture the driving context from both inside and outside of the car. In order to anticipate maneuvers, we propose a sensory-fusion deep learning architecture which jointly learns to anticipate and fuse multiple sensory streams. Our architecture consists of Recurrent Neural Networks (RNNs) that use Long Short-Term Memory (LSTM) units to capture long temporal dependencies. We propose a novel training procedure which allows the network to predict the future given only a partial temporal context. We introduce a diverse data set with 1180 miles of natural freeway and city driving, and show that we can anticipate maneuvers 3.5 seconds before they occur in real-time with a precision and recall of 90.5\% and 87.4\% respectively.

Ashesh Jain, Avi Singh, Hema S Koppula et al.

Anticipating the future actions of a human is a widely studied problem in robotics that requires spatio-temporal reasoning. In this work we propose a deep learning approach for anticipation in sensory-rich robotics applications. We introduce a sensory-fusion architecture which jointly learns to anticipate and fuse information from multiple sensory streams. Our architecture consists of Recurrent Neural Networks (RNNs) that use Long Short-Term Memory (LSTM) units to capture long temporal dependencies. We train our architecture in a sequence-to-sequence prediction manner, and it explicitly learns to predict the future given only a partial temporal context. We further introduce a novel loss layer for anticipation which prevents over-fitting and encourages early anticipation. We use our architecture to anticipate driving maneuvers several seconds before they happen on a natural driving data set of 1180 miles. The context for maneuver anticipation comes from multiple sensors installed on the vehicle. Our approach shows significant improvement over the state-of-the-art in maneuver anticipation by increasing the precision from 77.4% to 90.5% and recall from 71.2% to 87.4%.