As the general capabilities of artificial intelligence (AI) agents continue to evolve, their ability to learn to master multiple complex tasks through experience remains a key challenge. Current LLM agents, particularly those based on proprietary language models, typically rely on prompts to incorporate knowledge about the target tasks. This approach does not allow the agent to internalize this information and instead relies on ever-expanding prompts to sustain its functionality in diverse scenarios. This resembles a system of notes used by a person affected by anterograde amnesia, the inability to form new memories. In this paper, we propose a novel method to train AI agents to incorporate knowledge and skills for multiple tasks without the need for either cumbersome note systems or prior high-quality demonstration data. Our approach employs an iterative process where the agent collects new experiences, receives corrective feedback from humans in the form of hints, and integrates this feedback into its weights via a context distillation training procedure. We demonstrate the efficacy of our approach by implementing it in a Llama-3-based agent which, after only a few rounds of feedback, outperforms advanced models GPT-4o and DeepSeek-V3 in a taskset requiring correct sequencing of information retrieval, tool use, and question answering.

Sample inefficiency is a long-lasting challenge in deep reinforcement learning (DRL). Despite dramatic improvements have been made, the problem is far from being solved and is especially challenging in environments with sparse or delayed rewards. In our work, we propose to use Adversarial Estimates as a new, simple and efficient approach to mitigate this problem for a class of feedback-based DRL algorithms. Our approach leverages latent similarity search from a small set of human-collected trajectories to boost learning, using only five minutes of human-recorded experience. The results of our study show algorithms trained with Adversarial Estimates converge faster than their original version. Moreover, we discuss how our approach could enable learning in feedback-based algorithms in extreme scenarios with very sparse rewards.

Sampling-based search, a simple paradigm for utilizing test-time compute, involves generating multiple candidate responses and selecting the best one--typically by having models self-verify each response for correctness. In this paper, we study the scaling trends governing sampling-based search. Among our findings is that simply scaling up a minimalist implementation of sampling-based search, using only random sampling and direct self-verification, provides a practical inference method that, for example, elevates the reasoning capabilities of Gemini v1.5 Pro above that of o1-Preview on popular benchmarks. We partially attribute the scalability of sampling-based search to a phenomenon of implicit scaling, where sampling a larger pool of responses in turn improves self-verification accuracy. We further identify two useful principles for improving self-verification capabilities with test-time compute: (1) comparing across responses provides helpful signals about the locations of errors and hallucinations, and (2) different model output styles are useful for different contexts--chains of thought are useful for reasoning but harder to verify. We also find that, though accurate verification can be elicited, frontier models demonstrate remarkably weak out-of-box verification capabilities and introduce a benchmark to measure progress on these deficiencies.

Algorithms are increasingly used to automate or aid human decisions, yet recent research shows that these algorithms may exhibit bias across legally protected demographic groups. However, data on these groups may be unavailable to organizations or external auditors due to privacy legislation. This paper studies bias detection using an unsupervised clustering tool when data on demographic groups are unavailable. We collaborate with the Dutch Executive Agency for Education to audit an algorithm that was used to assign risk scores to college students at the national level in the Netherlands between 2012-2023. Our audit covers more than 250,000 students from the whole country. The unsupervised clustering tool highlights known disparities between students with a non-European migration background and Dutch origin. Our contributions are three-fold: (1) we assess bias in a real-world, large-scale and high-stakes decision-making process by a governmental organization; (2) we use simulation studies to highlight potential pitfalls of using the unsupervised clustering tool to detect true bias when demographic group data are unavailable and provide recommendations for valid inferences; (3) we provide the unsupervised clustering tool in an open-source library. Our work serves as a starting point for a deliberative assessment by human experts to evaluate potential discrimination in algorithmic-supported decision-making processes.

Listeners adapt language comprehension based on their mental representations of speakers, but how these representations are dynamically updated remains unclear. We investigated whether listeners probabilistically adapt their comprehension based on the likelihood of speakers producing stereotype-incongruent utterances. Our findings reveal two potential mechanisms: a speaker-general mechanism that adjusts overall expectations about speaker-content relationships, and a speaker-specific mechanism that updates individual speaker models. In two EEG experiments, participants heard speakers make stereotype-congruent or incongruent utterances, with incongruency base rate manipulated between blocks. In Experiment 1, speaker incongruency modulated both high-beta (21-30 Hz) and theta (4-6 Hz) oscillations: incongruent utterances decreased oscillatory power in low base rate condition but increased it in high base rate condition. The theta effect varied with listeners' openness trait: less open participants showed theta increases to speaker-incongruencies, suggesting maintenance of speaker-specific information, while more open participants showed theta decreases, indicating flexible model updating. In Experiment 2, we dissociated base rate from the target speaker by manipulating the overall base rate using an alternative non-target speaker. Only the high-beta effect persisted, showing power decrease for speaker-incongruencies in low base rate condition but no effect in high base rate condition. The high-beta oscillations might reflect the speaker-general adjustment, while theta oscillations may index the speaker-specific model updating. These findings provide evidence for how language processing is shaped by social cognition in real time.

In our evaluation, we assess the performance of GPT-[n] and o-[n] models on abstract multimodal puzzles from PuzzleVQA, which primarily test abstract reasoning. Additionally, we evaluate the models on AlgoPuzzleVQA, which require an algorithmic approach rather than brute-force solving. To ensure a comprehensive evaluation, we present the puzzles in both multiple-choice and openended question answering formats. Our findings indicate that despite their sophisticated capabilities in standard benchmarks, current models still struggle with seemingly simple multimodal puzzles (Figure 3). Contrary to previous benchmarks such as ARC-AGI, we observe a less dramatic reasoning curve without extreme jumps in performance. This limitation highlights the substantial gap between current artificial intelligence and human-like reasoning abilities. As the models continue to rapidly advance and scale as in Figure 1, this benchmark will serve as a critical indicator of progress toward more robust and generalized artificial intelligence. Overall, here are the key findings of our study: TL;DR 1. Performance steadily improves from GPT-4-Turbo to GPT-4o to o1. While the jump from GPT-4-Turbo to GPT-4o is moderate, the transition from GPT-4o to o1 marks a significant advancement but it comes at a cost of 750x more inference cost.

Plasticity loss, a critical challenge in neural network training, limits a model's ability to adapt to new tasks or shifts in data distribution. This paper introduces AID (Activation by Interval-wise Dropout), a novel method inspired by Dropout, designed to address plasticity loss. Unlike Dropout, AID generates subnetworks by applying Dropout with different probabilities on each preactivation interval. Theoretical analysis reveals that AID regularizes the network, promoting behavior analogous to that of deep linear networks, which do not suffer from plasticity loss. We validate the effectiveness of AID in maintaining plasticity across various benchmarks, including continual learning tasks on standard image classification datasets such as CIFAR10, CIFAR100, and TinyImageNet. Furthermore, we show that AID enhances reinforcement learning performance in the Arcade Learning Environment benchmark.

These authors contribute equally to this work. Abstract Artificial neural networks face the stability-plasticity dilemma in continual learning, while the brain can maintain memories and remain adaptable. However, the biological strategies for continual learning and their potential to inspire learning algorithms in neural networks are poorly understood. This study presents a minimal model of the fly olfactory circuit to investigate the biological strategies that support continual odor learning. We introduce the fly olfactory circuit as a plug-and-play component, termed the Fly Model, which can integrate with modern machine learning methods to address this dilemma. Our findings demonstrate that the Fly Model enhances both memory stability and learning plasticity, overcoming the limitations of current continual learning strategies. We validated its effectiveness across various challenging continual learning scenarios using commonly used datasets. The fly olfactory system serves as an elegant biological circuit for lifelong learning, offering a module that enhances continual learning with minimal additional computational cost for machine learning. When learning new tasks and updating parameters, these models inevitably overwrite previously learned patterns, resulting in "catastrophic forgetting" [1-3]. This critical flaw has become the Achilles' heel of neural network models, preventing them from realizing their full potential. Conversely, especially under long non-stationary data streams, the parameters of network models may become less effective at updating, resulting in a gradual decline in their ability to adapt to new information. This issue, known as plasticity loss, has garnered increasing attention in recent years [4-5].

Retrieval-Augmented Generation (RAG) has demonstrated remarkable success in enhancing Large Language Models (LLMs) through external knowledge integration, yet its application has primarily focused on textual content, leaving the rich domain of multi-modal video knowledge predominantly unexplored. This paper introduces VideoRAG, the first retrieval-augmented generation framework specifically designed for processing and understanding extremely long-context videos. Our core innovation lies in its dual-channel architecture that seamlessly integrates (i) graph-based textual knowledge grounding for capturing cross-video semantic relationships, and (ii) multi-modal context encoding for efficiently preserving visual features. This novel design empowers VideoRAG to process unlimited-length videos by constructing precise knowledge graphs that span multiple videos while maintaining semantic dependencies through specialized multi-modal retrieval paradigms. Through comprehensive empirical evaluation on our proposed LongerVideos benchmark-comprising over 160 videos totaling 134+ hours across lecture, documentary, and entertainment categories-VideoRAG demonstrates substantial performance compared to existing RAG alternatives and long video understanding methods. The source code of VideoRAG implementation and the benchmark dataset are openly available at: https://github.com/HKUDS/VideoRAG.

The paradigm of weak-to-strong generalization constitutes the training of a strong AI model on data labeled by a weak AI model, with the goal that the strong model nevertheless outperforms its weak supervisor on the target task of interest. For the setting of real-valued regression with the squared loss, recent work quantitatively characterizes the gain in performance of the strong model over the weak model in terms of the misfit between the strong and weak model. We generalize such a characterization to learning tasks whose loss functions correspond to arbitrary Bregman divergences when the strong class is convex. This extends the misfit-based characterization of performance gain in weak-to-strong generalization to classification tasks, as the cross-entropy loss can be expressed in terms of a Bregman divergence. In most practical scenarios, however, the strong model class may not be convex. We therefore weaken this assumption and study weak-to-strong generalization for convex combinations of $k$ strong models in the strong class, in the concrete setting of classification. This allows us to obtain a similar misfit-based characterization of performance gain, upto an additional error term that vanishes as $k$ gets large. Our theoretical findings are supported by thorough experiments on synthetic as well as real-world datasets.