Developing agents that can follow multimodal instructions remains a fundamental challenge in robotics and AI. Although large-scale pre-training on unlabeled datasets (no language instruction) has enabled agents to learn diverse behaviors, these agents often struggle with following instructions. While augmenting the dataset with instruction labels can mitigate this issue, acquiring such high-quality annotations at scale is impractical. To address this issue, we frame the problem as a semi-supervised learning task and introduce GROOT-2, a multimodal instructable agent trained using a novel approach that combines weak supervision with latent variable models. Our method consists of two key components: constrained self-imitating, which utilizes large amounts of unlabeled demonstrations to enable the policy to learn diverse behaviors, and human intention alignment, which uses a smaller set of labeled demonstrations to ensure the latent space reflects human intentions. GROOT-2's effectiveness is validated across four diverse environments, ranging from video games to robotic manipulation, demonstrating its robust multimodal instruction-following capabilities.

Species range maps (SRMs) are essential tools for research and policy-making in ecology, conservation, and environmental management. However, traditional SRMs rely on the availability of environmental covariates and high-quality species location observation data, both of which can be challenging to obtain due to geographic inaccessibility and resource constraints. We propose a novel approach combining millions of citizen science species observations with textual descriptions from Wikipedia, covering habitat preferences and range descriptions for tens of thousands of species. Our framework maps locations, species, and text descriptions into a common space, facilitating the learning of rich spatial covariates at a global scale and enabling zero-shot range estimation from textual descriptions. Evaluated on held-out species, our zero-shot SRMs significantly outperform baselines and match the performance of SRMs obtained using tens of observations. Our approach also acts as a strong prior when combined with observational data, resulting in more accurate range estimation with less data. We present extensive quantitative and qualitative analyses of the learned representations in the context of range estimation and other spatial tasks, demonstrating the effectiveness of our approach.

Rewards remain an uninterpretable way to specify tasks for Reinforcement Learning, as humans are often unable to predict the optimal behavior of any given reward function, leading to poor reward design and reward hacking. Language presents an appealing way to communicate intent to agents and bypass reward design, but prior efforts to do so have been limited by costly and unscalable labeling efforts. In this work, we propose a method for a completely unsupervised alternative to grounding language instructions in a zero-shot manner to obtain policies. We present a solution that takes the form of imagine, project, and imitate: The agent imagines the observation sequence corresponding to the language description of a task, projects the imagined sequence to our target domain, and grounds it to a policy. Video-language models allow us to imagine task descriptions that leverage knowledge of tasks learned from internet-scale video-text mappings. The challenge remains to ground these generations to a policy. In this work, we show that we can achieve a zero-shot language-to-behavior policy by first grounding the imagined sequences in real observations of an unsupervised RL agent and using a closed-form solution to imitation learning that allows the RL agent to mimic the grounded observations. Our method, RLZero, is the first to our knowledge to show zero-shot language to behavior generation abilities without any supervision on a variety of tasks on simulated domains. We further show that RLZero can also generate policies zero-shot from cross-embodied videos such as those scraped from YouTube.

We explore the Iterative Inference Hypothesis (IIH) within the context of transformer-based language models, aiming to understand how a model's latent representations are progressively refined and whether observable differences are present between correct and incorrect generations. Our findings provide empirical support for the IIH, showing that the nth token embedding in the residual stream follows a trajectory of decreasing loss. Additionally, we observe that the rate at which residual embeddings converge to a stable output representation reflects uncertainty in the token generation process. Finally, we introduce a method utilizing cross-entropy to detect this uncertainty and demonstrate its potential to distinguish between correct and incorrect token generations on a dataset of idioms.

This paper introduces M$^{3}$-20M, a large-scale Multi-Modal Molecular dataset that contains over 20 million molecules. Designed to support AI-driven drug design and discovery, M$^{3}$-20M is 71 times more in the number of molecules than the largest existing dataset, providing an unprecedented scale that can highly benefit training or fine-tuning large (language) models with superior performance for drug design and discovery. This dataset integrates one-dimensional SMILES, two-dimensional molecular graphs, three-dimensional molecular structures, physicochemical properties, and textual descriptions collected through web crawling and generated by using GPT-3.5, offering a comprehensive view of each molecule. To demonstrate the power of M$^{3}$-20M in drug design and discovery, we conduct extensive experiments on two key tasks: molecule generation and molecular property prediction, using large language models including GLM4, GPT-3.5, and GPT-4. Our experimental results show that M$^{3}$-20M can significantly boost model performance in both tasks. Specifically, it enables the models to generate more diverse and valid molecular structures and achieve higher property prediction accuracy than the existing single-modal datasets, which validates the value and potential of M$^{3}$-20M in supporting AI-driven drug design and discovery. The dataset is available at \url{https://github.com/bz99bz/M-3}.

The paper describes LLM unlearning without a retaining dataset, using the ORPO reinforcement learning method with inference enhanced by modified classifier-free guidance. Significant improvement in unlearning, without degradation of the model, is achieved through direct training on synthetic replacement data in CFG-aware training regime, with classifier-free guidance applied during the inference. This article is an extended version of the NeurIPS 2024 LLM-PC submission, which was awarded second prize.

Advancements in deep learning have demonstrated that the performance of neural networks scales predictably as a function of model size, data size, and computational resources [Hestness et al., 2017, Kaplan et al., 2020a, Rosenfeld et al., 2020, Henighan et al., 2020a]. These trends, known as neural scaling laws, have motivated research into understanding how scaling influences model performance in a range of domains, in particular, Large Language Models (LLMs) [Brown et al., 2020, Hoffmann et al., 2022]. However, scaling during inference--the process by which a trained model makes predictions on new data--has received less attention. Recent works have shown empirically that LLMs can gain substantial benefits from repeated prompts to perform better on difficult tasks such as coding and formal proofs, where verification of the correct answer can be done [Brown et al., 2024, Snell et al., 2024, Bansal et al., 2024]. These works demonstrate that the performance of weaker models can be amplified without further training, by simply repeating inference trials. A natural question then arises: Can we interpret, or predict the inference scaling behavior of a model with repeated attempts? To answer this question, we propose a simple toy model that isolates the inference scaling laws which dictate how certain performance metrics improve as a function of the number of inference attempts. Inspired by the work of Hutter [2021], which introduced a model to study scaling behavior for memorization and generalization, we devise a simple setting to capture the effect of repeated inference attempts, focusing on the coverage metric, also known as pass@k. In this work, we present analytical predictions for coverage from a probabilistic perspective and demonstrate how inference improves with the number of repeated trials in a predictable way, which matches the observed behavior in Brown et al. [2024] and Snell et al. [2024].

The commonsense reasoning capabilities of vision-language models (VLMs), especially in abductive reasoning and defeasible reasoning, remain poorly understood. Most benchmarks focus on typical visual scenarios, making it difficult to discern whether model performance stems from keen perception and reasoning skills, or reliance on pure statistical recall. We argue that by focusing on atypical events in videos, clearer insights can be gained on the core capabilities of VLMs. Explaining and understanding such out-of-distribution events requires models to extend beyond basic pattern recognition and regurgitation of their prior knowledge. To this end, we introduce BlackSwanSuite, a benchmark for evaluating VLMs' ability to reason about unexpected events through abductive and defeasible tasks. Our tasks artificially limit the amount of visual information provided to models while questioning them about hidden unexpected events, or provide new visual information that could change an existing hypothesis about the event. We curate a comprehensive benchmark suite comprising over 3,800 MCQ, 4,900 generative and 6,700 yes/no tasks, spanning 1,655 videos. After extensively evaluating various state-of-the-art VLMs, including GPT-4o and Gemini 1.5 Pro, as well as open-source VLMs such as LLaVA-Video, we find significant performance gaps of up to 32% from humans on these tasks. Our findings reveal key limitations in current VLMs, emphasizing the need for enhanced model architectures and training strategies.

Large language models (LLMs) have transformed natural language processing, yet face challenges in specialized tasks such as simulating opinions on environmental policies. This paper introduces a novel fine-tuning approach that integrates socio-demographic data from the UK Household Longitudinal Study, uniquely using profiling factors, such as age, gender, income, education, and region. This method enhances the accuracy and representation of generated views. By emulating diverse synthetic profiles, the fine-tuned models significantly outperform pre-trained counterparts, achieving measurable improvements in capturing demographic nuances. Evaluation metrics, including Chi-Squared, Cosine Similarity, Jaccard Index, and KL-divergence, reveal a strong alignment between synthetic and real-world opinions. This work demonstrates the potential of fine-tuned LLMs tailored to societal contexts to enable more ethical and precise policy simulations. Its broader implications include deploying LLMs in domains like healthcare and education, fostering inclusive and data-driven decision-making in both research and practice.

This week, OpenAI launched what its chief executive, Sam Altman, called "the smartest model in the world"--a generative-AI program whose capabilities are supposedly far greater, and more closely approximate how humans think, than those of any such software preceding it. The start-up has been building toward this moment since September 12, a day that, in OpenAI's telling, set the world on a new path toward superintelligence. That was when the company previewed early versions of a series of AI models, known as o1, constructed with novel methods that the start-up believes will propel its programs to unseen heights. Mark Chen, then OpenAI's vice president of research, told me a few days later that o1 is fundamentally different from the standard ChatGPT because it can "reason," a hallmark of human intelligence. Shortly thereafter, Altman pronounced "the dawn of the Intelligence Age," in which AI helps humankind fix the climate and colonize space. As of yesterday afternoon, the start-up has released the first complete version of o1, with fully fledged reasoning powers, to the public.