We introduce Refeeding State Embeddings aligned using Environmental Data (RESEED), a novel method for grounding language in environmental data.While large language models (LLMs) excel at many tasks, they continue to struggle with multi-step sequential reasoning.RESEED addresses this by producing latent embeddings aligned with the true state of the environment and refeeding these embeddings into the model before generating its output.To evaluate its effectiveness, we develop three new sequential reasoning benchmarks, each with a training set of paired state-text trajectories and several text-only evaluation sets that test generalization to longer trajectories.Across all benchmarks, RESEED significantly improves generalization and scalability over a text-only baseline.We further show that RE-SEED outperforms commercial LLMs on our benchmarks, highlighting the value of grounding language in the environment.1

Accommodating human preferences is essential for creating aligned LLM agents that deliver personalized and effective interactions. Recent work has shown the potential for LLMs acting as writing agents to infer a description of user preferences. Agent alignment then comes from conditioning on the inferred preference description. However, existing methods often produce generic preference descriptions that fail to capture the unique and individualized nature of human preferences. This paper introduces PROSE, a method designed to enhance the precision of preference descriptions inferred from user writing samples. PROSE incorporates two key elements: (1) iterative refinement of inferred preferences, and (2) verification of inferred preferences across multiple user writing samples. We evaluate PROSE with several LLMs (i.e., Qwen2.5 7B and 72B Instruct, GPT-mini, and GPT-4o) on a summarization and an email writing task. We find that PROSE more accurately infers nuanced human preferences, improving the quality of the writing agent’s generations over CIPHER (a state-of-the-art method for inferring preferences) by 33%. Lastly, we demonstrate that ICL and PROSE are complementary methods, and combining them provides up to a 9% improvement over ICL alone. Code: https://github.com/apple/ml-predict

In collaborative tasks, autonomous agents fall short of humans in their capability to quickly adapt to new and unfamiliar teammates. We posit that a limiting factor for zero-shot coordination is the lack of shared task abstractions, a mechanism humans rely on to implicitly align with teammates. To address this gap, we introduce HA^2: Hierarchical Ad Hoc Agents, a framework leveraging hierarchical reinforcement learning to mimic the structured approach humans use in collaboration. We evaluate HA^2 in the Overcooked environment, demonstrating statistically significant improvement over existing baselines when paired with both unseen agents and humans, providing better resilience to environmental shifts, and outperforming all state-of-the-art methods.

Effective collaboration between humans and AIs hinges on transparent communication and alignment of mental models. However, explicit, verbal communication is not always feasible. Under such circumstances, human-human teams often depend on implicit, nonverbal cues to glean important information about their teammates such as intent and expertise, thereby bolstering team alignment and adaptability. Among these implicit cues, two of the most salient and fundamental are a human’s actions in the environment and their visual attention. In this paper, we present a novel method to combine eye gaze data and behavioral data, and evaluate their respective predictive power for human proficiency, trust, and intent. We first collect a dataset of paired eye gaze and gameplay data in the fast-paced collaborative "Overcooked" environment. We then train models on this dataset to compare how the predictive powers differ between gaze data, gameplay data, and their combination. We additionally compare our method to prior works that aggregate eye gaze data and demonstrate how these aggregation methods can substantially reduce the predictive ability of eye gaze. Our results indicate that, while eye gaze data and gameplay data excel in different situations, a model that integrates both types consistently outperforms all baselines. This work paves the way for developing intuitive and responsive agents that can efficiently adapt to new teammates.

Recent work has raised concerns about the inherent limitations of text-only pretraining.In this paper, we first demonstrate that reporting bias, the tendency of people to not state the obvious, is one of the causes of this limitation, and then investigate to what extent multimodal training can mitigate this issue.To accomplish this, we 1) generate the Color Dataset (CoDa), a dataset of human-perceived color distributions for 521 common objects; 2) use CoDa to analyze and compare the color distribution found in text, the distribution captured by language models, and a human's perception of color; and 3) investigate the performance differences between text-only and multimodal models on CoDa.Our results show that the distribution of colors that a language model recovers correlates more strongly with the inaccurate distribution found in text than with the ground-truth, supporting the claim that reporting bias negatively impacts and inherently limits text-only training.We then demonstrate that multimodal models can leverage their visual training to mitigate these effects, providing a promising avenue for future research.* *Email has no accent, but includes the hyphen. 1 In this paper, we use LM to refer to both causal LMs as well as masked LMs.

BERT set many state-of-the-art results over varied NLU benchmarks by pre-training over two tasks: masked language modelling (MLM) and next sentence prediction (NSP), the latter of which has been highly criticized.In this paper, we 1) clarify NSP's effect on BERT pre-training, 2) explore fourteen possible auxiliary pre-training tasks, of which seven are novel to modern language models, and 3) investigate different ways to include multiple tasks into pre-training.We show that NSP is detrimental to training due to its context splitting and shallow semantic signal.We also identify six auxiliary pre-training tasks -sentence ordering, adjacent sentence prediction, TF prediction, TF-IDF prediction, a Fast-Sent variant, and a Quick Thoughts variant -that outperform a pure MLM baseline.Finally, we demonstrate that using multiple tasks in a multi-task pre-training framework provides better results than using any single auxiliary task.Using these methods, we outperform BERT Base on the GLUE benchmark using fewer than a quarter of the training tokens.