What is the best compromise in a situation where different people value different things? The most common method for answering this question is to look at all the options, add up the utility per person associated with each, and pick the option with the largest sum. This approach seems like the obvious, theory-neutral solution, though, in fact, it relies on substantive utilitarian commitments. A second, quite common approach, involves integrating egalitarian concerns, and picking an option that trades off maximizing utility with equalizing outcomes. But there is an important, though often-ignored, third approach as well: a contractualist approach, which models what rational actors would agree to. In this paper, we test the proposals of these three competing preference aggregation algorithms in social decision-making contexts that are systematically sampled across a wide range. While the dominant approach to value aggregation up to now has been utilitarian, we find that participants strongly prefer the aggregations recommended by the contractualist algorithm, which also dominates the egalitarian approach across a range of parameterizations. Finally, we compare the judgments of large language models (LLMs) to those of our (human) participants, finding important misalignment between model and human preferences.
AI agents may soon become capable of autonomously completing valuable, long-horizon tasks in diverse domains. Current benchmarks either do not measure real-world tasks, or are not sufficiently difficult to meaningfully measure frontier models. To this end, we present Terminal-Bench 2.0: a carefully curated hard benchmark composed of 89 tasks in computer terminal environments inspired by problems from real workflows. Each task features a unique environment, human-written solution, and comprehensive tests for verification. We show that frontier models and agents score less than 65% on the benchmark and conduct an error analysis to identify areas for model and agent improvement. We publish the dataset and evaluation harness to assist developers and researchers in future work at tbench.ai.
Reasoning is essential for purposeful action, yet most robotic foundation models map perception and instructions directly to control, limiting adaptability, generalization, and semantic grounding. We introduce Action Reasoning Models (ARMs), which integrate perception, planning, and control through a structured three-stage pipeline. Our model, Molmoact, encodes observations and instructions into depth perception tokens, generates 2D spatial plans, and predicts fine-grained actions, enabling explainable and steerable behavior. Molmoact-7B-D achieves 70.5% zero-shot accuracy on SimplerEnv Visual Matching (surpassing Pi-0 and GR00T N1.5), 86.6% average success on LIBERO, and real-world fine-tuning gains of +10% (single-arm) and +22.7% (bimanual) over Pi-0-FAST. It further improves out-of-distribution generalization by +23.3% and ranks highest in human-preference evaluations for open-instruction following and trajectory steering. We also release Molmoact Dataset, a dataset of 10k diverse robot trajectories that yields an average +5.5% performance boost when used for training. Together with open model weights and code, this establishes Molmoact as a state-of-the-art robotic foundation model and an open blueprint for building ARMs that transform perception into grounded, purposeful action. Further experimental details and result with Molmoact Dataset and human-preference evaluations included in supplementary video.
Today’s strongest video-language models (VLMs) remain proprietary.The strongest open-weight models either rely on synthetic data from proprietary VLMs, effectively distilling from them, or do not disclose their training data or recipe.As a result, the open-source community lacks the foundations needed to improve on the state-of-the-art video (and image) language models.Crucially, many downstream applications require more than just high-level video understanding; they require grounding—either by pointing or by tracking in pixels. Even proprietary models lack this capability.We present Molmo2, a new family of VLMs that are state-of-the-art amongst open-source models and demonstrate exceptional new capabilities in point-driven grounding in single image, multi-image, and video tasks.Our key contribution is a collection of 7 new video datasets and 2 multi-image datasets, including a dataset of highly detailed video captions for pre-training, a free-form video Q&A dataset for fine-tuning, a new object tracking dataset with complex queries, and an innovative new video pointing dataset, all collected without the use of closed VLMs.We also present a training recipe for this data utilizing an efficient packing and message-tree encoding scheme and show bi-directional attention on vision tokens and a novel token-weight strategy improve performance. Our best-in-class 8B model outperforms others in the class of open weight and data models on short videos, counting, and captioning, and is competitive on long-videos. On video-grounding Molmo2 outperforms larger proprietary models, including 32.9% (Molmo2) vs 17% (Gemini 2.5 Pro) on video pointing.
Deep research agents perform multi-step research to produce long-form, well-attributed answers. However, most open deep research agents are trained on easily verifiable short-form QA tasks via reinforcement learning with verifiable rewards, which does not extend to realistic long-form tasks. We address this with , where rubrics are constructed and maintained to with the policy model during training. This allows the rubrics to incorporate newly explored information from search and contrasting model responses, enabling better fact checking and more discriminative on-policy feedback. Using RLER, we develop , the first fully open model that is directly trained for open-ended, long-form deep research. Across four long-form deep research benchmarks in science, healthcare, and general domains, DR Tulu-8B substantially outperforms existing open deep research agents (by 15.6% over Tongyi DR on average) and matches or exceeds proprietary deep research agents (by 0.7% over OpenAI DR on average), while being significantly smaller and cheaper per query (1000x cheaper than OpenAI DR per query).
