Search

Today's most advanced multimodal models remain proprietary. The strongest open-weight models rely heavily on synthetic data from proprietary VLMs to achieve good performance, effectively distilling these closed models into open ones. As a result, the community is still missing foundational knowledge about how to build performant VLMs from scratch. We present Molmo, a new family of VLMs that are state-of-the-art in their class of openness. Our key innovation is a novel, highly detailed image caption dataset collected entirely from human annotators using speech-based descriptions. To enable a wide array of user interactions, we also introduce a diverse dataset mixture for fine-tuning that includes in-the-wild Q&A and innovative 2D pointing data. The success of our approach relies on careful choices for the model architecture details, a well-tuned training pipeline, and, most critically, the quality of our newly collected datasets, all of which will be released. The best-in-class 72B model within the Molmo family not only outperforms others in the class of open weight and data models but also compares favorably against proprietary systems like GPT-4o, Claude 3.5, and Gemini 1.5 on both academic benchmarks and human evaluation. We will be releasing all of our model weights, captioning and fine-tuning data, and source code in the near future. Select model weights, inference code, and demo are available at https://molmo.allenai.org.

This is the second of two papers investigating for which positive integers $m$ there exists a maximal antichain of size $m$ in the Boolean lattice $B_n$ (the power set of $[n]:=\{1,2,\dots,n\}$, ordered by inclusion). In the first part, the sizes of maximal antichains have been characterized. Here we provide an alternative construction with the benefit of showing that almost all sizes of maximal antichains can be obtained using antichains containing only $l$-sets and $(l+1)$-sets for some $l$.

The quantum Hall (QH) effect in two-dimensional electron systems (2DESs) is conventionally observed at liquid-helium temperatures, where lattice vibrations are strongly suppressed and bulk carrier scattering is dominated by disorder. However, due to large Landau level (LL) separation (~2000 K at B = 30 T), graphene can support the QH effect up to room temperature (RT), concomitant with a non-negligible population of acoustic phonons with a wave-vector commensurate to the inverse electronic magnetic length. Here, we demonstrate that graphene encapsulated in hexagonal boron nitride (hBN) realizes a novel transport regime, where dissipation in the QH phase is governed predominantly by electron-phonon scattering. Investigating thermally-activated transport at filling factor 2 up to RT in an ensemble of back-gated devices, we show that the high B-field behaviour correlates with their zero B-field transport mobility. By this means, we extend the well-accepted notion of phonon-limited resistivity in ultra-clean graphene to a hitherto unexplored high-field realm., Comment: 17 pages, 4 figures. Supplementary information available at https://doi.org/10.1038/s41467-023-35986-3

We present and analyze the employment of the Diproche system, a natural language proof checker, within a one-semester mathematics beginners lecture with 228 participants. The system is used to check the students' solution attempts to proving exercises in Boolean set theory and elementary number theory and to give them immediate feedback. The benefits of the employment of the system are assessed via a questionnaire at the end of the semester and via analyzing the solution attempts of a subgroup of the students. Based on our results we develop approaches for future improvements., Comment: In Proceedings ThEdu'21, arXiv:2202.02144

This is the second in a sequence of three papers investigating the question for which positive integers $m$ there exists a maximal antichain of size $m$ in the Boolean lattice $B_n$ (the power set of $[n]:=\{1,2,\dots,n\}$, ordered by inclusion). In the previous paper we characterized those $m$ between $\binom{n}{\lceil n/2\rceil}-\lceil n/2\rceil^2$ and the maximum size $\binom{n}{\lceil n/2 \rceil}$ that are not sizes of maximal antichains. In this paper we show that all smaller $m$ are sizes of maximal antichains., Comment: This paper has been merged with arXiv:2106.02226

Extending a classical theorem of Sperner, we characterize the integers $m$ such that there exists a maximal antichain of size $m$ in the Boolean lattice $B_n$, that is, the power set of $[n]:=\{1,2,\dots,n\}$, ordered by inclusion. As an important ingredient in the proof, we initiate the study of an extension of the Kruskal-Katona theorem which is of independent interest. For given positive integers $t$ and $k$, we ask which integers $s$ have the property that there exists a family $\mathcal F$ of $k$-sets with $\lvert\mathcal F\rvert=t$ such that the shadow of $\mathcal F$ has size $s$, where the shadow of $\mathcal F$ is the collection of $(k-1)$-sets that are contained in at least one member of $\mathcal F$. We provide a complete answer for $t\leqslant k+1$. Moreover, we prove that the largest integer which is not the shadow size of any family of $k$-sets is $\sqrt 2k^{3/2}+\sqrt[4]{8}k^{5/4}+O(k)$., Comment: This is a merger of arXiv:2106.02226v2 with arXiv:2106.02230

We describe in dialogue form a possible way of discovering and investigating 10-adic numbers starting from the naive question about a `largest natural number'. Among the topics we pursue are possibilities of extensions to transfinite 10-adic numbers, 10-adic representations of rational numbers, zero divisors, square roots and 10-adic roots of higher degree of natural numbers, and applications of 10-adic number representation in computer arithmetic. The participants of the dialogue are idealized embodiments of different philosophical attitudes towards mathematics. The article aims at illustrating how these attitudes interact, in both jarring and stimulating ways, and how they impact mathematical development., Comment: 28 pages

Everybody knows from school how to solve a quadratic equation of the form $x^2-px+q=0$ graphically. But this method can become tedious if several equations ought to be solved, as for each pair $(p,q)$ a new parabola has to be drawn. Stunningly, there is one single curve that can be used to solve every quadratic equation via drawing tangent lines through a given point $(p,q)$ to this curve. In this article we derive this method in an elementary way and generalize it to equations of the form $x^n-px+q=0$ for arbitrary $n \ge 2$. Moreover, the number of solutions of a specific equation of this form can be seen immediately with this technique. Concluding the article we point out connections to the duality of points and lines in the plane and to the the concept of Legendre transformation., Comment: 12 pages, 9 figures

"No two rainbows are the same. Neither are two packs of Skittles. Enjoy an odd mix!". Using an interpretation via spatial random walks, we quantify the probability that two randomly selected packs of Skittles candy are identical and determine the expected number of packs one has to purchase until the first match. We believe this problem to be appealing for middle and high school students as well as undergraduate students at University., Comment: 13 pages, 3 figures

AI has achieved remarkable mastery over games such as Chess, Go, and Poker, and even Jeopardy, but the rich variety of standardized exams has remained a landmark challenge. Even in 2016, the best AI system achieved merely 59.3% on an 8th Grade science exam challenge. This paper reports unprecedented success on the Grade 8 New York Regents Science Exam, where for the first time a system scores more than 90% on the exam's non-diagram, multiple choice (NDMC) questions. In addition, our Aristo system, building upon the success of recent language models, exceeded 83% on the corresponding Grade 12 Science Exam NDMC questions. The results, on unseen test questions, are robust across different test years and different variations of this kind of test. They demonstrate that modern NLP methods can result in mastery on this task. While not a full solution to general question-answering (the questions are multiple choice, and the domain is restricted to 8th Grade science), it represents a significant milestone for the field., Comment: AI Magazine 41 (4) Winter 2020. New analysis sections added

This paper describes AllenNLP, a platform for research on deep learning methods in natural language understanding. AllenNLP is designed to support researchers who want to build novel language understanding models quickly and easily. It is built on top of PyTorch, allowing for dynamic computation graphs, and provides (1) a flexible data API that handles intelligent batching and padding, (2) high-level abstractions for common operations in working with text, and (3) a modular and extensible experiment framework that makes doing good science easy. It also includes reference implementations of high quality approaches for both core semantic problems (e.g. semantic role labeling (Palmer et al., 2005)) and language understanding applications (e.g. machine comprehension (Rajpurkar et al., 2016)). AllenNLP is an ongoing open-source effort maintained by engineers and researchers at the Allen Institute for Artificial Intelligence., Comment: Describes the initial version of AllenNLP. Many features and models have been added since the first release. This is the paper to cite if you use AllenNLP in your research. Updated 5/31/2018 with version accepted to the NLP OSS workshop help at ACL 2018

We report on the fabrication and characterization of high-quality chemical vapor-deposited (CVD) bilayer graphene (BLG). In particular, we demonstrate that CVD-grown BLG can mechanically be detached from the copper foil by an hexagonal boron nitride (hBN) crystal after oxidation of the copper-to-BLG interface. Confocal Raman spectroscopy reveals an AB-stacking order of the BLG crystals and a high structural quality. From transport measurements on fully encapsulated hBN/BLG/hBN Hall bar devices we extract charge carrier mobilities up to 180,000 cm$^2$/(Vs) at 2 K and up to 40,000 cm$^2$/(Vs) at 300 K, outperforming state-of-the-art CVD bilayer graphene devices. Moreover, we show an on-off ration of more than 10,000 and a band gap opening with values of up to 15 meV for a displacement field of 0.2 V/nm in such CVD grown BLG., Comment: 5 pages, 4 figures

The Raman 2D line of graphene is widely used for device characterization and during device fabrication as it contains valuable information on e.g. the direction and magnitude of mechanical strain and doping. Here we present systematic asymmetries in the 2D line shape of exfoliated graphene and graphene grown by chemical vapor deposition. Both graphene crystals are fully encapsulated in van der Waals heterostructures, where hexagonal boron nitride and tungsten diselenide are used as substrate materials. In both material stacks, we find very low doping values and extremely homogeneous strain distributions in the graphene crystal, which is a hall mark of the outstanding electronic quality of these samples. By fitting double Lorentzian functions to the spectra to account for the contributions of inner and outer processes to the 2D peak, we find that the splitting of the sub-peaks, $6.6 \pm 0.5$ cm$^{-1}$(hBN-Gr-WSe2) and $8.9 \pm 1.0$ cm$^{-1}$ (hBN-Gr-hBN), is significantly lower than the values reported in previous studies on suspended graphene.

We present a magneto-Raman study on high-quality single-layer graphene grown by chemical vapor deposition (CVD) that is fully encapsulated in hexagonal boron nitride by a dry transfer technique. By analyzing the Raman D, G, and 2D peaks, we find that the structural quality of the samples is comparable to state-of-the-art exfoliated graphene flakes. From B field dependent Raman measurements, we extract the broadening and associated lifetime of the G peak due to anharmonic effects. Furthermore, we determine the decay width and lifetime of Landau level (LL) transitions from magneto-phonon resonances as a function of laser power. At low laser power, we find a minimal decay width of 140 1/cm highlighting the high electronic quality of the CVD-grown graphene. At higher laser power, we observe an increase of the LL decay width leading to a saturation with the corresponding lifetime saturating at a minimal value of 18 fs.

We report on ballistic transport over more than 28 \mu m in graphene grown by chemical vapor deposition (CVD) that is fully encapsulated in hexagonal boron nitride. The structures are fabricated by an advanced dry van-der-Waals transfer method and exhibit carrier mobilities of up to three million cm$^2$/(Vs). The ballistic nature of charge transport is probed by measuring the bend resistance in cross- and square-shaped devices. Temperature dependent measurements furthermore prove that ballistic transport is maintained exceeding 1 \mu m up to 200 K., Comment: 13 pages, 3 figures