9607 research outputs found
Sort by
PolyPythias: Stability and Outliers across Fifty Language Model Pre-Training Runs
No full textThe stability of language model pre-training and its effects on downstream performance are still understudied. Prior work shows that the training process can yield significantly different results in response to slight variations in initial conditions, e.g., the random seed. Crucially, the research community still lacks sufficient resources and tools to systematically investigate pre-training stability, particularly for decoder-only language models. We introduce the PolyPythias, a set of 45 new training runs for the Pythia model suite: 9 new seeds across 5 model sizes, from 14M to 410M parameters, resulting in about 7k new checkpoints that we release. Using these new 45 training runs, in addition to the 5 already available, we study the effects of different initial conditions determined by the seed—i.e., parameters' initialisation and data order—on (i) downstream performance, (ii) learned linguistic representations, and (iii) emergence of training phases. In addition to common scaling behaviours, our analyses generally reveal highly consistent training dynamics across both model sizes and initial conditions. Further, the new seeds for each model allow us to identify outlier training runs and delineate their characteristics. Our findings show the potential of using these methods to predict training stability
NPA Hierarchy for Quantum Isomorphism and Homomorphism Indistinguishability
No full textMančinska and Roberson [FOCS'20] showed that two graphs are quantum isomorphic if and only if they are homomorphism indistinguishable over the class of planar graphs. Atserias et al. [JCTB'19] proved that quantum isomorphism is undecidable in general. The NPA hierarchy gives a sequence of semidefinite programming relaxations of quantum isomorphism. Recently, Roberson and Seppelt [ICALP'23] obtained a homomorphism indistinguishability characterization of the feasibility of each level of the Lasserre hierarchy of semidefinite programming relaxations of graph isomorphism. We prove a quantum analogue of this result by showing that each level of the NPA hierarchy of SDP relaxations for quantum isomorphism of graphs is equivalent to homomorphism indistinguishability over an appropriate class of planar graphs. By combining the convergence of the NPA hierarchy with the fact that the union of these graph classes is the set of all planar graphs, we are able to give a new proof of the result of Mančinska and Roberson [FOCS'20] that avoids the use of the theory of quantum groups. This homomorphism indistinguishability characterization also allows us to give a randomized polynomial-time algorithm deciding exact feasibility of each fixed level of the NPA hierarchy of SDP relaxations for quantum isomorphism
The Relational Becoming of a Participatory Design Commoner
No full textThis research expands the understanding of the collective designer behind Scandinavian Participatory Design with the Latin American notion of relational ontology, which posits a worldview where everything and everyone are mutually constituted through evolving relationships. Drawing from three thematic workshops with design practitioners and scholars that explored intersections between commoning and designing, this research unveils the relational becoming of a participatory design commoner. This includes the production of a shared subjectivity—the collective designer—as integral to commoning design and designing commons. Specifically, we explore the role of participatory designers in commoning through their subjectivity in their infrastructuring actions and affective engagements with objective commons
Declarative Dynamic Object Reclassification
No full textIn object-oriented languages, dynamic object reclassification is a technique to change the class binding of an object at runtime. Current approaches express when and how to reclassify inside the program’s business code, while maintaining internal consistency. These approaches are less suited for programs that need to be consistent with an external context, such as autonomous systems interacting with a knowledge base. This paper proposes declarative dynamic object reclassification, a novel technique that provides a separation of concerns between a program’s business code and its adaptation logic for reclassification, expressed via a knowledge base. We present Featherweight Semantically Reflected Java, a minimal calculus for declarative dynamic object reclassification that enables the programmer to define consistency both internally (using a type system) and externally (using declarative classification queries). We use this calculus to study how internal and external consistency interact for declarative dynamic object reclassification. We further implement the technique by extending SMOL, a language for reflective programming via external knowledge bases
User Misconceptions of LLM-Based Conversational Programming Assistants.
No full textProgramming assistants powered by large language models (LLMs) have become widely available, with conversational assistants like ChatGPT proving particularly accessible to less experienced programmers. However, the varied capabilities of these tools across model versions and the mixed availability of extensions that enable web search, code execution, or retrieval-augmented generation create opportunities for user misconceptions about what systems can and cannot do. Such misconceptions may lead to over-reliance, unproductive practices, or insufficient quality control in LLM-assisted programming. Here, we aim to characterize misconceptions that users of conversational LLM-based assistants may have in programming contexts. Using a two-phase approach, we first brainstorm and catalog user misconceptions that may occur, and then conduct a qualitative analysis to examine whether these conceptual issues surface in naturalistic Python-programming conversations with an LLM-based chatbot drawn from an openly available dataset. Indeed, we see evidence that some users have misplaced expectations about the availability of LLM-based chatbot features like web access, code execution, or non-text output generation. We also see potential evidence for deeper conceptual issues around the scope of information required to debug, validate, and optimize programs. Our findings reinforce the need for designing LLM-based tools that more clearly communicate their programming capabilities to users.<br/
The (un)fair algorithm:: The emergence of a dual ethics for artificial intelligence in social work
No full textPositive Thinking: Countercation Effects in Colloidal Syntheses of Gold Nanoparticles
No full textGold nanoparticles (Au NPs) are intensively studied and widely applicable to catalysis, sensing, medical applications, and many more. In particular, citrate- and borohydride- mediated colloidal syntheses of Au NPs are extremely popular. While it can be reasonably expected that countercations have a role to play, there is surprisingly almost no study on the effect of countercations in citrate- and borohydride-mediated colloidal syntheses of Au NPs. It is here shown that the countercation (Li+, Na+, K+) from citrate, borohydride, but also from hydroxide species, plays an overlooked role in the stabilization of gold colloidal dispersions. The stability, size, and degree of shape control over the NP decrease in the order Li+ > Na+ > K+, due to a stronger interaction between the smaller cations and metal surfaces. The findings are directly relevant for further fundamental studies, an improved control of the syntheses and scale-up
How Creative Practitioners Use Tools to Capture Ideas: A Cross-Domain Study
No full textCreative practitioners rely on tools to capture and manage ideas as a foundational aspect of their work. However, we have little knowledge about how idea management practices vary in different creative domains. Combining insights from qualitative surveys (N 200) and follow-up in-depth interviews (n 60) with creative professionals from four domains (interaction design, research, music, and graphics) of creative work, we report on (1) how ideas are externalized in practitioners’ archives, (2) what they consider important when choosing tools to capture ideas, and (3) how these tool collections resemble and differ from each other. Our cross-domain study demonstrates that participants’ tool use reflects idea capture characteristic needs as well as domain-specific views about the creative process. We conclude with a discussion about capturing as an externalizing activity, practitioners’ use of the term ideas, and four suggestions for directions in the design of creativity support tools
How CO2STLY Is CHI? The Carbon Footprint of Generative AI in HCI Research and What We Should Do About It
No full textThe energy cost of developing and deploying Generative AI (GenAI) models has exploded with their mass adoption, as has the ensuing carbon emissions. The climate impact of this is currently unknown. In Human-Computer Interaction, GenAI models are rarely trained but often used. Based on detailed review of 282 papers, we estimate this footprint from energy consumption of the total use of GenAI for CHI 2024 research as between 10,769.63 and 10,925.12 kg CO2e — equal to driving a car for more than 100,000 km. We show that in CHI research, GenAI is most often used for Prototyping, Evaluation & User studies, and that Data Collection and Fine-tuning models incurs the highest CO2st.1 We find that CHI submissions are unlikely to report GenAI use transparently, which makes precise calculations difficult. By measuring the usage of a subset of the papers on local hardware, we obtain estimations of the energy consumption and carbon footprint. Based on this evidence, we discuss and demonstrate ways to mitigate the issues of GenAI carbon footprint and lack of transparency
The Case for External Graph Sketching
No full textAlgorithms in the data stream model use O(polylog(N)) space to compute some property of an input of size N, and many of these algorithms are implemented and used in practice. However, sketching algorithms in the graph semi-streaming model use O(V polylog(V)) space for a V-vertex graph, and the fact that implementations of these algorithms are not used in the academic literature or in industrial applications may be because this space requirement is too large for RAM on today's hardware.In this paper we introduce the external semi-streaming model, which addresses the aspects of the semi-streaming model that limit its practical impact. In this model, the input is in the form of a stream and O(V polylog(V)) space is available, but most of that space is accessible only via block I/O operations as in the external memory model. The goal in the external semi-streaming model is to simultaneously achieve small space and low I/O cost.We present a general transformation from any vertex-based sketch algorithm to one which has a low sketching cost in the new model. We prove that this automatic transformation is tight or nearly (up to a O(\log(V)) factor) tight via an I/O lower bound for the task of sketching the input stream.Using this transformation and other techniques, we present external semi-streaming algorithms for connectivity, bipartiteness testing, (1+\epsilon)-approximating MST weight, testing k-edge connectivity, (1+\epsilon)-approximating the minimum cut of a graph, computing \epsilon-cut sparsifiers, and approximating the density of the densest subgraph. These algorithms all use O(V poly(\log(V), \epsilon^{-1},k) space. For many of these problems, our external semi-streaming algorithms outperform the state of the art algorithms in both the sketching and external-memory models