University of Wisconsin Madison
AI Insights
- Recommendation/ranking cues are rarely co-present (9%) and seldom traceable (7%) or contestable (2%), Advertising is more often co-present (36%) and traceable (25%) but never contestable in our sample (0%); governance/integrity cues remain less common overall (29) but disproportionately provide contestability (31%). (ML: 0.98)
- The study found that accountability attributes concentrate unevenly by decision type. (ML: 0.97)
- The study only examined a limited number of platforms and cues. (ML: 0.96)
- Co-present: The first explanation/disclosure/action surface that users encounter when interacting with a platform. (ML: 0.94)
- Feedback-loop opacity and placebo controls are issues that need to be addressed in order to improve user agency and accountability. (ML: 0.94)
- Contest: The ability to take action against a recommendation or decision made by the platform. (ML: 0.94)
- Longitudinal drift audits can be used to track changes in design over time and assess the evolution of metrics for each attribute. (ML: 0.93)
- The study highlights the importance of considering interaction depth as an auditable metric at the interface level of transparency. (ML: 0.91)
- Trace: The ability to access explanations or disclosures after navigating through multiple steps. (ML: 0.87)
Abstract
People who use social media are learning about how the companies that run these platforms make their decisions on who gets to see what through visual indicators in the interface (UI) of each social media site. These indicators are different for each platform and are not always located in an easy-to-find location on the site. Therefore, it is hard for someone to compare different social media platforms or determine whether transparency leads to greater accountability or only leads to increased understanding. A new classification system has been developed to help provide a standard way of categorizing the way, that an algorithm is presented through UI elements and whether the company has provided any type of explanation as to why they are featured. This new classification system includes the following three areas of development: design form, information content, and user agency. This new classification system can be applied to the six social media platforms currently available and serves as a reference database for identifying common archetypes of features in the each social media platform's UI. The new classification system will assist in determining whether or not the transparency of an algorithm functions the way that it was intended when it was developed and provide future design ideas that can help improve the inspectibility, actionability, and contestability of algorithms.
People who use social media are learning about how the companies that run these platforms make their decisions on who gets to see what through visual indicators in the interface (UI) of each social media site. These indicators are different for each platform and are not always located in an easy-to-find location on the site. Therefore, it is hard for someone to compare different social media platforms or determine whether transparency leads to greater accountability or only leads to increased understanding. A new classification system has been developed to help provide a standard way of categorizing the way, that an algorithm is presented through UI elements and whether the company has provided any type of explanation as to why they are featured. This new classification system includes the following three areas of development: design form, information content, and user agency. This new classification system can be applied to the six social media platforms currently available and serves as a reference database for identifying common archetypes of features in the each social media platform's UI. The new classification system will assist in determining whether or not the transparency of an algorithm functions the way that it was intended when it was developed and provide future design ideas that can help improve the inspectibility, actionability, and contestability of algorithms.
Why we are recommending this paper?
Due to your Interest in Data Transparency
This paper directly addresses AI transparency, a core interest, by examining how social media platforms communicate algorithmic decisions to users. Understanding these cues is crucial for assessing potential biases and promoting more accountable AI systems.
Due to your Interest in Data Transparency
This paper directly addresses AI transparency, a core interest, by examining how social media platforms communicate algorithmic decisions to users. Understanding these cues is crucial for assessing potential biases and promoting more accountable AI systems.
Max Planck Institute for Intelligent Systems
AI Insights
- The method is applied to a language model and a mathematical reasoning task, and it is demonstrated that the identified features are indeed interpretable. (ML: 0.98)
- The bound on the mean similarity between features can be used to identify interpretable features. (ML: 0.97)
- The approach is based on the idea that features with high self-coherence are more likely to be interpretable. (ML: 0.95)
- The method is based on the idea that features with high self-coherence are more likely to be interpretable. (ML: 0.95)
- The paper presents a method for identifying interpretable features in neural networks using orthogonality regularization. (ML: 0.94)
- The paper presents a novel approach for identifying interpretable features in neural networks using orthogonality regularization. (ML: 0.94)
- A bound on the mean similarity between features is derived, and it is shown that this bound can be used to identify interpretable features. (ML: 0.93)
- Self-coherence: The sum of the squares of the coefficients in a feature. (ML: 0.89)
- Orthogonality regularization: A method for promoting orthogonality between features by adding a penalty term to the loss function. (ML: 0.85)
- K-sparse: A vector with at most K non-zero elements. (ML: 0.74)
Abstract
With recent progress on fine-tuning language models around a fixed sparse autoencoder, we disentangle the decoder matrix into almost orthogonal features. This reduces interference and superposition between the features, while keeping performance on the target dataset essentially unchanged. Our orthogonality penalty leads to identifiable features, ensuring the uniqueness of the decomposition. Further, we find that the distance between embedded feature explanations increases with stricter orthogonality penalty, a desirable property for interpretability. Invoking the $\textit{Independent Causal Mechanisms}$ principle, we argue that orthogonality promotes modular representations amenable to causal intervention. We empirically show that these increasingly orthogonalized features allow for isolated interventions. Our code is available under $\texttt{https://github.com/mrtzmllr/sae-icm}$.
With recent progress on fine-tuning language models around a fixed sparse autoencoder, we disentangle the decoder matrix into almost orthogonal features. This reduces interference and superposition between the features, while keeping performance on the target dataset essentially unchanged. Our orthogonality penalty leads to identifiable features, ensuring the uniqueness of the decomposition. Further, we find that the distance between embedded feature explanations increases with stricter orthogonality penalty, a desirable property for interpretability. Invoking the $\textit{Independent Causal Mechanisms}$ principle, we argue that orthogonality promotes modular representations amenable to causal intervention. We empirically show that these increasingly orthogonalized features allow for isolated interventions. Our code is available under $\texttt{https://github.com/mrtzmllr/sae-icm}$.
Why we are recommending this paper?
Due to your Interest in Data Representation
This research tackles the fundamental problem of bias within AI models by proposing a method to disentangle features and reduce interference, aligning with concerns about data representation and fairness. The Max Planck Institute's involvement adds significant credibility to this approach.
Due to your Interest in Data Representation
This research tackles the fundamental problem of bias within AI models by proposing a method to disentangle features and reduce interference, aligning with concerns about data representation and fairness. The Max Planck Institute's involvement adds significant credibility to this approach.
Maastricht University
AI Insights
- The text also discusses the relationship between groundedness and maximization of complete and transitive preference relations. (ML: 0.97)
- Some of the key concepts explored include consistency, monotonicity, and weak axiom of revealed preference (WARP). (ML: 0.97)
- The results have implications for understanding rationalizability and groundedness in choice theory. (ML: 0.95)
- GAIC: Grounded Axiom of Revealed Preference. (ML: 0.92)
- A choice function c is said to satisfy GMAIC if it maximizes a complete and transitive preference relation over non-empty subsets of X. (ML: 0.91)
- Groundedness: A choice function c satisfies groundedness if for all x β X, there exists a set S β X \{x such that I(S) = β . (ML: 0.89)
- GMAIC: Grounded Maximizing Axiom of Choice. (ML: 0.89)
- The provided text provides a comprehensive proof of various theorems and propositions related to choice theory. (ML: 0.89)
- The proofs cover topics such as injectivity, surjectivity, and double union closure of interpretation functions. (ML: 0.88)
- A choice function c is said to satisfy GAIC if it satisfies groundedness and the corresponding interpretation I satisfies consistency, monotonicity, and WARP. (ML: 0.88)
- The proofs demonstrate the relationship between different axioms and properties of choice functions. (ML: 0.86)
- The provided text appears to be a proof of various theorems and propositions related to choice theory, specifically in the context of rationalizability and groundedness. (ML: 0.86)
Abstract
This paper proposes a model of choice via agentic artificial intelligence (AI). A key feature is that the AI may misinterpret a menu before recommending what to choose. A single acyclicity condition guarantees that there is a monotonic interpretation and a strict preference relation that together rationalize the AI's recommendations. Since this preference is in general not unique, there is no safeguard against it misaligning with that of a decision maker. What enables the verification of such AI alignment is interpretations satisfying double monotonicity. Indeed, double monotonicity ensures full identifiability and internal consistency. But, an additional idempotence property is required to guarantee that recommendations are fully rational and remain grounded within the original feasible set.
This paper proposes a model of choice via agentic artificial intelligence (AI). A key feature is that the AI may misinterpret a menu before recommending what to choose. A single acyclicity condition guarantees that there is a monotonic interpretation and a strict preference relation that together rationalize the AI's recommendations. Since this preference is in general not unique, there is no safeguard against it misaligning with that of a decision maker. What enables the verification of such AI alignment is interpretations satisfying double monotonicity. Indeed, double monotonicity ensures full identifiability and internal consistency. But, an additional idempotence property is required to guarantee that recommendations are fully rational and remain grounded within the original feasible set.
Why we are recommending this paper?
Due to your Interest in AI Bias
This paper explores the potential for misinterpretation within AI systems, directly relating to the userβs interest in AI bias and how these systems can lead to unfair outcomes. The model proposed offers a valuable framework for understanding and mitigating these risks.
Due to your Interest in AI Bias
This paper explores the potential for misinterpretation within AI systems, directly relating to the userβs interest in AI bias and how these systems can lead to unfair outcomes. The model proposed offers a valuable framework for understanding and mitigating these risks.
Tsinghua University
AI Insights
- The results show that incorporating income fairness and exposure bias into the ranking process can lead to more equitable outcomes and better performance in terms of relevance and diversity. (ML: 0.98)
- The paper builds upon previous work by incorporating income fairness and exposure bias into the traditional learning-to-rank paradigm. (ML: 0.98)
- However, this approach can lead to unequal outcomes when combined with other factors such as exposure bias. (ML: 0.98)
- The proposed framework may not be effective in scenarios where the data is highly imbalanced or has a large number of irrelevant features. (ML: 0.97)
- The paper proposes a new way to do this by combining income fairness with another concept called exposure bias, which is like making sure everyone gets a fair shot at getting noticed. (ML: 0.97)
- Previous research on fairness-aware ranking has focused primarily on individual fairness, which ensures that similar individuals are treated similarly. (ML: 0.97)
- The paper proposes a novel framework for fairness-aware ranking, which incorporates income fairness and exposure bias into the traditional learning-to-rank paradigm. (ML: 0.97)
- The paper proposes a novel framework for fairness-aware ranking, which incorporates income fairness and exposure bias into the traditional learning-to-rank paradigm. (ML: 0.97)
- That's where income fairness comes in - it ensures that everyone has an equal chance to get the job, regardless of their background. (ML: 0.97)
- Imagine you're trying to rank a list of job candidates based on their qualifications. (ML: 0.97)
- The proposed framework demonstrates improved fairness and performance compared to existing methods. (ML: 0.97)
- But what if some candidates have more resources or connections than others? (ML: 0.96)
- Income fairness: The idea that individuals with similar abilities or characteristics should have equal opportunities to receive resources or rewards. (ML: 0.94)
- Exposure bias: A type of bias that occurs when certain groups are more likely to be exposed to a particular resource or opportunity, leading to unequal outcomes. (ML: 0.88)
Abstract
Ranking is central to information distribution in web search and recommendation. Nowadays, in ranking optimization, the fairness to item providers is viewed as a crucial factor alongside ranking relevance for users. There are currently numerous concepts of fairness and one widely recognized fairness concept is Exposure Fairness. However, it relies primarily on exposure determined solely by position, overlooking other factors that significantly influence income, such as time. To address this limitation, we propose to study ranking fairness when the provider utility is influenced by other contextual factors and is neither equal to nor proportional to item exposure. We give a formal definition of Income Fairness and develop a corresponding measurement metric. Simulated experiments show that existing-exposure-fairness-based ranking algorithms fail to optimize the proposed income fairness. Therefore, we propose the Dynamic-Income-Derivative-aware Ranking Fairness algorithm, which, based on the marginal income gain at the present timestep, uses Taylor-expansion-based gradients to simultaneously optimize effectiveness and income fairness. In both offline and online settings with diverse time-income functions, DIDRF consistently outperforms state-of-the-art methods.
Ranking is central to information distribution in web search and recommendation. Nowadays, in ranking optimization, the fairness to item providers is viewed as a crucial factor alongside ranking relevance for users. There are currently numerous concepts of fairness and one widely recognized fairness concept is Exposure Fairness. However, it relies primarily on exposure determined solely by position, overlooking other factors that significantly influence income, such as time. To address this limitation, we propose to study ranking fairness when the provider utility is influenced by other contextual factors and is neither equal to nor proportional to item exposure. We give a formal definition of Income Fairness and develop a corresponding measurement metric. Simulated experiments show that existing-exposure-fairness-based ranking algorithms fail to optimize the proposed income fairness. Therefore, we propose the Dynamic-Income-Derivative-aware Ranking Fairness algorithm, which, based on the marginal income gain at the present timestep, uses Taylor-expansion-based gradients to simultaneously optimize effectiveness and income fairness. In both offline and online settings with diverse time-income functions, DIDRF consistently outperforms state-of-the-art methods.
Why we are recommending this paper?
Due to your Interest in Data Fairness
This work focuses on ranking fairness, a critical aspect of data fairness and algorithmic bias within information retrieval systems. The Tsinghua University contribution provides a relevant perspective on optimizing ranking algorithms for equitable outcomes.
Due to your Interest in Data Fairness
This work focuses on ranking fairness, a critical aspect of data fairness and algorithmic bias within information retrieval systems. The Tsinghua University contribution provides a relevant perspective on optimizing ranking algorithms for equitable outcomes.
University of Florida
AI Insights
- It enables practitioners to localize disparities, compare algorithms, and reason about accuracy-fairness trade-offs beyond conflicting scalar metrics. (ML: 0.99)
- Fairness Diagnosis: The process of identifying and addressing disparities in machine learning models. (ML: 0.99)
- Domain Shift: A change in the distribution of data that can affect the performance of a machine learning model. (ML: 0.99)
- Accuracy-Fairness Trade-off: The balance between achieving high accuracy and maintaining fairness in machine learning models. (ML: 0.99)
- It provides actionable insights for practitioners to improve their machine learning models. (ML: 0.98)
- The system has the potential to be extended to other areas, such as multiclass classification and language tasks. (ML: 0.98)
- RISE is an effective tool for post-hoc fairness diagnosis under domain shift. (ML: 0.98)
- RISE is an interactive system for post-hoc fairness diagnosis under domain shift. (ML: 0.98)
- The demo illustrates how RISE supports actionable model analysis, training, and selection. (ML: 0.95)
- Limited to binary classification problems. (ML: 0.94)
Abstract
Evaluating fairness under domain shift is challenging because scalar metrics often obscure exactly where and how disparities arise. We introduce \textit{RISE} (Residual Inspection through Sorted Evaluation), an interactive visualization tool that converts sorted residuals into interpretable patterns. By connecting residual curve structures to formal fairness notions, RISE enables localized disparity diagnosis, subgroup comparison across environments, and the detection of hidden fairness issues. Through post-hoc analysis, RISE exposes accuracy-fairness trade-offs that aggregate statistics miss, supporting more informed model selection.
Evaluating fairness under domain shift is challenging because scalar metrics often obscure exactly where and how disparities arise. We introduce \textit{RISE} (Residual Inspection through Sorted Evaluation), an interactive visualization tool that converts sorted residuals into interpretable patterns. By connecting residual curve structures to formal fairness notions, RISE enables localized disparity diagnosis, subgroup comparison across environments, and the detection of hidden fairness issues. Through post-hoc analysis, RISE exposes accuracy-fairness trade-offs that aggregate statistics miss, supporting more informed model selection.
Why we are recommending this paper?
Due to your Interest in AI Fairness
This paper offers a practical tool β RISE β for diagnosing and visualizing fairness issues in machine learning models, directly addressing the userβs interest in identifying and mitigating bias. The University of Floridaβs research is highly relevant to the broader challenge of evaluating fairness without harm.
Due to your Interest in AI Fairness
This paper offers a practical tool β RISE β for diagnosing and visualizing fairness issues in machine learning models, directly addressing the userβs interest in identifying and mitigating bias. The University of Floridaβs research is highly relevant to the broader challenge of evaluating fairness without harm.
University of Amsterdam
AI Insights
- A cultural shift towards proactive quality work, recognition of quality work as a routine part of ML practice, and collaboration between technical and legal teams are necessary for sustainable and trustworthy ML systems. (ML: 0.99)
- Technical limitations, organisational structures, workflow fragmentation, and gaps in collaboration between data and legal teams contribute to these challenges. (ML: 0.98)
- Socio-technical practice: A concept that highlights the interplay between technical infrastructures and regulatory expectations in shaping data quality practices. (ML: 0.98)
- Regulatory-aligned data quality: The practice of ensuring that machine learning systems meet regulatory requirements for data protection, transparency, and accountability. (ML: 0.98)
- Regulatory-aligned data quality is a socio-technical process shaped by regulation, governance, and engineering constraints. (ML: 0.98)
- Gaps in collaboration between data and legal teams can lead to delays, uncertainty, and reactive approaches to compliance. (ML: 0.98)
- Technical limitations and organisational structures can hinder the effective management of regulatory-aligned data quality. (ML: 0.98)
- Practitioners need tools that connect engineering work with regulatory requirements to manage regulatory-aligned data quality effectively. (ML: 0.97)
- Clearer governance structures, stable processes, and shared language are essential for supporting regulatory-aligned data quality management. (ML: 0.97)
- Practitioners struggle to translate high-level regulatory principles into concrete engineering practice. (ML: 0.95)
Abstract
Ensuring data quality in machine learning (ML) systems has become increasingly complex as regulatory requirements expand. In the European Union (EU), frameworks such as the General Data Protection Regulation (GDPR) and the Artificial Intelligence Act (AI Act) articulate data quality requirements that closely parallel technical concerns in ML practice, while also extending to legal obligations related to accountability, risk management, and human rights protection. This paper presents a qualitative interview study with EU-based data practitioners working on ML systems in regulated contexts. Through semi-structured interviews, we investigate how practitioners interpret regulatory-aligned data quality, the challenges they encounter, and the supports they identify as necessary. Our findings reveal persistent gaps between legal principles and engineering workflows, fragmentation across data pipelines, limitations of existing tools, unclear responsibility boundaries between technical and legal teams, and a tendency toward reactive, audit-driven quality practices. We also identify practitioners' needs for compliance-aware tooling, clearer governance structures, and cultural shifts toward proactive data governance.
Ensuring data quality in machine learning (ML) systems has become increasingly complex as regulatory requirements expand. In the European Union (EU), frameworks such as the General Data Protection Regulation (GDPR) and the Artificial Intelligence Act (AI Act) articulate data quality requirements that closely parallel technical concerns in ML practice, while also extending to legal obligations related to accountability, risk management, and human rights protection. This paper presents a qualitative interview study with EU-based data practitioners working on ML systems in regulated contexts. Through semi-structured interviews, we investigate how practitioners interpret regulatory-aligned data quality, the challenges they encounter, and the supports they identify as necessary. Our findings reveal persistent gaps between legal principles and engineering workflows, fragmentation across data pipelines, limitations of existing tools, unclear responsibility boundaries between technical and legal teams, and a tendency toward reactive, audit-driven quality practices. We also identify practitioners' needs for compliance-aware tooling, clearer governance structures, and cultural shifts toward proactive data governance.
Why we are recommending this paper?
Due to your Interest in Data Ethics
Due to your Interest in Data Ethics
China Mobile Research Institute
AI Insights
- The authors argue that developing principled mechanisms to assess and regulate inference validity under epistemic drift is an urgent direction for future research. (ML: 0.98)
- unobservable reliability drift: A phenomenon where an estimator converges to a biased value due to unmodeled changes in the data-generating process. (ML: 0.96)
- A number of references are provided, including papers on dataset shift, concept drift adaptation, and practical Bayesian model evaluation using leave-one-out cross-validation and WAIC. (ML: 0.94)
- The paper discusses a phenomenon called 'unobservable reliability drift' where an estimator converges to a biased value due to unmodeled changes in the data-generating process. (ML: 0.92)
- The bias term bt is not modeled by the inference procedure. (ML: 0.91)
- drift models: Types of observational bias considered, including linear drift, random-walk drift, and no-drift control. (ML: 0.91)
- inference procedure: The method used to update the posterior distribution given new observations, which assumes stationarity and does not account for drift. (ML: 0.90)
- The authors provide a proof sketch of Proposition 1, which states that the estimator converges almost surely to ΞΈβ+limnββ1/nPn t=1bt, whenever the limit exists. (ML: 0.83)
- Proposition 1: A formal statement and proof sketch of the phenomenon of stable convergence under unobservable drift. (ML: 0.81)
- A minimal inference problem with a single scalar parameter ΞΈβ=0 is considered, where observations are generated according to yt=ΞΈβ+Ο΅t+bt. (ML: 0.79)
- The authors provide a formal statement and proof sketch of this phenomenon, known as Proposition 1. (ML: 0.78)
- The authors argue that this phenomenon is a worst-case scenario and may be mitigated by partial observability or zero long-term mean drift in real-world applications. (ML: 0.77)
Abstract
Modern science increasingly relies on ever-growing observational datasets and automated inference pipelines, under the implicit belief that accumulating more data makes scientific conclusions more reliable. Here we show that this belief can fail in a fundamental and irreversible way. We identify a structural regime in which standard inference procedures converge smoothly, remain well calibrated, and pass conventional diagnostic checks, yet systematically converge to incorrect conclusions. This failure arises when the reliability of observations degrades in a manner that is intrinsically unobservable to the inference process itself. Using minimal synthetic experiments, we demonstrate that in this regime additional data do not correct error but instead amplify it, while residual-based and goodness-of-fit diagnostics remain misleadingly normal. These results reveal an intrinsic limit of data-driven science: stability, convergence, and confidence are not sufficient indicators of epistemic validity. We argue that inference cannot be treated as an unconditional consequence of data availability, but must instead be governed by explicit constraints on the integrity of the observational process.
Modern science increasingly relies on ever-growing observational datasets and automated inference pipelines, under the implicit belief that accumulating more data makes scientific conclusions more reliable. Here we show that this belief can fail in a fundamental and irreversible way. We identify a structural regime in which standard inference procedures converge smoothly, remain well calibrated, and pass conventional diagnostic checks, yet systematically converge to incorrect conclusions. This failure arises when the reliability of observations degrades in a manner that is intrinsically unobservable to the inference process itself. Using minimal synthetic experiments, we demonstrate that in this regime additional data do not correct error but instead amplify it, while residual-based and goodness-of-fit diagnostics remain misleadingly normal. These results reveal an intrinsic limit of data-driven science: stability, convergence, and confidence are not sufficient indicators of epistemic validity. We argue that inference cannot be treated as an unconditional consequence of data availability, but must instead be governed by explicit constraints on the integrity of the observational process.
Why we are recommending this paper?
Due to your Interest in Data Ethics
Due to your Interest in Data Ethics
University of Pennsylvania
AI Insights
- The authors rely heavily on empirical evidence, but do not provide a clear theoretical framework for their approach. (ML: 0.99)
- The authors provide empirical evidence that diminishing improvements in single-step accuracy can compound, resulting in exponential growth in the length of tasks a model can complete. (ML: 0.98)
- The authors argue that traditional evaluation methods, such as benchmarking and forecasting, are insufficient to capture the full range of LLM capabilities. (ML: 0.97)
- The paper presents a new method for evaluating the capabilities of large language models (LLMs). (ML: 0.97)
- The authors provide empirical evidence that their approach can capture the full range of LLM capabilities and predict future trends in AI development. (ML: 0.96)
- They propose a new approach based on the concept of 'execution' and 'planning', which distinguishes between a model's ability to execute a complex plan and its ability to generate plans in the first place. (ML: 0.95)
- LLMs: Large Language Models Benchmarking: Evaluating AI performance under realistic conditions Forecasting: Predicting future trends and developments in AI The paper presents a new method for evaluating LLM capabilities, which takes into account the distinction between execution and planning. (ML: 0.95)
Abstract
Rapidly increasing AI capabilities have substantial real-world consequences, ranging from AI safety concerns to labor market consequences. The Model Evaluation & Threat Research (METR) report argues that AI capabilities have exhibited exponential growth since 2019. In this note, we argue that the data does not support exponential growth, even in shorter-term horizons. Whereas the METR study claims that fitting sigmoid/logistic curves results in inflection points far in the future, we fit a sigmoid curve to their current data and find that the inflection point has already passed. In addition, we propose a more complex model that decomposes AI capabilities into base and reasoning capabilities, exhibiting individual rates of improvement. We prove that this model supports our hypothesis that AI capabilities will exhibit an inflection point in the near future. Our goal is not to establish a rigorous forecast of our own, but to highlight the fragility of existing forecasts of exponential growth.
Rapidly increasing AI capabilities have substantial real-world consequences, ranging from AI safety concerns to labor market consequences. The Model Evaluation & Threat Research (METR) report argues that AI capabilities have exhibited exponential growth since 2019. In this note, we argue that the data does not support exponential growth, even in shorter-term horizons. Whereas the METR study claims that fitting sigmoid/logistic curves results in inflection points far in the future, we fit a sigmoid curve to their current data and find that the inflection point has already passed. In addition, we propose a more complex model that decomposes AI capabilities into base and reasoning capabilities, exhibiting individual rates of improvement. We prove that this model supports our hypothesis that AI capabilities will exhibit an inflection point in the near future. Our goal is not to establish a rigorous forecast of our own, but to highlight the fragility of existing forecasts of exponential growth.
Why we are recommending this paper?
Due to your Interest in AI Bias
Due to your Interest in AI Bias
EPFL
AI Insights
- Researchers should carefully consider the limitations and potential biases when using synthetic data in statistical inference problems. (ML: 0.98)
- Machine learning model: A type of algorithm that learns patterns and relationships in data to make predictions or decisions. (ML: 0.98)
- Validation data from the target distribution is essential to assess improvements. (ML: 0.97)
- Effective augmentation requires external information sources beyond the base dataset. (ML: 0.96)
- Statistical estimation problem: A problem where the goal is to estimate unknown population parameters from limited samples. (ML: 0.96)
- Synthetic data augmentation or partial imputation has been proposed to address data scarcity in statistical inference problems. (ML: 0.96)
- Synthetic data augmentation for statistical estimation requires specialized methods to ensure validity guarantees. (ML: 0.95)
- Synthetic data augmentation: The process of generating new data points by using a generative model, which can be used to augment existing datasets. (ML: 0.94)
- The applicability of these methods is limited due to their specificity and the need for substantial future work. (ML: 0.92)
- Limited applicability of specialized methods (ML: 0.88)
Abstract
Recent advances in generative modelling have led many to see synthetic data as the go-to solution for a range of problems around data access, scarcity, and under-representation. In this paper, we study three prominent use cases: (1) Sharing synthetic data as a proxy for proprietary datasets to enable statistical analyses while protecting privacy, (2) Augmenting machine learning training sets with synthetic data to improve model performance, and (3) Augmenting datasets with synthetic data to reduce variance in statistical estimation. For each use case, we formalise the problem setting and study, through formal analysis and case studies, under which conditions synthetic data can achieve its intended objectives. We identify fundamental and practical limits that constrain when synthetic data can serve as an effective solution for a particular problem. Our analysis reveals that due to these limits many existing or envisioned use cases of synthetic data are a poor problem fit. Our formalisations and classification of synthetic data use cases enable decision makers to assess whether synthetic data is a suitable approach for their specific data availability problem.
Recent advances in generative modelling have led many to see synthetic data as the go-to solution for a range of problems around data access, scarcity, and under-representation. In this paper, we study three prominent use cases: (1) Sharing synthetic data as a proxy for proprietary datasets to enable statistical analyses while protecting privacy, (2) Augmenting machine learning training sets with synthetic data to improve model performance, and (3) Augmenting datasets with synthetic data to reduce variance in statistical estimation. For each use case, we formalise the problem setting and study, through formal analysis and case studies, under which conditions synthetic data can achieve its intended objectives. We identify fundamental and practical limits that constrain when synthetic data can serve as an effective solution for a particular problem. Our analysis reveals that due to these limits many existing or envisioned use cases of synthetic data are a poor problem fit. Our formalisations and classification of synthetic data use cases enable decision makers to assess whether synthetic data is a suitable approach for their specific data availability problem.
Why we are recommending this paper?
Due to your Interest in Data Transparency
Due to your Interest in Data Transparency
JigsawStack, Inc
AI Insights
- Small language models may not be as effective as large language models in certain tasks. (ML: 0.99)
- The use of LLMs with tools may be limited by the availability of high-quality training data. (ML: 0.99)
- Researchers are exploring the potential of multimodal safety classification, which may have significant implications for industries that rely on text-based data. (ML: 0.98)
- Multimodal safety classification is a technique used to identify potential risks or hazards in text-based data. (ML: 0.98)
- The use of small language models is gaining traction as a valuable plug-in for large language models. (ML: 0.95)
- LLMs (Large Language Models) are artificial intelligence models that can process and generate human-like language. (ML: 0.94)
- The use of LLMs with tools and small language models is becoming increasingly prevalent in various applications. (ML: 0.94)
- There is a growing interest in multimodal safety classification, with the introduction of Llama Guard 4 by Meta AI. (ML: 0.89)
- LLMs with tools are becoming increasingly popular, and researchers are exploring their potential in various applications. (ML: 0.88)
- Agentic AI refers to artificial intelligence systems that can perform tasks autonomously, often requiring human oversight or intervention. (ML: 0.83)
Abstract
We present Interfaze, a system that treats modern LLM applications as a problem of building and acting over context, not just picking the right monolithic model. Instead of a single transformer, we combine (i) a stack of heterogeneous DNNs paired with small language models as perception modules for OCR involving complex PDFs, charts and diagrams, and multilingual ASR with (ii) a context-construction layer that crawls, indexes, and parses external sources (web pages, code, PDFs) into compact structured state, and (iii) an action layer that can browse, retrieve, execute code in a sandbox, and drive a headless browser for dynamic web pages. A thin controller sits on top of this stack and exposes a single, OpenAI-style endpoint: it decides which small models and actions to run and always forwards the distilled context to a user-selected LLM that produces the final response. On this architecture, Interfaze-Beta achieves 83.6% on MMLU-Pro, 91.4% on MMLU, 81.3% on GPQA-Diamond, 57.8% on LiveCodeBench v5, and 90.0% on AIME-2025, along with strong multimodal scores on MMMU (val) (77.3%), AI2D (91.5%), ChartQA (90.9%), and Common Voice v16 (90.8%). We show that most queries are handled primarily by the small-model and tool stack, with the large LLM operating only on distilled context, yielding competitive accuracy while shifting the bulk of computation away from the most expensive and monolithic models.
We present Interfaze, a system that treats modern LLM applications as a problem of building and acting over context, not just picking the right monolithic model. Instead of a single transformer, we combine (i) a stack of heterogeneous DNNs paired with small language models as perception modules for OCR involving complex PDFs, charts and diagrams, and multilingual ASR with (ii) a context-construction layer that crawls, indexes, and parses external sources (web pages, code, PDFs) into compact structured state, and (iii) an action layer that can browse, retrieve, execute code in a sandbox, and drive a headless browser for dynamic web pages. A thin controller sits on top of this stack and exposes a single, OpenAI-style endpoint: it decides which small models and actions to run and always forwards the distilled context to a user-selected LLM that produces the final response. On this architecture, Interfaze-Beta achieves 83.6% on MMLU-Pro, 91.4% on MMLU, 81.3% on GPQA-Diamond, 57.8% on LiveCodeBench v5, and 90.0% on AIME-2025, along with strong multimodal scores on MMMU (val) (77.3%), AI2D (91.5%), ChartQA (90.9%), and Common Voice v16 (90.8%). We show that most queries are handled primarily by the small-model and tool stack, with the large LLM operating only on distilled context, yielding competitive accuracy while shifting the bulk of computation away from the most expensive and monolithic models.
Why we are recommending this paper?
Due to your Interest in AI Transparency
Due to your Interest in AI Transparency
Johannes Gutenberg University
AI Insights
- The build-up of the recursive relation is a critical aspect of the semi-naive evaluation, where found facts are identified as new and unique and subsequently appended to the recursive relation. (ML: 0.97)
- The strategy makes a significant performance difference in individual cases, and different strategies win in different situations, so multiple strategies should be considered. (ML: 0.95)
- Different strategies to identify and append new and unique tuples vary in the order in which operations are executed, affecting performance. (ML: 0.94)
- Semi-naive evaluation: A method for evaluating recursive queries that incrementally builds up the recursive relation by identifying new and unique tuples and appending them to the base relation. (ML: 0.93)
- Multiple strategies should be considered depending on the workload characteristics and query pattern. (ML: 0.90)
- Physical representation: A data structure used to store and access a relation, which can be optimized for specific query patterns or workloads. (ML: 0.90)
- The choice of physical representation and deduplication strategy has a significant impact on the performance of recursive queries. (ML: 0.85)
- The decision to create exclusive physical representations must consider the relationship between body evaluation and bulk-loading effort, as well as potential restrictions due to a higher memory footprint. (ML: 0.84)
- Using four indexes with UPI consumes only 375MB across all data structures, except RX (612MB), highlighting UPIβs use for sharing indexes to save memory. (ML: 0.69)
- For unordered probes, creating exclusive physical representations does not pay off for the ordered structures SA, BP, and RX in any case. (ML: 0.59)
Abstract
Datalog is an increasingly popular recursive query language that is declarative by design, meaning its programs must be translated by an engine into the actual physical execution plan. When generating this plan, a central decision is how to physically represent all involved relations, an aspect in which existing Datalog engines are surprisingly restrictive and often resort to one-size-fits-all solutions. The reason for this is that the typical execution plan of a Datalog program not only performs a single type of operation against the physical representations, but a mixture of operations, such as insertions, lookups, and containment-checks. Further, the relevance of each operation type highly depends on the workload characteristics, which range from familiar properties such as the size, multiplicity, and arity of the individual relations to very specific Datalog properties, such as the "interweaving" of rules when relations occur multiple times, and in particular the recursiveness of the query which might generate new tuples on the fly during evaluation. This indicates that a variety of physical representations, each with its own strengths and weaknesses, is required to meet the specific needs of different workload situations. To evaluate this, we conduct an in-depth experimental study of the interplay between potentially suitable physical representations and seven dimensions of workload characteristics that vary across actual Datalog programs, revealing which properties actually matter. Based on these insights, we design an automatic selection mechanism that utilizes a set of decision trees to identify suitable physical representations for a given workload.
Datalog is an increasingly popular recursive query language that is declarative by design, meaning its programs must be translated by an engine into the actual physical execution plan. When generating this plan, a central decision is how to physically represent all involved relations, an aspect in which existing Datalog engines are surprisingly restrictive and often resort to one-size-fits-all solutions. The reason for this is that the typical execution plan of a Datalog program not only performs a single type of operation against the physical representations, but a mixture of operations, such as insertions, lookups, and containment-checks. Further, the relevance of each operation type highly depends on the workload characteristics, which range from familiar properties such as the size, multiplicity, and arity of the individual relations to very specific Datalog properties, such as the "interweaving" of rules when relations occur multiple times, and in particular the recursiveness of the query which might generate new tuples on the fly during evaluation. This indicates that a variety of physical representations, each with its own strengths and weaknesses, is required to meet the specific needs of different workload situations. To evaluate this, we conduct an in-depth experimental study of the interplay between potentially suitable physical representations and seven dimensions of workload characteristics that vary across actual Datalog programs, revealing which properties actually matter. Based on these insights, we design an automatic selection mechanism that utilizes a set of decision trees to identify suitable physical representations for a given workload.
Why we are recommending this paper?
Due to your Interest in Data Representation
Due to your Interest in Data Representation
The Ohio State University
AI Insights
- Fairness interventions can be effective, but their performance depends on the specific problem and dataset. (ML: 0.99)
- AI fairness issues remain challenging in computer vision domains, even as new methods are proposed and model capacity continues to increase. (ML: 0.98)
- The study highlights the importance of hyperparameter tuning and model selection in achieving fairness without harm. (ML: 0.98)
- Utility need not be sacrificed: data augmentation can deliver simultaneous gains in accuracy and subgroup parity. (ML: 0.97)
- The study introduces NH-Fair, a rigorously curated benchmark for evaluating fairness interventions in image models. (ML: 0.97)
- NH-Fair: A rigorously curated benchmark for evaluating fairness interventions in image models. (ML: 0.97)
- ERM: Empirical Risk Minimization, a method used to train machine learning models. (ML: 0.95)
- Data augmentation: A technique used to artificially increase the size of a training dataset by applying transformations to existing images. (ML: 0.95)
- Hyperparameter search: The process of finding the optimal values for model hyperparameters. (ML: 0.93)
- A carefully tuned ERM with hyperparameter search often rivals specialized debiasing methods. (ML: 0.92)
Abstract
Machine learning models trained on real-world data often inherit and amplify biases against certain social groups, raising urgent concerns about their deployment at scale. While numerous bias mitigation methods have been proposed, comparing the effectiveness of bias mitigation methods remains difficult due to heterogeneous datasets, inconsistent fairness metrics, isolated evaluation of vision versus multi-modal models, and insufficient hyperparameter tuning that undermines fair comparisons. We introduce NH-Fair, a unified benchmark for fairness without harm that spans both vision models and large vision-language models (LVLMs) under standardized data, metrics, and training protocols, covering supervised and zero-shot regimes. Our key contributions are: (1) a systematic ERM tuning study that identifies training choices with large influence on both utility and disparities, yielding empirically grounded guidelines to help practitioners reduce expensive hyperparameter tuning space in achieving strong fairness and accuracy; (2) evidence that many debiasing methods do not reliably outperform a well-tuned ERM baseline, whereas a composite data-augmentation method consistently delivers parity gains without sacrificing utility, emerging as a promising practical strategy. (3) an analysis showing that while LVLMs achieve higher average accuracy, they still exhibit subgroup disparities, and gains from scaling are typically smaller than those from architectural or training-protocol choices. NH-Fair provides a reproducible, tuning-aware pipeline for rigorous, harm-aware fairness evaluation.
Machine learning models trained on real-world data often inherit and amplify biases against certain social groups, raising urgent concerns about their deployment at scale. While numerous bias mitigation methods have been proposed, comparing the effectiveness of bias mitigation methods remains difficult due to heterogeneous datasets, inconsistent fairness metrics, isolated evaluation of vision versus multi-modal models, and insufficient hyperparameter tuning that undermines fair comparisons. We introduce NH-Fair, a unified benchmark for fairness without harm that spans both vision models and large vision-language models (LVLMs) under standardized data, metrics, and training protocols, covering supervised and zero-shot regimes. Our key contributions are: (1) a systematic ERM tuning study that identifies training choices with large influence on both utility and disparities, yielding empirically grounded guidelines to help practitioners reduce expensive hyperparameter tuning space in achieving strong fairness and accuracy; (2) evidence that many debiasing methods do not reliably outperform a well-tuned ERM baseline, whereas a composite data-augmentation method consistently delivers parity gains without sacrificing utility, emerging as a promising practical strategy. (3) an analysis showing that while LVLMs achieve higher average accuracy, they still exhibit subgroup disparities, and gains from scaling are typically smaller than those from architectural or training-protocol choices. NH-Fair provides a reproducible, tuning-aware pipeline for rigorous, harm-aware fairness evaluation.
Why we are recommending this paper?
Due to your Interest in Data Fairness
Due to your Interest in Data Fairness
ELLIS Alicante
AI Insights
- It highlights the importance of teaching students how to use AI effectively and critically, rather than simply relying on it as a tool for learning. (ML: 0.99)
- The article concludes by emphasizing the need for educators to strike a balance between using AI tools and promoting critical thinking skills in students. (ML: 0.99)
- However, there are concerns about the over-reliance on AI, which can lead to decreased critical thinking skills and academic performance. (ML: 0.99)
- However, there are also concerns about the impact of AI on student mental health, with some studies suggesting that over-reliance on AI can lead to increased stress and anxiety. (ML: 0.98)
- The article discusses the impact of artificial intelligence (AI) in education, highlighting both its benefits and drawbacks. (ML: 0.97)
- A study found that students who used AI tools excessively showed a significant decrease in their ability to solve problems independently. (ML: 0.96)
- The use of AI in education has increased significantly, with 70% of students using AI tools for learning. (ML: 0.96)
- The article also discusses the potential benefits of AI in education, including personalized learning, improved accessibility, and enhanced student engagement. (ML: 0.96)
- Chatbots: Computer programs that use AI to simulate conversation with humans, often used in customer service or education. (ML: 0.93)
- Artificial intelligence (AI): A type of computer system that can perform tasks that would typically require human intelligence, such as learning, problem-solving, and decision-making. (ML: 0.92)
Abstract
Artificial intelligence (AI) is rapidly being integrated into educational contexts, promising personalized support and increased efficiency. However, growing evidence suggests that the uncritical adoption of AI may produce unintended harms that extend beyond individual learning outcomes to affect broader societal goals. This paper examines the societal implications of AI in education through an integrative framework with four interrelated dimensions: cognition, agency, emotional well-being, and ethics. Drawing on research from education, cognitive science, psychology, and ethics, we synthesize existing evidence to show how AI-driven cognitive offloading, diminished learner agency, emotional disengagement, and surveillance-oriented practices can mutually reinforce one another. We argue that these dynamics risk undermining critical thinking, intellectual autonomy, emotional resilience, and trust, capacities that are foundational both for effective learning and also for democratic participation and informed civic engagement. Moreover, AI's impact is contingent on design and governance: pedagogically aligned, ethically grounded, and human-centered AI systems can scaffold effortful reasoning, support learner agency, and preserve meaningful social interaction. By integrating fragmented strands of prior research into a unified framework, this paper advances the discourse on responsible AI in education and offers actionable implications for educators, designers, and institutions. Ultimately, the paper contends that the central challenge is not whether AI should be used in education, but how it can be designed and governed to support learning while safeguarding the social and civic purposes of education.
Artificial intelligence (AI) is rapidly being integrated into educational contexts, promising personalized support and increased efficiency. However, growing evidence suggests that the uncritical adoption of AI may produce unintended harms that extend beyond individual learning outcomes to affect broader societal goals. This paper examines the societal implications of AI in education through an integrative framework with four interrelated dimensions: cognition, agency, emotional well-being, and ethics. Drawing on research from education, cognitive science, psychology, and ethics, we synthesize existing evidence to show how AI-driven cognitive offloading, diminished learner agency, emotional disengagement, and surveillance-oriented practices can mutually reinforce one another. We argue that these dynamics risk undermining critical thinking, intellectual autonomy, emotional resilience, and trust, capacities that are foundational both for effective learning and also for democratic participation and informed civic engagement. Moreover, AI's impact is contingent on design and governance: pedagogically aligned, ethically grounded, and human-centered AI systems can scaffold effortful reasoning, support learner agency, and preserve meaningful social interaction. By integrating fragmented strands of prior research into a unified framework, this paper advances the discourse on responsible AI in education and offers actionable implications for educators, designers, and institutions. Ultimately, the paper contends that the central challenge is not whether AI should be used in education, but how it can be designed and governed to support learning while safeguarding the social and civic purposes of education.
Why we are recommending this paper?
Due to your Interest in AI Ethics
Due to your Interest in AI Ethics
University of Southern California
AI Insights
- The study highlights the limitations of using large language models (LLMs) for social surveys and emphasizes the need for caution when interpreting results. (ML: 0.99)
- LLM: Large Language Model Misinformation: False or inaccurate information spread through various media channels Bias: Systematic error or distortion in a model's predictions or behavior The study demonstrates the importance of critically evaluating LLM-based social surveys and highlights the need for more research on mitigating biases in these models. (ML: 0.99)
- The authors suggest that targeted training, prompting, or architectural interventions may be necessary to reduce the systematic patterns documented in this study. (ML: 0.99)
- The study relies on online panels, which may limit representativeness and introduce sampling biases (ML: 0.99)
- The authors used a combination of distributional metrics, predictive modeling, interaction analysis, reasoning inspection, and training-data tracing to diagnose bias and distortion in LLM-based social surveys. (ML: 0.99)
- The authors emphasize that caution should be exercised when interpreting results from LLM-based social surveys, as surface-level alignment may not necessarily reflect underlying relational structure or feature effects. (ML: 0.99)
- The study found that surface-level alignment between real and simulated responses can mask important discrepancies in relational structure and feature effects. (ML: 0.96)
Abstract
Large language models (LLMs) are increasingly used as proxies for human judgment in computational social science, yet their ability to reproduce patterns of susceptibility to misinformation remains unclear. We test whether LLM-simulated survey respondents, prompted with participant profiles drawn from social survey data measuring network, demographic, attitudinal and behavioral features, can reproduce human patterns of misinformation belief and sharing. Using three online surveys as baselines, we evaluate whether LLM outputs match observed response distributions and recover feature-outcome associations present in the original survey data. LLM-generated responses capture broad distributional tendencies and show modest correlation with human responses, but consistently overstate the association between belief and sharing. Linear models fit to simulated responses exhibit substantially higher explained variance and place disproportionate weight on attitudinal and behavioral features, while largely ignoring personal network characteristics, relative to models fit to human responses. Analyses of model-generated reasoning and LLM training data suggest that these distortions reflect systematic biases in how misinformation-related concepts are represented. Our findings suggest that LLM-based survey simulations are better suited for diagnosing systematic divergences from human judgment than for substituting it.
Large language models (LLMs) are increasingly used as proxies for human judgment in computational social science, yet their ability to reproduce patterns of susceptibility to misinformation remains unclear. We test whether LLM-simulated survey respondents, prompted with participant profiles drawn from social survey data measuring network, demographic, attitudinal and behavioral features, can reproduce human patterns of misinformation belief and sharing. Using three online surveys as baselines, we evaluate whether LLM outputs match observed response distributions and recover feature-outcome associations present in the original survey data. LLM-generated responses capture broad distributional tendencies and show modest correlation with human responses, but consistently overstate the association between belief and sharing. Linear models fit to simulated responses exhibit substantially higher explained variance and place disproportionate weight on attitudinal and behavioral features, while largely ignoring personal network characteristics, relative to models fit to human responses. Analyses of model-generated reasoning and LLM training data suggest that these distortions reflect systematic biases in how misinformation-related concepts are represented. Our findings suggest that LLM-based survey simulations are better suited for diagnosing systematic divergences from human judgment than for substituting it.
Why we are recommending this paper?
Due to your Interest in Data Bias
Due to your Interest in Data Bias
π¬ Help Shape Our Pricing
We're exploring pricing options to make this project sustainable. Take 3 minutes to share what you'd be willing to pay (if anything). Your input guides our future investment.
Help us improve your experience!
This project is on its early stages your feedback can be pivotal on the future of the project.
Let us know what you think about this week's papers and suggestions!
Give Feedback