Nous utilisons des témoins pour analyser le trafic et l’utilisation de notre site web, afin de personnaliser votre expérience. Vous pouvez désactiver ces technologies à tout moment, mais cela peut restreindre certaines fonctionnalités du site. Consultez notre Politique de protection de la vie privée pour en savoir plus.
Paramètre des cookies
Vous pouvez activer et désactiver les types de cookies que vous souhaitez accepter. Cependant certains choix que vous ferez pourraient affecter les services proposés sur nos sites (ex : suggestions, annonces personnalisées, etc.).
Cookies essentiels
Ces cookies sont nécessaires au fonctionnement du site et ne peuvent être désactivés. (Toujours actif)
Cookies analyse
Acceptez-vous l'utilisation de cookies pour mesurer l'audience de nos sites ?
Multimedia Player
Acceptez-vous l'utilisation de cookies pour afficher et vous permettre de regarder les contenus vidéo hébergés par nos partenaires (YouTube, etc.) ?
Publications
Evaluating Interventional Reasoning Capabilities of Large Language Models
Numerous decision-making tasks require estimating causal effects under interventions on different parts of a system. As practitioners consid… (voir plus)er using large language models (LLMs) to automate decisions, studying their causal reasoning capabilities becomes crucial. A recent line of work evaluates LLMs ability to retrieve commonsense causal facts, but these evaluations do not sufficiently assess how LLMs reason about interventions. Motivated by the role that interventions play in causal inference, in this paper, we conduct empirical analyses to evaluate whether LLMs can accurately update their knowledge of a data-generating process in response to an intervention. We create benchmarks that span diverse causal graphs (e.g., confounding, mediation) and variable types, and enable a study of intervention-based reasoning. These benchmarks allow us to isolate the ability of LLMs to accurately predict changes resulting from their ability to memorize facts or find other shortcuts. We evaluate six LLMs on the benchmarks, finding that GPT models show promising accuracy at predicting the intervention effects.
Specifications play a crucial role in neural network verification. They define the precise input regions we aim to verify, typically represe… (voir plus)nted as L-infinity norm balls. While recent research suggests using neural activation patterns (NAPs) as specifications for verifying unseen test set data, it focuses on computing the most refined NAPs, often limited to very small regions in the input space. In this paper, we study the following problem: Given a neural network, find a minimal (coarsest) NAP that is sufficient for formal verification of the network's robustness. Finding the minimal NAP specification not only expands verifiable bounds but also provides insights into which neurons contribute to the model's robustness. To address this problem, we propose several exact and approximate approaches. Our exact approaches leverage the verification tool to find minimal NAP specifications in either a deterministic or statistical manner. Whereas the approximate methods efficiently estimate minimal NAPs using adversarial examples and local gradients, without making calls to the verification tool. This allows us to inspect potential causal links between neurons and the robustness of state-of-the-art neural networks, a task for which existing verification frameworks fail to scale. Our experimental results suggest that minimal NAP specifications require much smaller fractions of neurons compared to the most refined NAP specifications, yet they can significantly expand the verifiable boundaries to several orders of magnitude larger.
Computing differences between tree-structured data is a critical but challenging problem in software analysis. In this paper, we propose a n… (voir plus)ovel tree diffing approach called SatDiff, which reformulates the structural diffing problem into a MaxSAT problem. By encoding the necessary transformations from the source tree to the target tree, SatDiff generates correct, minimal, and type safe low-level edit scripts with formal guarantees. We then synthesize concise high-level edit scripts by effectively merging low-level edits in the appropriate topological order. Our empirical results demonstrate that SatDiff outperforms existing heuristic-based approaches by a significant margin in terms of conciseness while maintaining a reasonable runtime.
Governance frameworks should address the prospect of AI systems that cannot be safely tested Technical experts and policy-makers have increa… (voir plus)singly emphasized the need to address extinction risk from artificial intelligence (AI) systems that might circumvent safeguards and thwart attempts to control them (1). Reinforcement learning (RL) agents that plan over a long time horizon far more effectively than humans present particular risks. Giving an advanced AI system the objective to maximize its reward and, at some point, withholding reward from it, strongly incentivizes the AI system to take humans out of the loop, if it has the opportunity. The incentive to deceive humans and thwart human control arises not only for RL agents but for long-term planning agents (LTPAs) more generally. Because empirical testing of sufficiently capable LTPAs is unlikely to uncover these dangerous tendencies, our core regulatory proposal is simple: Developers should not be permitted to build sufficiently capable LTPAs, and the resources required to build them should be subject to stringent controls.
Governance frameworks should address the prospect of AI systems that cannot be safely tested Technical experts and policy-makers have increa… (voir plus)singly emphasized the need to address extinction risk from artificial intelligence (AI) systems that might circumvent safeguards and thwart attempts to control them (1). Reinforcement learning (RL) agents that plan over a long time horizon far more effectively than humans present particular risks. Giving an advanced AI system the objective to maximize its reward and, at some point, withholding reward from it, strongly incentivizes the AI system to take humans out of the loop, if it has the opportunity. The incentive to deceive humans and thwart human control arises not only for RL agents but for long-term planning agents (LTPAs) more generally. Because empirical testing of sufficiently capable LTPAs is unlikely to uncover these dangerous tendencies, our core regulatory proposal is simple: Developers should not be permitted to build sufficiently capable LTPAs, and the resources required to build them should be subject to stringent controls.
Abstract Sentences containing multiple semantic operators with overlapping scope often create ambiguities in interpretation, known as scope … (voir plus)ambiguities. These ambiguities offer rich insights into the interaction between semantic structure and world knowledge in language processing. Despite this, there has been little research into how modern large language models treat them. In this paper, we investigate how different versions of certain autoregressive language models—GPT-2, GPT-3/3.5, Llama 2, and GPT-4—treat scope ambiguous sentences, and compare this with human judgments. We introduce novel datasets that contain a joint total of almost 1,000 unique scope-ambiguous sentences, containing interactions between a range of semantic operators, and annotated for human judgments. Using these datasets, we find evidence that several models (i) are sensitive to the meaning ambiguity in these sentences, in a way that patterns well with human judgments, and (ii) can successfully identify human-preferred readings at a high level of accuracy (over 90% in some cases).1
Abstract Sentences containing multiple semantic operators with overlapping scope often create ambiguities in interpretation, known as scope … (voir plus)ambiguities. These ambiguities offer rich insights into the interaction between semantic structure and world knowledge in language processing. Despite this, there has been little research into how modern large language models treat them. In this paper, we investigate how different versions of certain autoregressive language models—GPT-2, GPT-3/3.5, Llama 2, and GPT-4—treat scope ambiguous sentences, and compare this with human judgments. We introduce novel datasets that contain a joint total of almost 1,000 unique scope-ambiguous sentences, containing interactions between a range of semantic operators, and annotated for human judgments. Using these datasets, we find evidence that several models (i) are sensitive to the meaning ambiguity in these sentences, in a way that patterns well with human judgments, and (ii) can successfully identify human-preferred readings at a high level of accuracy (over 90% in some cases).1
The jar test is the current standard method for predicting the performance of a conventional drinking water treatment (DWT) process and opti… (voir plus)mizing the coagulant dose. This test is time-consuming and requires human intervention, meaning it is infeasible for making continuous process predictions. As a potential alternative, we developed a machine learning (ML) model from historical DWT plant data that can operate continuously using real-time sensor data without human intervention for predicting clarified water turbidity 15 min in advance. We evaluated three types of models: multilayer perceptron (MLP), the long short-term memory (LSTM) recurrent neural network (RNN), and the gated recurrent unit (GRU) RNN. We also employed two training methodologies: the commonly used holdout method and the theoretically correct blocked cross-validation (BCV) method. We found that the RNN with GRU was the best model type overall and achieved a mean absolute error on an independent production set of as low as 0.044 NTU. We further found that models trained using BCV typically achieve errors equal to or lower than their counterparts trained using holdout. These results suggest that RNNs trained using BCV are superior for the development of ML models for DWT processes compared to those reported in earlier literature.
