Publications

Epistemic Integrity in Large Language Models
Bijean Ghafouri
Shahrad Mohammadzadeh
James Zhou
Pratheeksha Nair
Jacob-Junqi Tian
Mayank Goel
Reihaneh Rabbany
Jean-François Godbout
Kellin Pelrine
Large language models are increasingly relied upon as sources of information, but their propensity for generating false or misleading statem… (see more)ents with high confidence poses risks for users and society. In this paper, we confront the critical problem of epistemic miscalibration—where a model's linguistic assertiveness fails to reflect its true internal certainty. We introduce a new human-labeled dataset and a novel method for measuring the linguistic assertiveness of Large Language Models which cuts error rates by over 50% relative to previous benchmarks. Validated across multiple datasets, our method reveals a stark misalignment between how confidently models linguistically present information and their actual accuracy. Further human evaluations confirm the severity of this miscalibration. This evidence underscores the urgent risk of the overstated certainty Large Language Models hold which may mislead users on a massive scale. Our framework provides a crucial step forward in diagnosing and correcting this miscalibration, offering a path to safer and more trustworthy AI across domains.
2024-10-12
NeurIPS.cc/2024/Workshop/SafeGenAi (poster)
openreview.net
openreview.net
Epistemic Integrity in Large Language Models
Bijean Ghafouri
Shahrad Mohammadzadeh
James Zhou
Pratheeksha Nair
Jacob-Junqi Tian
Mayank Goel
Reihaneh Rabbany
Jean-François Godbout
Kellin Pelrine
Large language models are increasingly relied upon as sources of information, but their propensity for generating false or misleading statem… (see more)ents with high confidence poses risks for users and society. In this paper, we confront the critical problem of epistemic miscalibration—where a model's linguistic assertiveness fails to reflect its true internal certainty. We introduce a new human-labeled dataset and a novel method for measuring the linguistic assertiveness of Large Language Models which cuts error rates by over 50% relative to previous benchmarks. Validated across multiple datasets, our method reveals a stark misalignment between how confidently models linguistically present information and their actual accuracy. Further human evaluations confirm the severity of this miscalibration. This evidence underscores the urgent risk of the overstated certainty Large Language Models hold which may mislead users on a massive scale. Our framework provides a crucial step forward in diagnosing and correcting this miscalibration, offering a path to safer and more trustworthy AI across domains.
2024-10-12
NeurIPS.cc/2024/Workshop/SafeGenAi (poster)
openreview.net
openreview.net
Hallucination Detox: Sensitive Neuron Dropout (SeND) for Large Language Model Training
Shahrad Mohammadzadeh
Juan David Guerra
Marco Bonizzato
Reihaneh Rabbany
Golnoosh Farnadi
As large language models (LLMs) become increasingly deployed across various industries, concerns regarding their reliability, particularly d… (see more)ue to hallucinations-outputs that are factually inaccurate or irrelevant to user input-have grown. Our research investigates the relationship between the training process and the emergence of hallucinations to address a key gap in existing research that focuses primarily on post hoc detection and mitigation strategies. Using models from the Pythia suite (70M-12B parameters) and several hallucination detection metrics, we analyze hallucination trends throughout training and explore LLM internal dynamics. We introduce SEnsitive Neuron Dropout (SeND), a novel training protocol designed to mitigate hallucinations by reducing variance during training. SeND achieves this by deterministically dropping neurons with significant variability on a dataset, referred to as Sensitive Neurons. In addition, we develop an unsupervised hallucination detection metric, Efficient EigenScore (EES), which approximates the traditional EigenScore in 2x speed. This efficient metric is integrated into our protocol, allowing SeND to be both computationally scalable and effective at reducing hallucinations. Our empirical evaluation demonstrates that our approach improves LLM reliability at test time by up to 40% compared to normal training while also providing an efficient method to improve factual accuracy when adapting LLMs to domains such as Wikipedia and Medical datasets.
2024-10-12
NeurIPS.cc/2024/Workshop/SafeGenAi (poster)
openreview.net
openreview.net
Identifying and Addressing Delusions for Target-Directed Decision-Making
Harry Zhao
Mingde Zhao
Tristan Sylvain
Doina Precup
Yoshua Bengio
We are interested in target-directed agents, which produce targets during decision-time planning, to guide their behaviors and achieve bette… (see more)r generalization during evaluation. Improper training of these agents can result in delusions: the agent may come to hold false beliefs about the targets, which cannot be properly rejected, leading to unwanted behaviors and damaging out-of-distribution generalization. We identify different types of delusions by using intuitive examples in carefully controlled environments, and investigate their causes. We demonstrate how delusions can be addressed for agents trained by hindsight relabeling, a mainstream approach in for training target-directed RL agents. We validate empirically the effectiveness of the proposed solutions in correcting delusional behaviors and improving out-of-distribution generalization.
2024-10-12
NeurIPS.cc/2024/Workshop/SafeGenAi (poster)
openreview.net
openreview.net
Quantifying Likeness: A Simple Machine Learning Approach to Identifying Copyright Infringement in (AI-Generated) Artwork
Michaela Drouillard
Ryan Spencer
Nikée Nantambu-Allen
Tegan Maharaj
Through study of legal precedent, we propose a pragmatic way to quantify copyright infringement, via stylistic similarity, in AI-generated a… (see more)rtwork. Copyright infringement by AI systems is a topic of rapidly-increasing importance as generative AI becomes more widespread and commercial. In contrast to typical work in this field, more in line with a realistic legal setting, our approach quantifies similarity of a set of potentially-infringing "defendant" artworks to a set of copyrighted "plaintiff" artworks. We develop our approach by making use of one of the most litigated artistic creations of this century -- Mickey Mouse. We curate a dataset using Mickey as the plaintiff, and perform hyperparameter search, scaling, and robustness analyses with various defendent artworks from real legal cases to find settings that generalize well. We operationalize similarity via a simple discrimintative task which can be accomplished in a low-resource setting by non-experts -- our aim is to provide a `plug and play' method that is feasible for artists and/or legal experts to use with their own plaintiff sets of artworks. We further demonstrate the viability of our approach by quantifying similarity in a second curated dataset of Maria Prymachenko's art vs. AI-generated images. We conclude by discussing uses of our work in both legal and other settings, including provision of artist compensation.
2024-10-12
NeurIPS.cc/2024/Workshop/SafeGenAi (poster)
openreview.net
openreview.net
Simulation System Towards Solving Societal-Scale Manipulation
Maximilian Puelma Touzel
Sneheel Sarangi
Austin Welch
Gayatri K
Dan Zhao
Zachary Yang
Hao Yu
Tom Gibbs
Ethan Kosak-Hine
Andreea Musulan
Camille Thibault
Busra Tugce Gurbuz
Reihaneh Rabbany
Jean-François Godbout
Kellin Pelrine
The rise of AI-driven manipulation poses significant risks to societal trust and democratic processes. Yet, studying these effects in real-w… (see more)orld settings at scale is ethically and logistically impractical, highlighting a need for simulation tools that can model these dynamics in controlled settings to enable experimentation with possible defenses. We present a simulation environment designed to address this. We elaborate upon the Concordia framework that simulates offline, `real life' activity by adding online interactions to the simulation through social media with the integration of a Mastodon server. Through a variety of means we then improve simulation efficiency and information flow, and add a set of measurement tools, particularly longitudinal surveys of the agents' political positions. We demonstrate the simulator with a tailored example of how partisan manipulation of agents can affect election results.
2024-10-12
NeurIPS.cc/2024/Workshop/SafeGenAi (poster)
openreview.net
openreview.net
The Structural Safety Generalization Problem
Tom Gibbs
Julius Broomfield
George Ingebretsen
Ethan Kosak-Hine
Tia Nasir
Jason Zhang
Reihaneh Iranmanesh
Sara Pieri
Reihaneh Rabbany
Kellin Pelrine
It is widely known that AI is vulnerable to adversarial examples, from pixel perturbations to jailbreaks. We propose that there is a key, ea… (see more)sier class of problems that is also still unsolved: failures of safety to generalize over structure, despite semantic equivalence. We demonstrate this vulnerability by showing how recent AI systems are differently vulnerable both to multi-turn and multi-image attacks, compared to their single-turn and single-image counterparts with equivalent meaning. We suggest this is the same class of vulnerability as that found in yet unconnected threads of the literature: vulnerabilities to low-resource languages and indefensibility of strongly superhuman Go AIs to cyclic attacks. When viewed together, these reveal a common picture: models that are not only vulnerable to attacks, but vulnerable to attacks with near identical meaning in their benign and harmful components both, and only different in structure. In contrast to attacks with identical benign input (e.g., pictures that look like cats) but unknown semanticity of the harmful component (e.g., diverse noise that is all unintelligible to humans), these represent a class of attacks where semantic understanding and defense against one version should guarantee defense against others—yet current AI safety measures do not. This vulnerability represents a necessary but not sufficient condition towards defending against attacks whose harmful component has arbitrary semanticity. Consequently, by building on the data and approaches we highlight, we frame an intermediate problem for AI safety to solve, that represents a critical checkpoint towards safe AI while being far more tractable than trying to solve it directly and universally.
2024-10-12
NeurIPS.cc/2024/Workshop/SafeGenAi (poster)
openreview.net
openreview.net
Unlearning in- vs. out-of-distribution data in LLMs under gradient-based methods
Teodora Băluță
Pascal Lamblin
Danny Tarlow
Fabian Pedregosa
Gintare Karolina Dziugaite
Machine unlearning aims to solve the problem of removing the influence of selected training examples from a learned model. Despite the incre… (see more)asing attention to this problem, it remains an open research question how to evaluate unlearning in large language models (LLMs), and what are the critical properties of the data to be unlearned that affect the quality and efficiency of unlearning. This work formalizes a metric to evaluate unlearning quality in generative models, and uses it to assess the trade-offs between unlearning quality and performance. We demonstrate that unlearning out-of-distribution examples requires more unlearning steps but overall presents a better trade-off overall. For in-distribution examples, however, we observe a rapid decay in performance as unlearning progresses. We further evaluate how example's memorization and difficulty affect unlearning under a classical gradient ascent-based approach.
2024-10-12
NeurIPS.cc/2024/Workshop/SafeGenAi (poster)
openreview.net
openreview.net
Cell ontology guided transcriptome foundation model
Xinyu Yuan
Zhihao Zhan
Zuobai Zhang
Manqi Zhou
Jianan Zhao
Boyu Han
Yue Li
Jian Tang
Transcriptome foundation models (TFMs) hold great promises of deciphering the transcriptomic language that dictate diverse cell functions by… (see more) self-supervised learning on large-scale single-cell gene expression data, and ultimately unraveling the complex mechanisms of human diseases. However, current TFMs treat cells as independent samples and ignore the taxonomic relationships between cell types, which are available in cell ontology graphs. We argue that effectively leveraging this ontology information during the TFM pre-training can improve learning biologically meaningful gene co-expression patterns while preserving TFM as a general purpose foundation model for downstream zero-shot and fine-tuning tasks. To this end, we present **s**ingle **c**ell, **Cell-o**ntology guided TFM (scCello). We introduce cell-type coherence loss and ontology alignment loss, which are minimized along with the masked gene expression prediction loss during the pre-training. The novel loss component guide scCello to learn the cell-type-specific representation and the structural relation between cell types from the cell ontology graph, respectively. We pre-trained scCello on 22 million cells from CellxGene database leveraging their cell-type labels mapped to the cell ontology graph from Open Biological and Biomedical Ontology Foundry. Our TFM demonstrates competitive generalization and transferability performance over the existing TFMs on biologically important tasks including identifying novel cell types of unseen cells, prediction of cell-type-specific marker genes, and cancer drug responses.
2024-10-11
NeurIPS.cc/2024/Workshop/FM4Science (poster)
openreview.net
openreview.net
MATES: A Deep Learning-Based Model for Locus-specific Quantification of Transposable Elements in Single Cell
Ruohan Wang
Yumin Zheng
Zijian Zhang
Kailu Song
Erxi Wu
Xiaopeng Zhu
Tao P. Wu
Jun Ding
Transposable elements (TEs) are crucial for genetic diversity and gene regulation. Current single-cell quantification methods often align mu… (see more)lti-mapping reads to either ‘best-mapped’ or ‘random-mapped’ locations and categorize them at sub-family levels, overlooking the biological necessity for accurate, locus-specific TE quantification. Moreover, these existing methods are primarily designed for and focused on transcriptomics data, which restricts their adaptability to single-cell data of other modalities. To address these challenges, here we introduce MATES, a novel deep-learning approach that accurately allocates multi-mapping reads to specific loci of TEs, utilizing context from adjacent read alignments flanking the TE locus. When applied to diverse single-cell omics datasets, MATES shows improved performance over existing methods, enhancing the accuracy of TE quantification and aiding in the identification of marker TEs for identified cell populations. This development enables exploring single-cell heterogeneity and gene regulation through the lens of TEs, offering a transformative tool for the single-cell genomics community.
2024-10-11
Nature Communications (published)
doi.org
Physical Simulation for Multi-agent Multi-machine Tending
Abdalwhab Abdalwhab
Giovanni Beltrame
David St-Onge
2024-10-11
ArXiv (preprint)
arxiv.org
arxiv.org
PoisonBench: Assessing Large Language Model Vulnerability to Data Poisoning
Tingchen Fu
Mrinank Sharma
Philip Torr
Shay B. Cohen
David Scott Krueger
Fazl Barez
Preference learning is a central component for aligning current LLMs, but this process can be vulnerable to data poisoning attacks. To addre… (see more)ss this concern, we introduce PoisonBench, a benchmark for evaluating large language models' susceptibility to data poisoning during preference learning. Data poisoning attacks can manipulate large language model responses to include hidden malicious content or biases, potentially causing the model to generate harmful or unintended outputs while appearing to function normally. We deploy two distinct attack types across eight realistic scenarios, assessing 21 widely-used models. Our findings reveal concerning trends: (1) Scaling up parameter size does not inherently enhance resilience against poisoning attacks; (2) There exists a log-linear relationship between the effects of the attack and the data poison ratio; (3) The effect of data poisoning can generalize to extrapolated triggers that are not included in the poisoned data. These results expose weaknesses in current preference learning techniques, highlighting the urgent need for more robust defenses against malicious models and data manipulation.
2024-10-11
ArXiv (preprint)
arxiv.org
arxiv.org