On the Role of Prompt Multiplicity in LLM Hallucination Evaluation
Prakhar Ganesh
Reza Shokri
Golnoosh Farnadi
Large language models (LLMs) are known to "hallucinate" by generating false or misleading outputs. Existing hallucination benchmarks often o… (voir plus)verlook prompt sensitivity, due to stable accuracy scores despite prompt variations. However, such stability can be misleading. In this work, we introduce prompt multiplicity--the multiplicity of individual hallucinations depending on the input prompt--and study its role in LLM hallucination benchmarks. We find severe multiplicity, with even more than 50% of responses changing between correct and incorrect answers simply based on the prompt for certain benchmarks, like Med-HALT. Prompt multiplicity also gives us the lens to distinguish between randomness in generation and consistent factual inaccuracies, providing a more nuanced understanding of LLM hallucinations and their real-world harms. By situating our discussion within existing hallucination taxonomies--supporting their quantification--and exploring its relationship with uncertainty in generation, we highlight how prompt multiplicity fills a critical gap in the literature on LLM hallucinations.
2025-03-05
ICLR.cc/2025/Workshop/BuildingTrust (accepté)
openreview.net
openreview.net
Training Plug n' Play Knowledge Modules with Deep Context Distillation
Lucas Caccia
Alan Ansell
Ivan Vulić
Edoardo Ponti
Alessandro Sordoni
Dynamically integrating new or rapidly evolving information after Language Model (LM) pre-training remains challenging, particularly in low-… (voir plus)data scenarios or when dealing with private and specialized documents. In-context learning and retrieval-augmented generation (RAG) face limitations, including their high inference costs and their inability to capture global document information. In this paper, we propose a way of modularizing knowledge by training Knowledge Modules (KMs). KMs are lightweight components implemented as parameter-efficient LoRA modules, which are trained to store information about new documents and can be easily plugged into models on demand. We show that next-token prediction performs poorly in training KMs. We instead propose Deep Context Distillation: we learn KMs parameters such as to simulate hidden states and logits of a teacher that takes the document in context. Our method outperforms standard next-token prediction and pre-instruction training techniques, across two datasets. Finally, we highlight synergies between KMs and retrieval-augmented generation.
2025-03-05
ICLR.cc/2025/Workshop/MCDC (accepté)
openreview.net
openreview.net
Training Plug-n-Play Knowledge Modules with Deep Context Distillation
Lucas Caccia
Alan Ansell
Ivan Vulić
Edoardo Ponti
Alessandro Sordoni
2025-03-05
ICLR.cc/2025/Workshop/MCDC (accepté)
openreview.net
openreview.net
TypyBench: Evaluating LLM Type Inference for Untyped Python Repositories
Yuhe Jiang
Xun Deng
Jiacheng Yang
Honghua Dong
Gennady Pekhimenko
Fan Long
Xujie Si
Type inference for dynamic languages like Python is a persistent challenge in software engineering. While large language models (LLMs) have … (voir plus)shown promise in code understanding, their type inference capabilities remain underexplored. We introduce `TypyBench`, a benchmark designed to evaluate LLMs' type inference across entire Python repositories. `TypyBench` features two novel metrics: `TypeSim`, which captures nuanced semantic relationships between predicted and ground truth types, and `TypeCheck`, which assesses type consistency across codebases. Our evaluation of various LLMs on a curated dataset of 50 high-quality Python repositories reveals that, although LLMs achieve decent `TypeSim` scores, they struggle with complex nested types and exhibit significant type consistency errors. These findings suggest that future research should shift focus from improving type similarity to addressing repository-level consistency. `TypyBench` provides a foundation for this new direction, offering insights into model performance across different type complexities and usage contexts.
2025-03-05
ICLR.cc/2025/Workshop/DL4C (publié)
openreview.net
openreview.net
Understanding (Un)Reliability of Steering Vectors in Language Models
Joschka Braun
Carsten Eickhoff
David Scott Krueger
Seyed Ali Bahrainian
Dmitrii Krasheninnikov
Steering vectors are a lightweight method to control language model behavior by adding a learned bias to the activations at inference time. … (voir plus)Although steering demonstrates promising performance, recent work shows that it can be unreliable or even counterproductive in some cases. This paper studies the influence of prompt types and the geometry of activation differences on steering reliability. First, we find that all seven prompt types used in our experiments produce a net positive steering effect, but exhibit high variance across samples, and often give an effect opposite of the desired one. No prompt type clearly outperforms the others, and yet the steering vectors resulting from the different prompt types often differ directionally (as measured by cosine similarity). Second, we show that higher cosine similarity between training set activation differences predicts more effective steering. Finally, we observe that datasets where positive and negative activations are better separated are more steerable. Our results suggest that vector steering is unreliable when the target behavior is not represented by a coherent direction.
2025-03-05
ICLR.cc/2025/Workshop/BuildingTrust (accepté)
openreview.net
openreview.net
UNLEARNING GEO-CULTURAL STEREOTYPES IN MULTILINGUAL LLMS
Alireza Dehghanpour Farashah
Aditi Khandelwal
Negar Rostamzadeh
Golnoosh Farnadi
As multilingual generative models become more widely used, most safety and fairness evaluation techniques still focus on English-language re… (voir plus)sources, while overlooking important cross-cultural factors. This limitation raises concerns about fairness and safety, particularly regarding geoculturally situated stereotypes that hinder the models’ global inclusivity. In this work, we present preliminary findings on the impact of stereotype unlearning across languages, specifically in English, French, and Hindi. Using an adapted version of the SeeGULL dataset, we analyze how unlearning stereotypes in one language influences other languages within multilingual large language models. Our study evaluates two model families, Llama-3.1-8B and Aya-Expanse-8B, to assess whether unlearning in one linguistic context transfers across languages, potentially mitigating or exacerbating biases in multilingual settings.
2025-03-05
ICLR.cc/2025/Workshop/BuildingTrust (accepté)
openreview.net
openreview.net
WebMMU: A Benchmark for Multimodal Multilingual Website Understanding and Code Generation
Rabiul Awal
Mahsa Massoud
Zichao Li
Aarash Feizi
Suyuchen Wang
Chris Pal
Aishwarya Agrawal
David Vazquez
Siva Reddy
Juan A. Rodriguez
Perouz Taslakian
Spandana Gella
Sai Rajeswar
Understanding diverse web data and automating web development presents an exciting challenge for agentic AI. While existing benchmarks addre… (voir plus)ss isolated web-based tasks—such as website-based Visual Question Answering (VQA) and UI-to-code generation—they lack a unified evaluation suite for assessing web agents that interact with and reason about web environments. We introduce WebMMU, a large-scale benchmark for evaluating AI-driven web agents across multilingual website VQA, HTML/CSS/JavaScript code editing, and sketch-to-code generation. WebMMU provides a comprehensive evaluation suite with real-world website data, multi-step reasoning tasks, and functional UI understanding. Benchmarking state-of-the-art multimodal models on WebMMU reveals significant limitations in web-based reasoning, layout understanding, and structured code generation, particularly in preserving UI hierarchy, handling multilingual content, and producing robust, functional code. While most existing models are optimized for English-only settings, WebMMU highlights the challenges of cross-lingual adaptation in real-world web development. These findings expose critical gaps in current models’ ability to understand website structures, execute user instructions, and generate high-quality web code, underscoring the need for more advanced multimodal reasoning in AI-driven web understanding and development.
2025-03-05
ICLR.cc/2025/Workshop/DL4C (publié)
openreview.net
openreview.net
Automated diagnosis of usual interstitial pneumonia on chest CT via the mean curvature of isophotes
Peter Savadjiev
Morteza Rezanejad
Sahir Bhatnagar
David Camirand
Claude Kauffmann
Kaleem Siddiqi
Ronald J Dandurand
Patrick Bourgouin
Carl Chartrand-Lefebvre
Alexandre Semionov
2025-03-04
medRxiv (prépublication)
doi.org
AI Automatons: AI Systems Intended to Imitate Humans
Alexandra Olteanu
Solon Barocas
Su Lin Blodgett
Lisa Egede
Alicia DeVrio
Myra Cheng
There is a growing proliferation of AI systems designed to mimic people's behavior, work, abilities, likenesses, or humanness -- systems we … (voir plus)dub AI automatons. Individuals, groups, or generic humans are being simulated to produce creative work in their styles, to respond to surveys in their places, to probe how they would use a new system before deployment, to provide users with assistance and companionship, and to anticipate their possible future behavior and interactions with others, just to name a few applications. The research, design, deployment, and availability of such AI systems have, however, also prompted growing concerns about a wide range of possible legal, ethical, and other social impacts. To both 1) facilitate productive discussions about whether, when, and how to design and deploy such systems, and 2) chart the current landscape of existing and prospective AI automatons, we need to tease apart determinant design axes and considerations that can aid our understanding of whether and how various design choices along these axes could mitigate -- or instead exacerbate -- potential adverse impacts that the development and use of AI automatons could give rise to. In this paper, through a synthesis of related literature and extensive examples of existing AI systems intended to mimic humans, we develop a conceptual framework to help foreground key axes of design variations and provide analytical scaffolding to foster greater recognition of the design choices available to developers, as well as the possible ethical implications these choices might have.
2025-03-04
ArXiv (prépublication)
arxiv.org
arxiv.org
AI Automatons: AI Systems Intended to Imitate Humans
Alexandra Olteanu
Solon Barocas
Su Lin Blodgett
Lisa Egede
Alicia DeVrio
Myra Cheng
There is a growing proliferation of AI systems designed to mimic people's behavior, work, abilities, likenesses, or humanness -- systems we … (voir plus)dub AI automatons. Individuals, groups, or generic humans are being simulated to produce creative work in their styles, to respond to surveys in their places, to probe how they would use a new system before deployment, to provide users with assistance and companionship, and to anticipate their possible future behavior and interactions with others, just to name a few applications. The research, design, deployment, and availability of such AI systems have, however, also prompted growing concerns about a wide range of possible legal, ethical, and other social impacts. To both 1) facilitate productive discussions about whether, when, and how to design and deploy such systems, and 2) chart the current landscape of existing and prospective AI automatons, we need to tease apart determinant design axes and considerations that can aid our understanding of whether and how various design choices along these axes could mitigate -- or instead exacerbate -- potential adverse impacts that the development and use of AI automatons could give rise to. In this paper, through a synthesis of related literature and extensive examples of existing AI systems intended to mimic humans, we develop a conceptual framework to help foreground key axes of design variations and provide analytical scaffolding to foster greater recognition of the design choices available to developers, as well as the possible ethical implications these choices might have.
2025-03-04
ArXiv (prépublication)
arxiv.org
arxiv.org
Beyond Cosine Decay: On the effectiveness of Infinite Learning Rate Schedule for Continual Pre-training
Vaibhav Singh
Paul Janson
Paria Mehrbod
Adam Ibrahim
Irina Rish
Eugene Belilovsky
Benjamin Thérien
The ever-growing availability of unlabeled data presents both opportunities and challenges for training artificial intelligence systems. Whi… (voir plus)le self-supervised learning (SSL) has emerged as a powerful paradigm for extracting meaningful representations from vast amounts of unlabeled data, existing methods still struggle to adapt to the non-stationary, non-IID nature of real-world data streams without forgetting previously learned knowledge. Recent works have adopted a repeated cosine annealing schedule for large-scale continual pre-training; however, these schedules (1) inherently cause forgetting during the re-warming phase and (2) have not been systematically compared to existing continual SSL methods. In this work, we systematically compare the widely used cosine schedule with the recently proposed infinite learning rate schedule and empirically find the latter to be a more effective alternative. Our extensive empirical evaluation across diverse image and language datasets demonstrates that the infinite learning rate schedule consistently enhances continual pre-training performance compared to a repeated cosine decay without being restricted to a fixed iteration budget. For instance, in a small-scale MAE pre-training setup, it outperforms several strong baselines from the literature. We then scale up our experiments to larger MAE pre-training and autoregressive language model pre-training. Our results show that the infinite learning rate schedule remains effective at scale, surpassing repeated cosine decay for both MAE pre-training and zero-shot LM benchmarks.
2025-03-04
ArXiv (prépublication)
arxiv.org
arxiv.org
Beyond Cosine Decay: On the effectiveness of Infinite Learning Rate Schedule for Continual Pre-training
Vaibhav Singh
Paul Janson
Paria Mehrbod
Adam Ibrahim
Irina Rish
Eugene Belilovsky
Benjamin Thérien
The ever-growing availability of unlabeled data presents both opportunities and challenges for training artificial intelligence systems. Whi… (voir plus)le self-supervised learning (SSL) has emerged as a powerful paradigm for extracting meaningful representations from vast amounts of unlabeled data, existing methods still struggle to adapt to the non-stationary, non-IID nature of real-world data streams without forgetting previously learned knowledge. Recent works have adopted a repeated cosine annealing schedule for large-scale continual pre-training; however, these schedules (1) inherently cause forgetting during the re-warming phase and (2) have not been systematically compared to existing continual SSL methods. In this work, we systematically compare the widely used cosine schedule with the recently proposed infinite learning rate schedule and empirically find the latter to be a more effective alternative. Our extensive empirical evaluation across diverse image and language datasets demonstrates that the infinite learning rate schedule consistently enhances continual pre-training performance compared to a repeated cosine decay without being restricted to a fixed iteration budget. For instance, in a small-scale MAE pre-training setup, it outperforms several strong baselines from the literature. We then scale up our experiments to larger MAE pre-training and autoregressive language model pre-training. Our results show that the infinite learning rate schedule remains effective at scale, surpassing repeated cosine decay for both MAE pre-training and zero-shot LM benchmarks.
2025-03-04
ArXiv (prépublication)
arxiv.org
arxiv.org