Publications

Controlling Large Language Model Agents with Entropic Activation Steering

Nathan Rahn

Pierluca D'Oro

Marc Gendron-Bellemare

The generality of pretrained large language models (LLMs) has prompted increasing interest in their use as in-context learning agents. To be… (see more) successful, such agents must form beliefs about how to achieve their goals based on limited interaction with their environment, resulting in uncertainty about the best action to take at each step. In this paper, we study how LLM agents form and act on these beliefs by conducting experiments in controlled sequential decision-making tasks. To begin, we find that LLM agents are overconfident: They draw strong conclusions about what to do based on insufficient evidence, resulting in inadequately explorative behavior. We dig deeper into this phenomenon and show how it emerges from a collapse in the entropy of the action distribution implied by sampling from the LLM. We then demonstrate that existing token-level sampling techniques are by themselves insufficient to make the agent explore more. Motivated by this fact, we introduce Entropic Activation Steering (EAST), an activation steering method for in-context LLM agents. EAST computes a steering vector as an entropy-weighted combination of representations, and uses it to manipulate an LLM agent's uncertainty over actions by intervening on its activations during the forward pass. We show that EAST can reliably increase the entropy in an LLM agent's actions, causing more explorative behavior to emerge. Finally, EAST modifies the subjective uncertainty an LLM agent expresses, paving the way to interpreting and controlling how LLM agents represent uncertainty about their decisions.

2024-06-24

ICML.cc/2024/Workshop/MI (poster)

Interpretability in Action: Exploratory Analysis of VPT, a Minecraft Agent

Karolis Jucys

George Adamopoulos

Mehrab Hamidi

Stephanie Milani

Mohammad Reza Samsami

Artem Zholus

Sonia Joseph

Blake Richards

Irina Rish

Özgür Şimşek

Understanding the mechanisms behind decisions taken by large foundation models in sequential tasks is critical to ensuring that such systems… (see more) operate transparently and safely. However, interpretability methods have not yet been applied extensively to large-scale agents based on reinforcement learning. In this work, we perform exploratory analysis on the Video PreTraining (VPT) Minecraft playing agent, one of the largest open-source vision-based agents. We try to illuminate its reasoning mechanisms by applying various interpretability techniques. First, we analyze the attention mechanism while the agent solves its training task --- crafting a diamond pickaxe. The agent seems to pay attention to the 4 last frames and several key-frames further back. This provides clues as to how it maintains coherence in the task that takes 3-10 minutes, despite the agent's short memory span of only six seconds. Second, we perform various interventions, which help us uncover a worrying case of goal misgeneralization: VPT mistakenly identifies a villager wearing brown clothes as a tree trunk and punches it to death, when positioned stationary under green tree leaves. We demonstrate similar misbehavior in a related agent (STEVE-1), which motivates the use of VPT as a model organism for large-scale vision-based agent interpretability.

2024-06-24

ICML.cc/2024/Workshop/MI (poster)

Gintare Karolina Dziugaite

Robust Unlearning via Mechanistic Localizations

Phillip Huang Guo

Aaquib Syed

Abhay Sheshadri

Aidan Ewart

Methods for machine unlearning in large language models seek to remove undesirable knowledge or capabilities without compromising general la… (see more)nguage modeling performance. This work investigates the use of mechanistic interpretability to improve the precision and effectiveness of unlearning. We demonstrate that localizing unlearning to components with particular mechanisms in factual recall leads to more robust unlearning across different input/output formats, relearning, and latent knowledge, and reduces unintended side effects compared to nonlocalized unlearning. Additionally, we analyze the strengths and weaknesses of different automated (rather than manual) interpretability methods for guiding unlearning, finding that their corresponding unlearned models require smaller edit sizes to achieve unlearning but are much less robust.

2024-06-24

ICML.cc/2024/Workshop/MI (spotlight)

The Responsible Foundation Model Development Cheatsheet: A Review of Tools&Resources

Shayne Longpre

Stella Biderman

Alon Albalak

Hailey Schoelkopf

Daniel McDuff

Sayash Kapoor

Kevin Klyman

Kyle Lo

Gabriel Ilharco

Nay San

Maribeth Rauh

Aviya Skowron

Bertie Vidgen

Laura Weidinger

Arvind Narayanan

Victor Sanh

David Ifeoluwa Adelani

Percy Liang

Rishi Bommasani

Peter Henderson 0002 … (see 3 more)

Sasha Luccioni

Yacine Jernite

Luca Soldaini

2024-06-24

ArXiv (preprint)

Samira Abbasgholizadeh-Rahimi

Existing Digital Health Technology Index Summary Report for Older Adults Living with Neurocognitive Disorders (Mild and Major) and Their Informal Caregivers: An Environmental Scan

Ambily Jose

Maxime Sasseville

Ellen Gorus

Anik Giguère

Anne Bourbonnais

Clémence Balley

Ronald Buyl

Marie-Pierre Gagnon

Digital health has added numerous promising solutions to enhance the health and wellness of people with neurocognitive disorders (NCDs) and … (see more)their informal caregivers. (1) Background: It is important to obtain a comprehensive view of currently available technologies, their outcomes, and conditions of success to inform recommendations regarding digital health solutions for people with NCDs and their caregivers. This environmental scan was performed to identify the features of existing digital health solutions relevant to the targeted population. This work reviews currently available digital health solutions and their related characteristics to develop a decision support tool for older adults living with mild or major neurocognitive disorders and their informal caregivers. This knowledge will aid the development of a decision support tool to assist older adults and their informal caregivers in their search for adequate digital health solutions according to their needs and preferences based on trustable information. (2) Methods: We conducted an environmental scan to identify digital health solutions from a systematic review and targeted searches in the grey literature covering the regions of Canada and Europe. Technological tools were scanned based on a preformatted extraction grid. We assessed their relevance based on selected attributes and summarized the findings. (3) Results: We identified 100 available digital health solutions. The majority (56%) were not specific to NCDs. Only 28% provided scientific evidence of their effectiveness. Remote patient care, movement tracking, and cognitive exercises were the most common purposes of digital health solutions. Most solutions were presented as decision aid tools, pill dispensers, apps, web, or a combination of these platforms. (4) Conclusions: This environmental scan allowed for identifying current digital health solutions for older adults with mild or major neurocognitive disorders and their informal caregivers. Findings from the environmental scan highlight the need for additional approaches to strengthen digital health interventions for the well-being of older adults with mild and major NCDs and their informal and formal healthcare providers.

2024-06-22

Geriatrics (published)

Laurence Perreault-Levasseur

Caustics: A Python Package for Accelerated Strong Gravitational Lensing Simulations

Connor Stone

Alexandre Adam

Adam Coogan

M. J. Yantovski-Barth

Andreas Filipp

Landung Setiawan

Cordero Core

Ronan Legin

Charles Wilson

Gabriel Missael Barco

Yashar Hezaveh

2024-06-21

ArXiv (preprint)

DASB -- Discrete Audio and Speech Benchmark

Pooneh Mousavi

Luca Della Libera

Jarod Duret

Artem Ploujnikov

Cem Subakan

Mirco Ravanelli

Discrete audio tokens have recently gained considerable attention for their potential to connect audio and language processing, enabling the… (see more) creation of modern multimodal large language models. Ideal audio tokens must effectively preserve phonetic and semantic content along with paralinguistic information, speaker identity, and other details. While several types of audio tokens have been recently proposed, identifying the optimal tokenizer for various tasks is challenging due to the inconsistent evaluation settings in existing studies. To address this gap, we release the Discrete Audio and Speech Benchmark (DASB), a comprehensive leaderboard for benchmarking discrete audio tokens across a wide range of discriminative tasks, including speech recognition, speaker identification and verification, emotion recognition, keyword spotting, and intent classification, as well as generative tasks such as speech enhancement, separation, and text-to-speech. Our results show that, on average, semantic tokens outperform compression tokens across most discriminative and generative tasks. However, the performance gap between semantic tokens and standard continuous representations remains substantial, highlighting the need for further research in this field.

2024-06-20

ArXiv (preprint)

Language Model-In-The-Loop: Data Optimal Approach to Recommend Actions in Text Games

Arjun V Sudhakar

Prasanna Parthasarathi

Janarthanan Rajendran

Sarath Chandar

Large Language Models (LLMs) have demonstrated superior performance in language understanding benchmarks. A recent use case for LLMs involve… (see more)s training decision-making agents over textual information. The existing approach leverages LLM's linguistic priors for action candidate recommendations in text games, i.e., to operate without environment-provided actions. However, adapting LLMs to specific games/tasks requires a massive amount of annotated human gameplay. Moreover, in the existing approach, the language model was kept frozen during an agent's training process, which limits learning from in-game knowledge about the world. Hence, we explore strategies to adapt the language model for candidate recommendation with in-game transition in an online learning fashion to mitigate reliance on human-annotated gameplays, which are costly to acquire. In this paper, we propose in-game transition selection methods to adapt the LLM in the loop, reducing the dependency on using human-annotated gameplays while improving performance and convergence. Our method demonstrates a 53% relative improvement in average game score over the previous state-of-the-art model, achieving more than twice the convergence rate in a full-annotated dataset setting. Furthermore, even with only 10% of human annotation, we surpassed the 100\% state-of-the-art performance benchmark.

2024-06-20

rl-conference.cc/RLC/2024/Workshop/TAFM (published)

Performance Control in Early Exiting to Deploy Large Models at the Same Cost of Smaller Ones

Mehrnaz Mofakhami

Reza Bayat

Ioannis Mitliagkas

Joao Monteiro

Valentina Zantedeschi

2024-06-20

ICML.cc/2024/Workshop/ES-FoMo-II (poster)

Advantage Alignment Algorithms

Juan Agustin Duque

Milad Aghajohari

Tim Cooijmans

Tianyu Zhang

Aaron Courville

2024-06-19

ICML.cc/2024/Workshop/ARLET (poster)

APPL: A Prompt Programming Language for Harmonious Integration of Programs and Large Language Model Prompts

Honghua Dong

Qidong Su

Yubo Gao

Zhaoyu Li

Yangjun Ruan

Gennady G. Pekhimenko

Chris J. Maddison

Xujie Si

Large Language Models (LLMs) have become increasingly capable of handling diverse tasks with the aid of well-crafted prompts and integration… (see more) of external tools, but as task complexity rises, the workflow involving LLMs can be complicated and thus challenging to implement and maintain. To address this challenge, we propose APPL, A Prompt Programming Language that acts as a bridge between computer programs and LLMs, allowing seamless embedding of prompts into Python functions, and vice versa. APPL provides an intuitive and Python-native syntax, an efficient parallelized runtime with asynchronous semantics, and a tracing module supporting effective failure diagnosis and replaying without extra costs. We demonstrate that APPL programs are intuitive, concise, and efficient through three representative scenarios: Chain-of-Thought with self-consistency (CoT-SC), ReAct tool use agent, and multi-agent chat. Experiments on three parallelizable workflows further show that APPL can effectively parallelize independent LLM calls, with a significant speedup ratio that almost matches the estimation.

2024-06-19

ArXiv (preprint)