Publications

Training Dynamics Underlying Language Model Scaling Laws: Loss Deceleration and Zero-Sum Learning
Andrei Mircea
Supriyo Chakraborty
Nima Chitsazan
Irina Rish
Ekaterina Lobacheva
This work aims to understand how scaling improves language models, specifically in terms of training dynamics. We find that language models … (voir plus)undergo loss deceleration early in training; an abrupt slowdown in the rate of loss improvement, resulting in piecewise linear behaviour of the loss curve in log-log space. Scaling up the model mitigates this transition by (1) decreasing the loss at which deceleration occurs, and (2) improving the log-log rate of loss improvement after deceleration. We attribute loss deceleration to a type of degenerate training dynamics we term zero-sum learning (ZSL). In ZSL, per-example gradients become systematically opposed, leading to destructive interference in per-example changes in loss. As a result, improving loss on one subset of examples degrades it on another, bottlenecking overall progress. Loss deceleration and ZSL provide new insights into the training dynamics underlying language model scaling laws, and could potentially be targeted directly to improve language models independent of scale. We make our code and artefacts available at: https://github.com/mirandrom/zsl
2024-12-31
ACL (1) (publié)
doi.org
arxiv.org
Training Language Models to Self-Correct via Reinforcement Learning
Aviral Kumar
Vincent Zhuang
Rishabh Agarwal
Yi Su
John D Co-Reyes
Avi Singh
Kate Baumli
Shariq Iqbal
Colton Bishop
Rebecca Roelofs
Lei M Zhang
Kay McKinney
Disha Shrivastava
Cosmin Paduraru
George Tucker
Doina Precup
Feryal Behbahani
Aleksandra Faust
Self-correction is a highly desirable capability of large language models (LLMs), yet it has consistently been found to be largely ineffecti… (voir plus)ve in modern LLMs. Existing approaches for training self-correction either require multiple models or rely on a more capable model or other forms of supervision. To this end, we develop a multi-turn online reinforcement learning (RL) approach, SCoRe, that significantly improves an LLM's self-correction ability using entirely self-generated data. To build SCoRe, we first show that variants of supervised fine-tuning (SFT) on offline model-generated correction traces are insufficient for instilling self-correction behavior. In particular, we observe that training via SFT either suffers from a distribution mismatch between the training data and the model's own responses or implicitly prefers only a certain mode of correction behavior that is often not effective at test time. SCoRe addresses these challenges by training under the model's own distribution of self-generated correction traces and using appropriate regularization to steer the learning process into learning a self-correction strategy that is effective at test time as opposed to simply fitting high-reward responses for a given prompt. This regularization prescribes running a first phase of RL on a base model to generate a policy initialization that is less susceptible to collapse and then using a reward bonus to amplify self-correction during training. When applied to Gemini 1.0 Pro and 1.5 Flash models, we find that SCoRe achieves state-of-the-art self-correction performance, improving the base models' self-correction by 15.6% and 9.1% respectively on the MATH and HumanEval benchmarks.
2024-12-31
ICLR (publié)
doi.org
openreview.net
Training of LLM-Based List-Wise Multilingual Reranker
Hao Yu
David Ifeoluwa Adelani
2024-12-31
Proceedings of the 5th Workshop on Multilingual Representation Learning (MRL 2025) (publié)
doi.org
TrajGPT: Irregular Time-Series Representation Learning of Health Trajectory
Ziyang Song
Qincheng Lu
He Zhu
David Buckeridge
Yue Li
In the healthcare domain, time-series data are often irregularly sampled with varying intervals through outpatient visits, posing challenges… (voir plus) for existing models designed for equally spaced sequential data. To address this, we propose Trajectory Generative Pre-trained Transformer (TrajGPT) for representation learning on irregularly-sampled healthcare time series. TrajGPT introduces a novel Selective Recurrent Attention (SRA) module that leverages a data-dependent decay to adaptively filter irrelevant past information. As a discretized ordinary differential equation (ODE) framework, TrajGPT captures underlying continuous dynamics and enables a time-specific inference for forecasting arbitrary target timesteps without auto-regressive prediction. Experimental results based on the longitudinal EHR data PopHR from Montreal health system and eICU from PhysioNet showcase TrajGPT's superior zero-shot performance in disease forecasting, drug usage prediction, and sepsis detection. The inferred trajectories of diabetic and cardiac patients reveal meaningful comorbidity conditions, underscoring TrajGPT as a useful tool for forecasting patient health evolution.
2024-12-31
IEEE Journal of Biomedical and Health Informatics (inconnu)
doi.org
TransCeption: Enhancing medical image segmentation with an inception-like transformer design for efficient feature fusion
Reza Azad
Yiwei Jia
Ehsan Khodapanah Aghdam
Julien Cohen-Adad
Dorit Merhof
2024-12-31
Comput. Vis. Media (publié)
doi.org
Tree semantic segmentation from aerial image time series
Venkatesh Ramesh
Arthur Ouaknine
David Rolnick
Earth's forests play an important role in the fight against climate change, and are in turn negatively affected by it. Effective monitoring … (voir plus)of different tree species is essential to understanding and improving the health and biodiversity of forests. In this work, we address the challenge of tree species identification by performing semantic segmentation of trees using an aerial image dataset spanning over a year. We compare models trained on single images versus those trained on time series to assess the impact of tree phenology on segmentation performances. We also introduce a simple convolutional block for extracting spatio-temporal features from image time series, enabling the use of popular pretrained backbones and methods. We leverage the hierarchical structure of tree species taxonomy by incorporating a custom loss function that refines predictions at three levels: species, genus, and higher-level taxa. Our findings demonstrate the superiority of our methodology in exploiting the time series modality and confirm that enriching labels using taxonomic information improves the semantic segmentation performance.
2024-12-31
Environmental Data Science (publié)
doi.org
arxiv.org
Understanding and Meeting Practitioner Needs When Measuring Representational Harms Caused by LLM-Based Systems
Emma Harvey
Emily Sheng
Su Lin Blodgett
Alexandra Chouldechova
Jean Garcia-Gathright
A.R. Olteanu
Hanna Wallach
The NLP research community has made publicly available numerous instruments for measuring representational harms caused by large language mo… (voir plus)del (LLM)-based systems. These instruments have taken the form of datasets, metrics, tools, and more. In this paper, we examine the extent to which such instruments meet the needs of practitioners tasked with evaluating LLM-based systems. Via semi-structured interviews with 12 such practitioners, we find that practitioners are often unable to use publicly available instruments for measuring representational harms. We identify two types of challenges. In some cases, instruments are not useful because they do not meaningfully measure what practitioners seek to measure or are otherwise misaligned with practitioner needs. In other cases, instruments - even useful instruments - are not used by practitioners due to practical and institutional barriers impeding their uptake. Drawing on measurement theory and pragmatic measurement, we provide recommendations for addressing these challenges to better meet practitioner needs.
2024-12-31
ACL (Findings) (publié)
doi.org
arxiv.org
Unified Game Moderation: Soft-Prompting and LLM-Assisted Label Transfer for Resource-Efficient Toxicity Detection
Zachary Yang
Domenico Tullo
Reihaneh Rabbany
Toxicity detection in gaming communities faces significant scaling challenges when expanding across multiple games and languages, particular… (voir plus)ly in real-time environments where computational efficiency is crucial. We present two key findings to address these challenges while building upon our previous work on ToxBuster, a BERT-based real-time toxicity detection system. First, we introduce a soft-prompting approach that enables a single model to effectively handle multiple games by incorporating game-context tokens, matching the performance of more complex methods like curriculum learning while offering superior scalability. Second, we develop an LLM-assisted label transfer framework using GPT-4o-mini to extend support to seven additional languages. Evaluations on real game chat data across French, German, Portuguese, and Russian achieve macro F1-scores ranging from 32.96% to 58.88%, with particularly strong performance in German, surpassing the English benchmark of 45.39%. In production, this unified approach significantly reduces computational resources and maintenance overhead compared to maintaining separate models for each game and language combination. At Ubisoft, this model successfully identifies an average of 50 players, per game, per day engaging in sanctionable behavior.
2024-12-31
arXiv (prépublication)
doi.org
arxiv.org
Variation Matters: from Mitigating to Embracing Zero-Shot NAS Ranking Function Variation
Pavel Rumiantsev
Mark J. Coates
Neural Architecture Search (NAS) is a powerful automatic alternative to manual design of a neural network. In the zero-shot version, a fast … (voir plus)ranking function is used to compare architectures without training them. The outputs of the ranking functions often vary significantly due to different sources of randomness, including the evaluated architecture's weights' initialization or the batch of data used for calculations. A common approach to addressing the variation is to average a ranking function output over several evaluations. We propose taking into account the variation in a different manner, by viewing the ranking function output as a random variable representing a proxy performance metric. During the search process, we strive to construct a stochastic ordering of the performance metrics to determine the best architecture. Our experiments show that the proposed stochastic ordering can effectively boost performance of a search on standard benchmark search spaces.
2024-12-31
Trans. Mach. Learn. Res. (publié)
doi.org
openreview.net
Visual-Tactile Inference of 2.5D Object Shape From Marker Texture.
Affan Jilani
François Robert Hogan
Charlotte Morissette
Gregory Dudek
Michael Jenkin
Kaleem Siddiqi
Visual-tactile sensing affords abundant capabilities for contact-rich object manipulation tasks including grasping and placing. Here we intr… (voir plus)oduce a shape-from-texture inspired contact shape estimation approach for visual-tactile sensors equipped with visually distinct membrane markers. Under a perspective projection camera model, measurements related to the change in marker separation upon contact are used to recover surface shape. Our approach allows for shape sensing in real time, without requiring network training or complex assumptions related to lighting, sensor geometry or marker placement. Experiments show that the surface contact shape recovered is qualitatively and quantitatively consistent with those obtained through the use of photometric stereo, the current state of the art for shape recovery in visual-tactile sensors. Importantly, our approach is applicable to a large family of sensors not equipped with photometric stereo hardware, and also to those with semi-transparent membranes. The recovery of surface shape affords new capabilities to these sensors for robotic applications, such as the estimation of contact and slippage in object manipulation tasks (Hogan etal., 2022) and the use of force matching for kinesthetic teaching using multimodal visual-tactile sensing (Ablett etal., 2024).
2024-12-31
IEEE Robotics and Automation Letters (publié)
doi.org
Warmup Generations: A Task-Agnostic Approach for Guiding Sequence-to-Sequence Learning with Unsupervised Initial State Generation
Senyu Li
Zipeng Sun
Jiayi Wang
Xue Liu
Pontus Stenetorp
Siva Reddy
David Ifeoluwa Adelani
Traditional supervised fine-tuning (SFT) strategies for sequence-to-sequence tasks often train models to directly generate the target output… (voir plus). Recent work has shown that guiding models with intermediate steps, such as keywords, outlines, or reasoning chains, can significantly improve performance, coherence, and interpretability. However, these methods often depend on predefined intermediate formats and annotated data, limiting their scalability and generalizability. In this work, we introduce a task-agnostic framework that enables models to generate intermediate "warmup" sequences. These warmup sequences, serving as an initial state for subsequent generation, are optimized to enhance the probability of generating the target sequence without relying on external supervision or human-designed structures. Drawing inspiration from reinforcement learning principles, our method iteratively refines these intermediate steps to maximize their contribution to the final output, similar to reward-driven optimization in reinforcement learning with human feedback. Experimental results across tasks such as translation, summarization, and multi-choice question answering for logical reasoning show that our approach outperforms traditional SFT methods, and offers a scalable and flexible solution for sequence-to-sequence tasks.
2024-12-31
Association for Computational Linguistics (publié)
doi.org
arxiv.org
In Which Areas of Technical AI Safety Could Geopolitical Rivals Cooperate?
BEN BUCKNALL
Saad Siddiqui
LARA THURNHERR
CONOR MCGURK
BEN HARACK
Anka Reuel
PATRICIA PASKOV
CASEY MAHONEY
Sören Mindermann
Scott Singer
VINAY HIREMATH
Charbel-Raphael Segerie
OSCAR DELANEY
Alessandro Abate
Fazl Barez
Michael K. Cohen
Philip Torr
FERENC HUSZÁR
ANISOARA CALINESCU
GABRIEL DAVIS JONES … (voir 2 de plus)
Yoshua Bengio
Robert Trager
International cooperation is common in AI research, including between geopolitical rivals. While many experts advocate for greater internati… (voir plus)onal cooperation on AI safety to address shared global risks, some view cooperation on AI with suspicion, arguing that it can pose unacceptable risks to national security. However, the extent to which cooperation on AI safety poses such risks, as well as provides benefits, depends on the specific area of cooperation. In this paper, we consider technical factors that impact the risks of international cooperation on AI safety research, focusing on the degree to which such cooperation can advance dangerous capabilities, result in the sharing of sensitive information, or provide opportunities for harm. We begin by why nations historically cooperate on strategic technologies and analyse current US-China cooperation in AI as a case study. We further argue that existing frameworks for managing associated risks can be supplemented with consideration of key risks specific to cooperation on technical AI safety research. Through our analysis, we find that research into AI verification mechanisms and shared protocols may be suitable areas for such cooperation. Through this analysis we aim to help researchers and governments identify and mitigate the risks of international cooperation on AI safety research, so that the benefits of cooperation can be fully realised.
2024-12-31
ACM Conference on Fairness, Accountability, and Transparency (publié)
doi.org
arxiv.org