Sarath Chandar

Pranshu Malviya, Gonçalo Mordido, Aristide Baratin et al. · deepmind

Adaptive gradient-based optimizers, notably Adam, have left their mark in training large-scale deep learning models, offering fast convergence and robustness to hyperparameter settings. However, they often struggle with generalization, attributed to their tendency to converge to sharp minima in the loss landscape. To address this, we propose a new memory-augmented version of Adam that encourages exploration towards flatter minima by incorporating a buffer of critical momentum terms during training. This buffer prompts the optimizer to overshoot beyond narrow minima, promoting exploration. Through comprehensive analysis in simple settings, we illustrate the efficacy of our approach in increasing exploration and bias towards flatter minima. We empirically demonstrate that it can improve model performance for image classification on ImageNet and CIFAR10/100, language modelling on Penn Treebank, and online learning tasks on TinyImageNet and 5-dataset. Our code is available at \url{https://github.com/chandar-lab/CMOptimizer}.

Megh Thakkar, Tolga Bolukbasi, Sriram Ganapathy et al. · cmu, deepmind

Language Models (LMs) pre-trained with self-supervision on large text corpora have become the default starting point for developing models for various NLP tasks. Once the pre-training corpus has been assembled, all data samples in the corpus are treated with equal importance during LM pre-training. However, due to varying levels of relevance and quality of data, equal importance to all the data samples may not be the optimal choice. While data reweighting has been explored in the context of task-specific supervised learning and LM fine-tuning, model-driven reweighting for pre-training data has not been explored. We fill this important gap and propose PRESENCE, a method for jointly reweighting samples by leveraging self-influence (SI) scores as an indicator of sample importance and pre-training. PRESENCE promotes novelty and stability for model pre-training. Through extensive analysis spanning multiple model sizes, datasets, and tasks, we present PRESENCE as an important first step in the research direction of sample reweighting for pre-training language models.

Jarrid Rector-Brooks, Kanika Madan, Moksh Jain et al. · mila

Generative flow networks (GFlowNets) are amortized variational inference algorithms that treat sampling from a distribution over compositional objects as a sequential decision-making problem with a learnable action policy. Unlike other algorithms for hierarchical sampling that optimize a variational bound, GFlowNet algorithms can stably run off-policy, which can be advantageous for discovering modes of the target distribution. Despite this flexibility in the choice of behaviour policy, the optimal way of efficiently selecting trajectories for training has not yet been systematically explored. In this paper, we view the choice of trajectories for training as an active learning problem and approach it using Bayesian techniques inspired by methods for multi-armed bandits. The proposed algorithm, Thompson sampling GFlowNets (TS-GFN), maintains an approximate posterior distribution over policies and samples trajectories from this posterior for training. We show in two domains that TS-GFN yields improved exploration and thus faster convergence to the target distribution than the off-policy exploration strategies used in past work.

Shagun Sodhani, Mojtaba Faramarzi, Sanket Vaibhav Mehta et al. · deepmind

This primer is an attempt to provide a detailed summary of the different facets of lifelong learning. We start with Chapter 2 which provides a high-level overview of lifelong learning systems. In this chapter, we discuss prominent scenarios in lifelong learning (Section 2.4), provide 8 Introduction a high-level organization of different lifelong learning approaches (Section 2.5), enumerate the desiderata for an ideal lifelong learning system (Section 2.6), discuss how lifelong learning is related to other learning paradigms (Section 2.7), describe common metrics used to evaluate lifelong learning systems (Section 2.8). This chapter is more useful for readers who are new to lifelong learning and want to get introduced to the field without focusing on specific approaches or benchmarks. The remaining chapters focus on specific aspects (either learning algorithms or benchmarks) and are more useful for readers who are looking for specific approaches or benchmarks. Chapter 3 focuses on regularization-based approaches that do not assume access to any data from previous tasks. Chapter 4 discusses memory-based approaches that typically use a replay buffer or an episodic memory to save subset of data across different tasks. Chapter 5 focuses on different architecture families (and their instantiations) that have been proposed for training lifelong learning systems. Following these different classes of learning algorithms, we discuss the commonly used evaluation benchmarks and metrics for lifelong learning (Chapter 6) and wrap up with a discussion of future challenges and important research directions in Chapter 7.

Kamran Chitsaz, Quentin Fournier, Gonçalo Mordido et al.

The increasing scale of Transformer models has led to an increase in their pre-training computational requirements. While quantization has proven to be effective after pre-training and during fine-tuning, applying quantization in Transformers during pre-training has remained largely unexplored at scale for language modeling. This study aims to explore the impact of quantization for efficient pre-training of Transformers, with a focus on linear layer components. By systematically applying straightforward linear quantization to weights, activations, gradients, and optimizer states, we assess its effects on model efficiency, stability, and performance during training. By offering a comprehensive recipe of effective quantization strategies to be applied during the pre-training of Transformers, we promote high training efficiency from scratch while retaining language modeling ability. Code is available at https://github.com/chandar-lab/EfficientLLMs.

Gabriele Prato, Jerry Huang, Prasannna Parthasarathi et al.

In the age of artificial intelligence, the role of large language models (LLMs) is becoming increasingly central. Despite their growing prevalence, their capacity to consolidate knowledge from different training documents - a crucial ability in numerous applications - remains unexplored. This paper presents the first study examining the capability of LLMs to effectively combine such information within their parameter space. We introduce EpiK-Eval, a novel question-answering benchmark tailored to evaluate LLMs' proficiency in formulating a coherent and consistent knowledge representation from segmented narratives. Evaluations across various LLMs reveal significant weaknesses in this domain. We contend that these shortcomings stem from the intrinsic nature of prevailing training objectives. Consequently, we advocate for refining the approach towards knowledge consolidation, as it harbors the potential to dramatically improve their overall effectiveness and performance. The findings from this study offer insights for developing more robust and reliable LLMs. Our code and benchmark are available at https://github.com/chandar-lab/EpiK-Eval

Andreas Madsen, Siva Reddy, Sarath Chandar · mila

A common approach to explaining NLP models is to use importance measures that express which tokens are important for a prediction. Unfortunately, such explanations are often wrong despite being persuasive. Therefore, it is essential to measure their faithfulness. One such metric is if tokens are truly important, then masking them should result in worse model performance. However, token masking introduces out-of-distribution issues, and existing solutions that address this are computationally expensive and employ proxy models. Furthermore, other metrics are very limited in scope. This work proposes an inherently faithfulness measurable model that addresses these challenges. This is achieved using a novel fine-tuning method that incorporates masking, such that masking tokens become in-distribution by design. This differs from existing approaches, which are completely model-agnostic but are inapplicable in practice. We demonstrate the generality of our approach by applying it to 16 different datasets and validate it using statistical in-distribution tests. The faithfulness is then measured with 9 different importance measures. Because masking is in-distribution, importance measures that themselves use masking become consistently more faithful. Additionally, because the model makes faithfulness cheap to measure, we can optimize explanations towards maximal faithfulness; thus, our model becomes indirectly inherently explainable.

Abdelrahman Zayed, Prasanna Parthasarathi, Goncalo Mordido et al. · mila

Data-driven predictive solutions predominant in commercial applications tend to suffer from biases and stereotypes, which raises equity concerns. Prediction models may discover, use, or amplify spurious correlations based on gender or other protected personal characteristics, thus discriminating against marginalized groups. Mitigating gender bias has become an important research focus in natural language processing (NLP) and is an area where annotated corpora are available. Data augmentation reduces gender bias by adding counterfactual examples to the training dataset. In this work, we show that some of the examples in the augmented dataset can be not important or even harmful for fairness. We hence propose a general method for pruning both the factual and counterfactual examples to maximize the model's fairness as measured by the demographic parity, equality of opportunity, and equality of odds. The fairness achieved by our method surpasses that of data augmentation on three text classification datasets, using no more than half of the examples in the augmented dataset. Our experiments are conducted using models of varying sizes and pre-training settings.

Jerry Huang, Prasanna Parthasarathi, Mehdi Rezagholizadeh et al. · mila

Despite their widespread adoption, large language models (LLMs) remain prohibitive to use under resource constraints, with their ever growing sizes only increasing the barrier for use. One noted issue is the high latency associated with auto-regressive generation, rendering large LLMs use dependent on advanced computing infrastructure. Assisted decoding, where a smaller draft model guides a larger target model's generation, has helped alleviate this, but remains dependent on alignment between the two models. Thus if the draft model is insufficiently capable on some domain relative to the target model, performance can degrade. Alternatively, one can leverage multiple draft models to better cover the expertise of the target, but when multiple black-box draft models are available, selecting an assistant without details about its construction can be difficult. To better understand this decision making problem, we observe it as a contextual bandit, where a policy must choose a draft model based on a context. We show that even without prior knowledge of the draft models, creating an offline dataset from only outputs of independent draft/target models and training a policy over the alignment of these outputs can accelerate performance on multiple domains provided the candidates are effective. Further results show this to hold on various settings with multiple assisted decoding candidates, highlighting its flexibility and the advantageous role that such decision making can play.

Arjun Vaithilingam Sudhakar, Prasanna Parthasarathi, Janarthanan Rajendran et al. · mila

Large Language Models (LLMs) have demonstrated superior performance in language understanding benchmarks. CALM, a popular approach, leverages linguistic priors of LLMs -- GPT-2 -- for action candidate recommendations to improve the performance in text games in Jericho without environment-provided actions. However, CALM adapts GPT-2 with annotated human gameplays and keeps the LLM fixed during the learning of the text based games. In this work, we explore and evaluate updating LLM used for candidate recommendation during the learning of the text based game as well to mitigate the reliance on the human annotated gameplays, which are costly to acquire. We observe that by updating the LLM during learning using carefully selected in-game transitions, we can reduce the dependency on using human annotated game plays for fine-tuning the LLMs. We conducted further analysis to study the transferability of the updated LLMs and observed that transferring in-game trained models to other games did not result in a consistent transfer.

Louis Clouâtre, Prasanna Parthasarathi, Amal Zouaq et al. · mila

Providing better language tools for low-resource and endangered languages is imperative for equitable growth. Recent progress with massively multilingual pretrained models has proven surprisingly effective at performing zero-shot transfer to a wide variety of languages. However, this transfer is not universal, with many languages not currently understood by multilingual approaches. It is estimated that only 72 languages possess a "small set of labeled datasets" on which we could test a model's performance, the vast majority of languages not having the resources available to simply evaluate performances on. In this work, we attempt to clarify which languages do and do not currently benefit from such transfer. To that end, we develop a general approach that requires only unlabelled text to detect which languages are not well understood by a cross-lingual model. Our approach is derived from the hypothesis that if a model's understanding is insensitive to perturbations to text in a language, it is likely to have a limited understanding of that language. We construct a cross-lingual sentence similarity task to evaluate our approach empirically on 350, primarily low-resource, languages.

Simon Guiroy, Christopher Pal, Gonçalo Mordido et al.

Meta-Learning algorithms for few-shot learning aim to train neural networks capable of generalizing to novel tasks using only a few examples. Early-stopping is critical for performance, halting model training when it reaches optimal generalization to the new task distribution. Early-stopping mechanisms in Meta-Learning typically rely on measuring the model performance on labeled examples from a meta-validation set drawn from the training (source) dataset. This is problematic in few-shot transfer learning settings, where the meta-test set comes from a different target dataset (OOD) and can potentially have a large distributional shift with the meta-validation set. In this work, we propose Activation Based Early-stopping (ABE), an alternative to using validation-based early-stopping for meta-learning. Specifically, we analyze the evolution, during meta-training, of the neural activations at each hidden layer, on a small set of unlabelled support examples from a single task of the target tasks distribution, as this constitutes a minimal and justifiably accessible information from the target problem. Our experiments show that simple, label agnostic statistics on the activations offer an effective way to estimate how the target generalization evolves over time. At each hidden layer, we characterize the activation distributions, from their first and second order moments, then further summarized along the feature dimensions, resulting in a compact yet intuitive characterization in a four-dimensional space. Detecting when, throughout training time, and at which layer, the target activation trajectory diverges from the activation trajectory of the source data, allows us to perform early-stopping and improve generalization in a large array of few-shot transfer learning settings, across different algorithms, source and target datasets.

Louis Clouâtre, Prasanna Parthasarathi, Amal Zouaq et al. · mila

Many recent perturbation studies have found unintuitive results on what does and does not matter when performing Natural Language Understanding (NLU) tasks in English. Coding properties, such as the order of words, can often be removed through shuffling without impacting downstream performances. Such insight may be used to direct future research into English NLP models. As many improvements in multilingual settings consist of wholesale adaptation of English approaches, it is important to verify whether those studies replicate or not in multilingual settings. In this work, we replicate a study on the importance of local structure, and the relative unimportance of global structure, in a multilingual setting. We find that the phenomenon observed on the English language broadly translates to over 120 languages, with a few caveats.

Hadi Nekoei, Akilesh Badrinaaraayanan, Amit Sinha et al.

Decentralized cooperative multi-agent deep reinforcement learning (MARL) can be a versatile learning framework, particularly in scenarios where centralized training is either not possible or not practical. One of the critical challenges in decentralized deep MARL is the non-stationarity of the learning environment when multiple agents are learning concurrently. A commonly used and efficient scheme for decentralized MARL is independent learning in which agents concurrently update their policies independently of each other. We first show that independent learning does not always converge, while sequential learning where agents update their policies one after another in a sequence is guaranteed to converge to an agent-by-agent optimal solution. In sequential learning, when one agent updates its policy, all other agent's policies are kept fixed, alleviating the challenge of non-stationarity due to simultaneous updates in other agents' policies. However, it can be slow because only one agent is learning at any time. Therefore it might also not always be practical. In this work, we propose a decentralized cooperative MARL algorithm based on multi-timescale learning. In multi-timescale learning, all agents learn simultaneously, but at different learning rates. In our proposed method, when one agent updates its policy, other agents are allowed to update their policies as well, but at a slower rate. This speeds up sequential learning, while also minimizing non-stationarity caused by other agents updating concurrently. Multi-timescale learning outperforms state-of-the-art decentralized learning methods on a set of challenging multi-agent cooperative tasks in the epymarl(Papoudakis et al., 2020) benchmark. This can be seen as a first step towards more general decentralized cooperative deep MARL methods based on multi-timescale learning.

Yi Wan, Ali Rahimi-Kalahroudi, Janarthanan Rajendran et al.

In recent years, a growing number of deep model-based reinforcement learning (RL) methods have been introduced. The interest in deep model-based RL is not surprising, given its many potential benefits, such as higher sample efficiency and the potential for fast adaption to changes in the environment. However, we demonstrate, using an improved version of the recently introduced Local Change Adaptation (LoCA) setup, that well-known model-based methods such as PlaNet and DreamerV2 perform poorly in their ability to adapt to local environmental changes. Combined with prior work that made a similar observation about the other popular model-based method, MuZero, a trend appears to emerge, suggesting that current deep model-based methods have serious limitations. We dive deeper into the causes of this poor performance, by identifying elements that hurt adaptive behavior and linking these to underlying techniques frequently used in deep model-based RL. We empirically validate these insights in the case of linear function approximation by demonstrating that a modified version of linear Dyna achieves effective adaptation to local changes. Furthermore, we provide detailed insights into the challenges of building an adaptive nonlinear model-based method, by experimenting with a nonlinear version of Dyna.

Hadi Nekoei, Xutong Zhao, Janarthanan Rajendran et al.

Cooperative Multi-agent Reinforcement Learning (MARL) algorithms with Zero-Shot Coordination (ZSC) have gained significant attention in recent years. ZSC refers to the ability of agents to coordinate zero-shot (without additional interaction experience) with independently trained agents. While ZSC is crucial for cooperative MARL agents, it might not be possible for complex tasks and changing environments. Agents also need to adapt and improve their performance with minimal interaction with other agents. In this work, we show empirically that state-of-the-art ZSC algorithms have poor performance when paired with agents trained with different learning methods, and they require millions of interaction samples to adapt to these new partners. To investigate this issue, we formally defined a framework based on a popular cooperative multi-agent game called Hanabi to evaluate the adaptability of MARL methods. In particular, we created a diverse set of pre-trained agents and defined a new metric called adaptation regret that measures the agent's ability to efficiently adapt and improve its coordination performance when paired with some held-out pool of partners on top of its ZSC performance. After evaluating several SOTA algorithms using our framework, our experiments reveal that naive Independent Q-Learning (IQL) agents in most cases adapt as quickly as the SOTA ZSC algorithm Off-Belief Learning (OBL). This finding raises an interesting research question: How to design MARL algorithms with high ZSC performance and capability of fast adaptation to unseen partners. As a first step, we studied the role of different hyper-parameters and design choices on the adaptability of current MARL algorithms. Our experiments show that two categories of hyper-parameters controlling the training data diversity and optimization process have a significant impact on the adaptability of Hanabi agents.

Xutong Zhao, Yangchen Pan, Chenjun Xiao et al.

Efficient exploration is critical in cooperative deep Multi-Agent Reinforcement Learning (MARL). In this work, we propose an exploration method that effectively encourages cooperative exploration based on the idea of sequential action-computation scheme. The high-level intuition is that to perform optimism-based exploration, agents would explore cooperative strategies if each agent's optimism estimate captures a structured dependency relationship with other agents. Assuming agents compute actions following a sequential order at \textit{each environment timestep}, we provide a perspective to view MARL as tree search iterations by considering agents as nodes at different depths of the search tree. Inspired by the theoretically justified tree search algorithm UCT (Upper Confidence bounds applied to Trees), we develop a method called Conditionally Optimistic Exploration (COE). COE augments each agent's state-action value estimate with an action-conditioned optimistic bonus derived from the visitation count of the global state and joint actions of preceding agents. COE is performed during training and disabled at deployment, making it compatible with any value decomposition method for centralized training with decentralized execution. Experiments across various cooperative MARL benchmarks show that COE outperforms current state-of-the-art exploration methods on hard-exploration tasks.

Gabriele Prato, Shagun Sodhani, Alessandro Sordoni et al.

The standard practice for training large language models involves packing multiple documents together to optimize computational efficiency. However, the impact of this process on the models' capabilities remains largely unexplored. To address this gap, we investigate how different document-packing strategies influence the latent multi-hop reasoning abilities of LLMs. Our findings indicate that packing can improve model performance compared to training on individual documents, at the expense of more compute. To further understand the underlying mechanisms, we conduct an ablation study, identifying key factors that explain the advantages of packing. Ultimately, our research deepens the understanding of LLM training dynamics and provides practical insights for optimizing model development.

Ali Rahimi-Kalahroudi, Janarthanan Rajendran, Ida Momennejad et al.

One of the key behavioral characteristics used in neuroscience to determine whether the subject of study -- be it a rodent or a human -- exhibits model-based learning is effective adaptation to local changes in the environment, a particular form of adaptivity that is the focus of this work. In reinforcement learning, however, recent work has shown that modern deep model-based reinforcement-learning (MBRL) methods adapt poorly to local environment changes. An explanation for this mismatch is that MBRL methods are typically designed with sample-efficiency on a single task in mind and the requirements for effective adaptation are substantially higher, both in terms of the learned world model and the planning routine. One particularly challenging requirement is that the learned world model has to be sufficiently accurate throughout relevant parts of the state-space. This is challenging for deep-learning-based world models due to catastrophic forgetting. And while a replay buffer can mitigate the effects of catastrophic forgetting, the traditional first-in-first-out replay buffer precludes effective adaptation due to maintaining stale data. In this work, we show that a conceptually simple variation of this traditional replay buffer is able to overcome this limitation. By removing only samples from the buffer from the local neighbourhood of the newly observed samples, deep world models can be built that maintain their accuracy across the state-space, while also being able to effectively adapt to local changes in the reward function. We demonstrate this by applying our replay-buffer variation to a deep version of the classical Dyna method, as well as to recent methods such as PlaNet and DreamerV2, demonstrating that deep model-based methods can adapt effectively as well to local changes in the environment.

Gabriele Prato, Yale Song, Janarthanan Rajendran et al.

Transformers have become one of the dominant architectures in the field of computer vision. However, there are yet several challenges when applying such architectures to video data. Most notably, these models struggle to model the temporal patterns of video data effectively. Directly targeting this issue, we introduce PatchBlender, a learnable blending function that operates over patch embeddings across the temporal dimension of the latent space. We show that our method is successful at enabling vision transformers to encode the temporal component of video data. On Something-Something v2 and MOVi-A, we show that our method improves the baseline performance of video Transformers. PatchBlender has the advantage of being compatible with almost any Transformer architecture and since it is learnable, the model can adaptively turn on or off the prior. It is also extremely lightweight compute-wise, 0.005% the GFLOPs of a ViT-B.

Hadi Nekoei, Alexandre Blondin Massé, Rachid Hassani et al.

Reinforcement learning (RL) is a powerful framework for optimizing decision-making in complex systems under uncertainty, an essential challenge in real-world settings, particularly in the context of the energy transition. A representative example is remote microgrids that supply power to communities disconnected from the main grid. Enabling the energy transition in such systems requires coordinated control of renewable sources like wind turbines, alongside fuel generators and batteries, to meet demand while minimizing fuel consumption and battery degradation under exogenous and intermittent load and wind conditions. These systems must often conform to extensive regulations and complex operational constraints. To ensure that RL agents respect these constraints, it is crucial to provide interpretable guarantees. In this paper, we introduce Shielded Controller Units (SCUs), a systematic and interpretable approach that leverages prior knowledge of system dynamics to ensure constraint satisfaction. Our shield synthesis methodology, designed for real-world deployment, decomposes the environment into a hierarchical structure where each SCU explicitly manages a subset of constraints. We demonstrate the effectiveness of SCUs on a remote microgrid optimization task with strict operational requirements. The RL agent, equipped with SCUs, achieves a 24% reduction in fuel consumption without increasing battery degradation, outperforming other baselines while satisfying all constraints. We hope SCUs contribute to the safe application of RL to the many decision-making challenges linked to the energy transition.

Maryam Hashemzadeh, Elias Stengel-Eskin, Sarath Chandar et al. · microsoft-research

While Large Language Models (LLMs) have demonstrated significant promise as agents in interactive tasks, their substantial computational requirements and restricted number of calls constrain their practical utility, especially in long-horizon interactive tasks such as decision-making or in scenarios involving continuous ongoing tasks. To address these constraints, we propose a method for transferring the performance of an LLM with billions of parameters to a much smaller language model (770M parameters). Our approach involves constructing a hierarchical agent comprising a planning module, which learns through Knowledge Distillation from an LLM to generate sub-goals, and an execution module, which learns to accomplish these sub-goals using elementary actions. In detail, we leverage an LLM to annotate an oracle path with a sequence of sub-goals towards completing a goal. Subsequently, we utilize this annotated data to fine-tune both the planning and execution modules. Importantly, neither module relies on real-time access to an LLM during inference, significantly reducing the overall cost associated with LLM interactions to a fixed cost. In ScienceWorld, a challenging and multi-task interactive text environment, our method surpasses standard imitation learning based solely on elementary actions by 16.7% (absolute). Our analysis highlights the efficiency of our approach compared to other LLM-based methods. Our code and annotated data for distillation can be found on GitHub.

Arjun V Sudhakar, Hadi Nekoei, Mathieu Reymond et al.

Traditional multi-agent reinforcement learning (MARL) systems can develop cooperative strategies through repeated interactions. However, these systems are unable to perform well on any other setting than the one they have been trained on, and struggle to successfully cooperate with unfamiliar collaborators. This is particularly visible in the Hanabi benchmark, a popular 2-to-5 player cooperative card-game which requires complex reasoning and precise assistance to other agents. Current MARL agents for Hanabi can only learn one specific game-setting (e.g., 2-player games), and play with the same algorithmic agents. This is in stark contrast to humans, who can quickly adjust their strategies to work with unfamiliar partners or situations. In this paper, we introduce Recurrent Replay Relevance Distributed DQN (R3D2), a generalist agent for Hanabi, designed to overcome these limitations. We reformulate the task using text, as language has been shown to improve transfer. We then propose a distributed MARL algorithm that copes with the resulting dynamic observation- and action-space. In doing so, our agent is the first that can play all game settings concurrently, and extend strategies learned from one setting to other ones. As a consequence, our agent also demonstrates the ability to collaborate with different algorithmic agents -- agents that are themselves unable to do so. The implementation code is available at: $\href{https://github.com/chandar-lab/R3D2-A-Generalist-Hanabi-Agent}{R3D2-A-Generalist-Hanabi-Agent}$

Hadi Nekoei, Aman Jaiswal, Patrice Bechard et al.

Large language model (LLM) agents perform well in sequential decision-making tasks, but improving them on unfamiliar domains often requires costly online interactions or fine-tuning on large expert datasets. These strategies are impractical for closed-source models and expensive for open-source ones, with risks of catastrophic forgetting. Offline trajectories offer reusable knowledge, yet demonstration-based methods struggle because raw traces are long, noisy, and tied to specific tasks. We present Just-in-time Episodic Feedback Hinter (JEF Hinter), an agentic system that distills offline traces into compact, context-aware hints. A zooming mechanism highlights decisive steps in long trajectories, capturing both strategies and pitfalls. Unlike prior methods, JEF Hinter leverages both successful and failed trajectories, extracting guidance even when only failure data is available, while supporting parallelized hint generation and benchmark-independent prompting. At inference, a retriever selects relevant hints for the current state, providing targeted guidance with transparency and traceability. Experiments on MiniWoB++, WorkArena-L1, and WebArena-Lite show that JEF Hinter consistently outperforms strong baselines, including human- and document-based hints.

Prashant Govindarajan, Mathieu Reymond, Antoine Clavaud et al.

In silico design and optimization of new materials primarily relies on high-accuracy atomic simulators that perform density functional theory (DFT) calculations. While recent works showcase the strong potential of machine learning to accelerate the material design process, they mostly consist of generative approaches that do not use direct DFT signals as feedback to improve training and generation mainly due to DFT's high computational cost. To aid the adoption of direct DFT signals in the materials design loop through online reinforcement learning (RL), we propose CrystalGym, an open-source RL environment for crystalline material discovery. Using CrystalGym, we benchmark common value- and policy-based reinforcement learning algorithms for designing various crystals conditioned on target properties. Concretely, we optimize for challenging properties like the band gap, bulk modulus, and density, which are directly calculated from DFT in the environment. While none of the algorithms we benchmark solve all CrystalGym tasks, our extensive experiments and ablations show different sample efficiencies and ease of convergence to optimality for different algorithms and environment settings. Additionally, we include a case study on the scope of fine-tuning large language models with reinforcement learning for improving DFT-based rewards. Our goal is for CrystalGym to serve as a test bed for reinforcement learning researchers and material scientists to address these real-world design problems with practical applications. We therefore introduce a novel class of challenges for reinforcement learning methods dealing with time-consuming reward signals, paving the way for future interdisciplinary research for machine learning motivated by real-world applications.

Abdelrahman Zayed, Goncalo Mordido, Ioana Baldini et al.

The widespread use of large language models has brought up essential questions about the potential biases these models might learn. This led to the development of several metrics aimed at evaluating and mitigating these biases. In this paper, we first demonstrate that prompt-based fairness metrics exhibit poor agreement, as measured by correlation, raising important questions about the reliability of fairness assessment using prompts. Then, we outline six relevant reasons why such a low correlation is observed across existing metrics. Based on these insights, we propose a method called Correlated Fairness Output (CAIRO) to enhance the correlation between fairness metrics. CAIRO augments the original prompts of a given fairness metric by using several pre-trained language models and then selects the combination of the augmented prompts that achieves the highest correlation across metrics. We show a significant improvement in Pearson correlation from 0.3 and 0.18 to 0.90 and 0.98 across metrics for gender and religion biases, respectively. Our code is available at https://github.com/chandar-lab/CAIRO.

Steven Y. Feng, Varun Gangal, Jason Wei et al.

Data augmentation has recently seen increased interest in NLP due to more work in low-resource domains, new tasks, and the popularity of large-scale neural networks that require large amounts of training data. Despite this recent upsurge, this area is still relatively underexplored, perhaps due to the challenges posed by the discrete nature of language data. In this paper, we present a comprehensive and unifying survey of data augmentation for NLP by summarizing the literature in a structured manner. We first introduce and motivate data augmentation for NLP, and then discuss major methodologically representative approaches. Next, we highlight techniques that are used for popular NLP applications and tasks. We conclude by outlining current challenges and directions for future research. Overall, our paper aims to clarify the landscape of existing literature in data augmentation for NLP and motivate additional work in this area. We also present a GitHub repository with a paper list that will be continuously updated at https://github.com/styfeng/DataAug4NLP

Hadi Nekoei, Akilesh Badrinaaraayanan, Aaron Courville et al.

Current deep reinforcement learning (RL) algorithms are still highly task-specific and lack the ability to generalize to new environments. Lifelong learning (LLL), however, aims at solving multiple tasks sequentially by efficiently transferring and using knowledge between tasks. Despite a surge of interest in lifelong RL in recent years, the lack of a realistic testbed makes robust evaluation of LLL algorithms difficult. Multi-agent RL (MARL), on the other hand, can be seen as a natural scenario for lifelong RL due to its inherent non-stationarity, since the agents' policies change over time. In this work, we introduce a multi-agent lifelong learning testbed that supports both zero-shot and few-shot settings. Our setup is based on Hanabi -- a partially-observable, fully cooperative multi-agent game that has been shown to be challenging for zero-shot coordination. Its large strategy space makes it a desirable environment for lifelong RL tasks. We evaluate several recent MARL methods, and benchmark state-of-the-art LLL algorithms in limited memory and computation regimes to shed light on their strengths and weaknesses. This continual learning paradigm also provides us with a pragmatic way of going beyond centralized training which is the most commonly used training protocol in MARL. We empirically show that the agents trained in our setup are able to coordinate well with unseen agents, without any additional assumptions made by previous works. The code and all pre-trained models are available at https://github.com/chandar-lab/Lifelong-Hanabi.

Nolan Bard, Jakob N. Foerster, Sarath Chandar et al.

From the early days of computing, games have been important testbeds for studying how well machines can do sophisticated decision making. In recent years, machine learning has made dramatic advances with artificial agents reaching superhuman performance in challenge domains like Go, Atari, and some variants of poker. As with their predecessors of chess, checkers, and backgammon, these game domains have driven research by providing sophisticated yet well-defined challenges for artificial intelligence practitioners. We continue this tradition by proposing the game of Hanabi as a new challenge domain with novel problems that arise from its combination of purely cooperative gameplay with two to five players and imperfect information. In particular, we argue that Hanabi elevates reasoning about the beliefs and intentions of other agents to the foreground. We believe developing novel techniques for such theory of mind reasoning will not only be crucial for success in Hanabi, but also in broader collaborative efforts, especially those with human partners. To facilitate future research, we introduce the open-source Hanabi Learning Environment, propose an experimental framework for the research community to evaluate algorithmic advances, and assess the performance of current state-of-the-art techniques.

Mido Assran, Adrien Bardes, David Fan et al. · meta-ai

A major challenge for modern AI is to learn to understand the world and learn to act largely by observation. This paper explores a self-supervised approach that combines internet-scale video data with a small amount of interaction data (robot trajectories), to develop models capable of understanding, predicting, and planning in the physical world. We first pre-train an action-free joint-embedding-predictive architecture, V-JEPA 2, on a video and image dataset comprising over 1 million hours of internet video. V-JEPA 2 achieves strong performance on motion understanding (77.3 top-1 accuracy on Something-Something v2) and state-of-the-art performance on human action anticipation (39.7 recall-at-5 on Epic-Kitchens-100) surpassing previous task-specific models. Additionally, after aligning V-JEPA 2 with a large language model, we demonstrate state-of-the-art performance on multiple video question-answering tasks at the 8 billion parameter scale (e.g., 84.0 on PerceptionTest, 76.9 on TempCompass). Finally, we show how self-supervised learning can be applied to robotic planning tasks by post-training a latent action-conditioned world model, V-JEPA 2-AC, using less than 62 hours of unlabeled robot videos from the Droid dataset. We deploy V-JEPA 2-AC zero-shot on Franka arms in two different labs and enable picking and placing of objects using planning with image goals. Notably, this is achieved without collecting any data from the robots in these environments, and without any task-specific training or reward. This work demonstrates how self-supervised learning from web-scale data and a small amount of robot interaction data can yield a world model capable of planning in the physical world.

Andreas Madsen, Sarath Chandar, Siva Reddy · mila

Instruction-tuned Large Language Models (LLMs) excel at many tasks and will even explain their reasoning, so-called self-explanations. However, convincing and wrong self-explanations can lead to unsupported confidence in LLMs, thus increasing risk. Therefore, it's important to measure if self-explanations truly reflect the model's behavior. Such a measure is called interpretability-faithfulness and is challenging to perform since the ground truth is inaccessible, and many LLMs only have an inference API. To address this, we propose employing self-consistency checks to measure faithfulness. For example, if an LLM says a set of words is important for making a prediction, then it should not be able to make its prediction without these words. While self-consistency checks are a common approach to faithfulness, they have not previously been successfully applied to LLM self-explanations for counterfactual, feature attribution, and redaction explanations. Our results demonstrate that faithfulness is explanation, model, and task-dependent, showing self-explanations should not be trusted in general. For example, with sentiment classification, counterfactuals are more faithful for Llama2, feature attribution for Mistral, and redaction for Falcon 40B.

Mohammad Reza Samsami, Artem Zholus, Janarthanan Rajendran et al.

Current model-based reinforcement learning (MBRL) agents struggle with long-term dependencies. This limits their ability to effectively solve tasks involving extended time gaps between actions and outcomes, or tasks demanding the recalling of distant observations to inform current actions. To improve temporal coherence, we integrate a new family of state space models (SSMs) in world models of MBRL agents to present a new method, Recall to Imagine (R2I). This integration aims to enhance both long-term memory and long-horizon credit assignment. Through a diverse set of illustrative tasks, we systematically demonstrate that R2I not only establishes a new state-of-the-art for challenging memory and credit assignment RL tasks, such as BSuite and POPGym, but also showcases superhuman performance in the complex memory domain of Memory Maze. At the same time, it upholds comparable performance in classic RL tasks, such as Atari and DMC, suggesting the generality of our method. We also show that R2I is faster than the state-of-the-art MBRL method, DreamerV3, resulting in faster wall-time convergence.

Abdelrahman Zayed, Goncalo Mordido, Samira Shabanian et al.

The increasing size of large language models (LLMs) has introduced challenges in their training and inference. Removing model components is perceived as a solution to tackle the large model sizes, however, existing pruning methods solely focus on performance, without considering an essential aspect for the responsible use of LLMs: model fairness. It is crucial to address the fairness of LLMs towards diverse groups, such as women, Black people, LGBTQ+, Jewish communities, among others, as they are being deployed and available to a wide audience. In this work, first, we investigate how attention heads impact fairness and performance in pre-trained transformer-based language models. We then propose a novel method to prune the attention heads that negatively impact fairness while retaining the heads critical for performance, i.e. language modeling capabilities. Our approach is practical in terms of time and resources, as it does not require fine-tuning the final pruned, and fairer, model. Our findings demonstrate a reduction in gender bias by 19%, 19.5%, 39.5%, 34.7%, 23%, and 8% for DistilGPT-2, GPT-2, GPT-Neo of two different sizes, GPT-J, and Llama 2 models, respectively, in comparison to the biased model, with only a slight decrease in performance.

Reza Bayat, Ali Rahimi-Kalahroudi, Mohammad Pezeshki et al.

A key challenge in AI alignment is guiding large language models (LLMs) to follow desired behaviors at test time. Activation steering, which modifies internal model activations during inference, offers a potential solution. However, prior work in dense activation spaces struggles with superposition, wherein multiple features become entangled, limiting interpretability and precise control. In contrast, sparse representations provide an untapped opportunity for more interpretable behavior modulation. In this work, we introduce sparse activation steering (SAS), a method that leverages sparse autoencoders (SAEs) to steer LLM behavior in sparse spaces. By isolating behavior-specific features through a contrastive prompt-pairing approach, we define a set of features that can selectively reinforce or suppress behaviors. Experiments on Gemma 2 LLMs show that SAS vectors enable nuanced behavioral modulation and finer-grained control. Furthermore, scaling SAEs improves monosemanticity of SAS vectors, suggesting more reliable and interpretable interventions.

Jerry Huang, Prasanna Parthasarathi, Mehdi Rezagholizadeh et al. · mila

Large language models (LLMs) show an innate skill for solving language based tasks. But insights have suggested an inability to adjust for information or task-solving skills becoming outdated, as their knowledge, stored directly within their parameters, remains static in time. Tool use helps by offloading work to systems that the LLM can access through an interface, but LLMs that use them still must adapt to nonstationary environments for prolonged use, as new tools can emerge and existing tools can change. Nevertheless, tools require less specialized knowledge, therefore we hypothesize they are better suited for continual learning (CL) as they rely less on parametric memory for solving tasks and instead focus on learning when to apply pre-defined tools. To verify this, we develop a synthetic benchmark and follow this by aggregating existing NLP tasks to form a more realistic testing scenario. While we demonstrate scaling model size is not a solution, regardless of tool usage, continual learning techniques can enable tool LLMs to both adapt faster while forgetting less, highlighting their potential as continual learners.

Megh Thakkar, Quentin Fournier, Matthew Riemer et al. · ibm-research

There is a growing interest in training domain-expert LLMs that excel in specific technical fields compared to their general-purpose instruction-tuned counterparts. However, these expert models often experience a loss in their safety abilities in the process, making them capable of generating harmful content. As a solution, we introduce an efficient and effective merging-based alignment method called \textsc{MergeAlign} that interpolates the domain and alignment vectors, creating safer domain-specific models while preserving their utility. We apply \textsc{MergeAlign} on Llama3 variants that are experts in medicine and finance, obtaining substantial alignment improvements with minimal to no degradation on domain-specific benchmarks. We study the impact of model merging through model similarity metrics and contributions of individual models being merged. We hope our findings open new research avenues and inspire more efficient development of safe expert LLMs.

Razvan Pascanu, Clare Lyle, Ionut-Vlad Modoranu et al. · deepmind

Due to the nonlinear nature of Deep Neural Networks (DNNs), one can not guarantee convergence to a unique global minimum of the loss when using optimizers relying only on local information, such as SGD. Indeed, this was a primary source of skepticism regarding the feasibility of DNNs in the early days of the field. The past decades of progress in deep learning have revealed this skepticism to be misplaced, and a large body of empirical evidence shows that sufficiently large DNNs following standard training protocols exhibit well-behaved optimization dynamics that converge to performant solutions. This success has biased the community to use convex optimization as a mental model for learning, leading to a focus on training efficiency, either in terms of required iteration, FLOPs or wall-clock time, when improving optimizers. We argue that, while this perspective has proven extremely fruitful, another perspective specific to DNNs has received considerably less attention: the optimizer not only influences the rate of convergence, but also the qualitative properties of the learned solutions. Restated, the optimizer can and will encode inductive biases and change the effective expressivity of a given class of models. Furthermore, we believe the optimizer can be an effective way of encoding desiderata in the learning process. We contend that the community should aim at understanding the biases of already existing methods, as well as aim to build new optimizers with the explicit intent of inducing certain properties of the solution, rather than solely judging them based on their convergence rates. We hope our arguments will inspire research to improve our understanding of how the learning process can impact the type of solution we converge to, and lead to a greater recognition of optimizers design as a critical lever that complements the roles of architecture and data in shaping model outcomes.

Lola Le Breton, Quentin Fournier, Mariam El Mezouar et al.

Recent innovations in architecture, pre-training, and fine-tuning have led to the remarkable in-context learning and reasoning abilities of large auto-regressive language models such as LLaMA and DeepSeek. In contrast, encoders like BERT and RoBERTa have not seen the same level of progress despite being foundational for many downstream NLP applications. To bridge this gap, we introduce NeoBERT, a next-generation encoder that redefines the capabilities of bidirectional models by integrating state-of-the-art advancements in architecture, modern data, and optimized pre-training methodologies. NeoBERT is designed for seamless adoption: it serves as a plug-and-play replacement for existing base models, relies on an optimal depth-to-width ratio, and leverages an extended context length of 4,096 tokens. Despite its compact 250M parameter footprint, it achieves state-of-the-art results on the massive MTEB benchmark, outperforming BERT large, RoBERTa large, NomicBERT, and ModernBERT under identical fine-tuning conditions. In addition, we rigorously evaluate the impact of each modification on GLUE and design a uniform fine-tuning and evaluation framework for MTEB. We release all code, data, checkpoints, and training scripts to accelerate research and real-world adoption.

Andreas Madsen, Himabindu Lakkaraju, Siva Reddy et al. · mila

Interpretability is the study of explaining models in understandable terms to humans. At present, interpretability is divided into two paradigms: the intrinsic paradigm, which believes that only models designed to be explained can be explained, and the post-hoc paradigm, which believes that black-box models can be explained. At the core of this debate is how each paradigm ensures its explanations are faithful, i.e., true to the model's behavior. This is important, as false but convincing explanations lead to unsupported confidence in artificial intelligence (AI), which can be dangerous. This paper's position is that we should think about new paradigms while staying vigilant regarding faithfulness. First, by examining the history of paradigms in science, we see that paradigms are constantly evolving. Then, by examining the current paradigms, we can understand their underlying beliefs, the value they bring, and their limitations. Finally, this paper presents 3 emerging paradigms for interpretability. The first paradigm designs models such that faithfulness can be easily measured. Another optimizes models such that explanations become faithful. The last paradigm proposes to develop models that produce both a prediction and an explanation.

Jerry Huang, Prasanna Parthasarathi, Mehdi Rezagholizadeh et al. · mila

The growth in prominence of large language models (LLMs) in everyday life can be largely attributed to their generative abilities, yet some of this is also owed to the risks and costs associated with their use. On one front is their tendency to hallucinate false or misleading information, limiting their reliability. On another is the increasing focus on the computational limitations associated with traditional self-attention based LLMs, which has brought about new alternatives, in particular recurrent models, meant to overcome them. Yet it remains uncommon to consider these two concerns simultaneously. Do changes in architecture exacerbate/alleviate existing concerns about hallucinations? Do they affect how and where they occur? Through an extensive evaluation, we study how these architecture-based inductive biases affect the propensity to hallucinate. While hallucination remains a general phenomenon not limited to specific architectures, the situations in which they occur and the ease with which specific types of hallucinations can be induced can significantly differ based on the model architecture. These findings highlight the need for better understanding both these problems in conjunction with each other, as well as consider how to design more universal techniques for handling hallucinations.

Mohammad Reza Samsami, Mats Leon Richter, Juan Rodriguez et al. · mila

Large language models must balance their weight-encoded knowledge with in-context information from prompts to generate accurate responses. This paper investigates this interplay by analyzing how models of varying capacities within the same family handle intentionally misleading in-context information. Our experiments demonstrate that larger models exhibit higher resilience to deceptive prompts, showcasing an advanced ability to interpret and integrate prompt information with their internal knowledge. Furthermore, we find that larger models outperform smaller ones in following legitimate instructions, indicating that their resilience is not due to disregarding in-context information. We also show that this phenomenon is likely not a result of memorization but stems from the models' ability to better leverage implicit task-relevant information from the prompt alongside their internally stored knowledge.

Simon Guiroy, Mats Richter, Sarath Chandar et al.

To create state-of-the-art models for many downstream tasks, it has become common practice to fine-tune a pre-trained large vision model. However, it remains an open question of how to best determine which of the many possible model checkpoints resulting from a large training run to use as the starting point. This becomes especially important when data for the target task of interest is scarce, unlabeled and out-of-distribution. In such scenarios, common methods relying on in-distribution validation data become unreliable or inapplicable. This work proposes a novel approach for model selection that operates reliably on just a few unlabeled examples from the target task. Our approach is based on a novel concept: Neural Coherence, which entails characterizing a model's activation statistics for source and target domains, allowing one to define model selection methods with high data-efficiency. We provide experiments where models are pre-trained on ImageNet1K and examine target domains consisting of Food-101, PlantNet-300K and iNaturalist. We also evaluate it in many meta-learning settings. Our approach significantly improves generalization across these different target domains compared to established baselines. We further demonstrate the versatility of Neural Coherence as a powerful principle by showing its effectiveness in training data selection.

Milad Aghajohari, Kamran Chitsaz, Amirhossein Kazemnejad et al. · mila

Reinforcement learning (RL) has recently become a strong recipe for training reasoning LLMs that produce long chains of thought (LongCoT). Yet the standard RL "thinking environment", where the state is the prompt plus all prior reasoning tokens, makes the state unbounded and forces attention-based policies to pay quadratic compute as thoughts lengthen. We revisit the environment itself. We propose Markovian Thinking, a paradigm in which the policy advances reasoning while conditioning on a constant-size state, decoupling thinking length from context size. As an immediate consequence this yields linear compute with constant memory. We instantiate this idea with Delethink, an RL environment that structures reasoning into fixed-size chunks. Within each chunk, the model thinks as usual; at the boundary, the environment resets the context and reinitializes the prompt with a short carryover. Through RL, the policy learns to write a textual state near the end of each chunk sufficient for seamless continuation of reasoning after reset. Trained in this environment, an R1-Distill 1.5B model reasons in 8K-token chunks yet thinks up to 24K tokens, matching or surpassing LongCoT-RL trained with a 24K budget. With test-time scaling, Delethink continues to improve where LongCoT plateaus. The effect of linear compute is substantial: we empirically estimate at 96K average thinking length LongCoT-RL costs 27 H100-months vs. 7 for Delethink. Analysis at RL initialization shows off-the-shelf reasoning models (1.5B-120B) often sample Markovian traces zero-shot across diverse benchmarks, providing positive samples that make RL effective at scale. Our results show that redesigning the thinking environment is a powerful lever: it enables very long reasoning without quadratic overhead and opens a path toward efficient, scalable reasoning LLMs.

Prasanna Parthasarathi, Mathieu Reymond, Boxing Chen et al. · mila

Large language models (LLMs) are increasingly deployed for tasks requiring complex reasoning, prompting significant interest in improving their reasoning abilities through post-training. Especially RL based methods using verifiable reward, like the state-of-the-art GRPO, have shown to tremendously improve reasoning behaviors when applied as post-training methods. However, the lack of an explicit reward or critic model limits GRPO's ability to assign fine-grained credit across token sequences. In this work, we present GRPO-$λ$, a novel extension to GRPO that enhances credit assignment in RL finetuning of LLMs for complex reasoning tasks. We approximate learning from $λ$-return with a reformulation of eligibility traces using token-level log-probabilities applied after each sequence generation, and a novel critic-free approximation of the temporal-difference error. We introduce a few variations for the weighting of the $λ$-return, and their applications to the eligibility-trace, where all the variations provide significant gains over GRPO. We compare GRPO-$λ$ against GRPO by training models from 1.5B to 7B parameters on $4$ different math reasoning datasets. The training plots demonstrate 30-40% improved performance during RL training on both LLaMA-3.1 and Qwen-2.5 architectures. Finally, we show that with GRPO-$λ$, the resulting average performance on AIME24, Math500, OlympiadMath, MinervaMath, and AMC improves over GRPO by over $3$ points and a $4.5$ points improvement on the 7B model.

Kamran Chitsaz, Roshan Balaji, Quentin Fournier et al.

Designing de-novo molecules with desired property profiles requires efficient exploration of the vast chemical space ranging from $10^{23}$ to $10^{60}$ possible synthesizable candidates. While various deep generative models have been developed to design small molecules using diverse input representations, Molecular Large Language Models (Mol-LLMs) based on string representations have emerged as a scalable approach capable of exploring billions of molecules. However, there remains limited understanding regarding how standard language modeling practices such as textual representations, tokenization strategies, model size, and dataset scale impact molecular generation performance. In this work, we systematically investigate these critical aspects by introducing NovoMolGen, a family of transformer-based foundation models pretrained on 1.5 billion molecules for de-novo molecule generation. Through extensive empirical analyses, we identify a weak correlation between performance metrics measured during pretraining and actual downstream performance, revealing important distinctions between molecular and general NLP training dynamics. NovoMolGen establishes new state-of-the-art results, substantially outperforming prior Mol-LLMs and specialized generative models in both unconstrained and goal-directed molecular generation tasks, thus providing a robust foundation for advancing efficient and effective molecular modeling strategies.

Istabrak Abbes, Gopeshh Subbaraj, Matthew Riemer et al.

Training large language models (LLMs) typically involves pre-training on massive corpora, only to restart the process entirely when new data becomes available. A more efficient and resource-conserving approach would be continual pre-training, where models are updated with new data rather than retraining from scratch. However, the introduction of new data often causes distribution shifts, leading to performance degradation on previously learned tasks. In this paper, we take a deeper look at two popular proposals for addressing this distribution shift within the continual learning literature: experience replay and gradient alignment. We consider continual pre-training of models within the Llama family of architectures at a large scale across languages with 100 billion tokens of training data in each language, finding that both replay and gradient alignment lead to more stable learning without forgetting. This conclusion holds both as we vary the model scale and as we vary the number and diversity of tasks. Moreover, we are the first to demonstrate the effectiveness of gradient alignment techniques in the context of LLM pre-training and propose an efficient implementation of meta-experience replay (MER) that imbues experience replay with the benefits of gradient alignment despite negligible compute and memory overhead. Our scaling analysis across model sizes and replay rates indicates that small rates of replaying old examples are definitely a more valuable use of compute than investing in model size, but that it is more compute efficient to scale the size of the model than invest in high rates of replaying old examples.

Milan Bhan, Jean-Noel Vittaut, Nicolas Chesneau et al.

Large Language Models (LLMs) can generate plausible free text self-explanations to justify their answers. However, these natural language explanations may not accurately reflect the model's actual reasoning process, indicating a lack of faithfulness. Existing faithfulness evaluation methods rely primarily on behavioral tests or computational block analysis without examining the semantic content of internal neural representations. This paper proposes NeuroFaith, a flexible framework that measures the faithfulness of LLM free text self-explanation by identifying key concepts within explanations and mechanistically testing whether these concepts actually influence the model's predictions. We show the versatility of NeuroFaith across 2-hop reasoning and classification tasks. Additionally, a linear faithfulness probe based on NeuroFaith is developed to detect unfaithful self-explanations from representation space and improve faithfulness through steering. NeuroFaith provides a principled approach to evaluating and enhancing the faithfulness of LLM free text self-explanations, addressing critical needs for trustworthy AI systems.

Pranshu Malviya, Goncalo Mordido, Aristide Baratin et al.

Efficiently exploring complex loss landscapes is key to the performance of deep neural networks. While momentum-based optimizers are widely used in state-of-the-art setups, classical momentum can still struggle with large, misaligned gradients, leading to oscillations. To address this, we propose Torque-Aware Momentum (TAM), which introduces a damping factor based on the angle between the new gradients and previous momentum, stabilizing the update direction during training. Empirical results show that TAM, which can be combined with both SGD and Adam, enhances exploration, handles distribution shifts more effectively, and improves generalization performance across various tasks, including image classification and large language model fine-tuning, when compared to classical momentum-based optimizers.

Pranshu Malviya, Jerry Huang, Aristide Baratin et al.

Determining the optimal model for a given task often requires training multiple models from scratch, which becomes impractical as dataset and model sizes grow. A more efficient alternative is to expand smaller pre-trained models, but this approach is underutilized due to a limited understanding of its impact on the training dynamics. Existing methods for quantifying this impact have notable limitations, including computation cost. To address this, we introduce a new perspective based on the loss landscape, which has been shown to contain a manifold of linearly connected minima. Specifically, we propose a metric that estimates the size of this manifold to study the impact of model expansion. Our experiments reveal a strong correlation between performance gains and our manifold metric, enabling more informed model comparison and offering a first step toward a geometry-driven approach for reliable model expansion. Notably, our metric outperforms other baselines, even when different types of expansion with equivalent number of parameters are applied to a model.

Megh Thakkar, Quentin Fournier, Matthew D Riemer et al.

Large language models are first pre-trained on trillions of tokens and then instruction-tuned or aligned to specific preferences. While pre-training remains out of reach for most researchers due to the compute required, fine-tuning has become affordable thanks to parameter-efficient methods such as LoRA and QLoRA. Alignment is known to be sensitive to the many factors involved, including the quantity and quality of data, the alignment method, and the adapter rank. However, there has not yet been an extensive study of their effect on downstream performance. To address this gap, we conduct an in-depth investigation of the impact of popular choices for three crucial axes: (i) the alignment dataset (HH-RLHF and BeaverTails), (ii) the alignment technique (SFT and DPO), and (iii) the model (LLaMA-1, Vicuna-v1.3, Mistral-7b, and Mistral-7b-Instruct). Our extensive setup spanning over 300 experiments reveals consistent trends and unexpected findings. We observe how more informative data helps with preference alignment, cases where supervised fine-tuning outperforms preference optimization, and how aligning to a distinct preference boosts performance on downstream tasks. Through our in-depth analyses, we put forward key guidelines to help researchers perform more effective parameter-efficient LLM alignment.