Alessandro Sordoni

Alessandro Sordoni, Xingdi Yuan, Marc-Alexandre Côté et al. · microsoft-research

Large language models (LLMs) can be seen as atomic units of computation mapping sequences to a distribution over sequences. Thus, they can be seen as stochastic language layers in a language network, where the learnable parameters are the natural language prompts at each layer. By stacking two such layers and feeding the output of one layer to the next, we obtain a Deep Language Network (DLN). We first show how to effectively perform prompt optimization for a 1-Layer language network (DLN-1). Then, we present an extension that applies to 2-layer DLNs (DLN-2), where two prompts must be learned. The key idea is to consider the output of the first layer as a latent variable, which requires inference, and prompts to be learned as the parameters of the generative distribution. We first test the effectiveness of DLN-1 in multiple reasoning and natural language understanding tasks. Then, we show that DLN-2 can reach higher performance than a single layer, showing promise that we might reach comparable performance to GPT-4, even when each LLM in the network is smaller and less powerful.

Prateek Yadav, Colin Raffel, Mohammed Muqeeth et al. · ibm-research

The availability of performant pre-trained models has led to a proliferation of fine-tuned expert models that are specialized to a particular domain or task. Model MoErging methods aim to recycle expert models to create an aggregate system with improved performance or generalization. A key component of MoErging methods is the creation of a router that decides which expert model(s) to use for a particular input or application. The promise, effectiveness, and large design space of MoErging has spurred the development of many new methods over the past few years. This rapid pace of development has made it challenging to compare different MoErging methods, which are rarely compared to one another and are often validated in different experimental setups. To remedy such gaps, we present a comprehensive survey of MoErging methods that includes a novel taxonomy for cataloging key design choices and clarifying suitable applications for each method. Apart from surveying MoErging research, we inventory software tools and applications that make use of MoErging. We additionally discuss related fields of study such as model merging, multitask learning, and mixture-of-experts models. Taken as a whole, our survey provides a unified overview of existing MoErging methods and creates a solid foundation for future work in this burgeoning field.

Arian Hosseini, Ankit Vani, Dzmitry Bahdanau et al. · mila

Pretrained large generative language models have shown great performance on many tasks, but exhibit low compositional generalization abilities. Scaling such models has been shown to improve their performance on various NLP tasks even just by conditioning them on a few examples to solve the task without any fine-tuning (also known as in-context learning). In this work, we look at the gap between the in-distribution (ID) and out-of-distribution (OOD) performance of such models in semantic parsing tasks with in-context learning. In the ID settings, the demonstrations are from the same split (test or train) that the model is being evaluated on, and in the OOD settings, they are from the other split. We look at how the relative generalization gap of in-context learning evolves as models are scaled up. We evaluate four model families, OPT, BLOOM, CodeGen and Codex on three semantic parsing datasets, CFQ, SCAN and GeoQuery with different number of exemplars, and observe a trend of decreasing relative generalization gap as models are scaled up.

He Bai, Tong Wang, Alessandro Sordoni et al. · apple-ml

Class-based language models (LMs) have been long devised to address context sparsity in $n$-gram LMs. In this study, we revisit this approach in the context of neural LMs. We hypothesize that class-based prediction leads to an implicit context aggregation for similar words and thus can improve generalization for rare words. We map words that have a common WordNet hypernym to the same class and train large neural LMs by gradually annealing from predicting the class to token prediction during training. Empirically, this curriculum learning strategy consistently improves perplexity over various large, highly-performant state-of-the-art Transformer-based models on two datasets, WikiText-103 and Arxiv. Our analysis shows that the performance improvement is achieved without sacrificing performance on rare words. Finally, we document other attempts that failed to yield empirical gains, and discuss future directions for the adoption of class-based LMs on a larger scale.

Xinyi Wang, Lucas Caccia, Oleksiy Ostapenko et al.

Large language models (LLMs) have recently attracted considerable interest for their ability to perform complex reasoning tasks, such as chain-of-thought (CoT) reasoning. However, most of the existing approaches to enhance this ability rely heavily on data-driven methods, while neglecting the structural aspects of the model's reasoning capacity. To encourage a more structural generation of CoT steps, we propose a hierarchical generation scheme: we let the LM generate a planning token at the start of each reasoning step, intuitively serving as a high-level plan of the current step, and add their embeddings to the model parameters. Our approach requires a negligible increase in trainable parameters (0.001%) and can be applied through either full fine-tuning or a more parameter-efficient scheme. We demonstrate our method's effectiveness by applying it to three different LLMs, showing notable accuracy improvements across three math word problem datasets and one multihop QA dataset with respect to standard fine-tuning baselines.

Lucas Caccia, Edoardo Ponti, Zhan Su et al.

Parameter-efficient fine-tuning (PEFT) for cross-task generalization consists in pre-training adapters on a multi-task training set before few-shot adaptation to test tasks. Polytropon [Ponti et al., 2023] ($\texttt{Poly}$) jointly learns an inventory of adapters and a routing function that selects a (variable-size) subset of adapters for each task during both pre-training and few-shot adaptation. In this paper, we investigate the role that adapter routing plays in its success and design new variants based on our findings. First, we build on the intuition that finer-grained routing provides more expressivity. Hence, we propose $\texttt{MHR}$ (Multi-Head Routing) which combines subsets of adapter parameters and outperforms $\texttt{Poly}$ under a comparable parameter budget; by only fine-tuning the routing function and not the adapters ($\texttt{MHR}$-$z$) we achieve competitive performance with extreme parameter efficiency. Second, we find that $\texttt{Poly}$/$\texttt{MHR}$ performance is a result of better multi-task optimization, rather than modular inductive biases that facilitate adapter recombination and local adaptation, as previously hypothesized. In fact, we find that $\texttt{MHR}$ exhibits high gradient alignment between training tasks. We find that routing is most beneficial during multi-task pre-training rather than during few-shot adaptation and propose $\texttt{MHR}$-$μ$, which discards routing and fine-tunes the average of the pre-trained adapters on each downstream tasks. This establishes $\texttt{MHR}$-$μ$ as an effective method for single-adapter fine-tuning. We also show that $\texttt{MHR}$-$μ$ can be used as an effective zero-shot transfer method by training the average of the pre-trained adapters for a few additional steps on the multi-task training set: this yields gains up to 3% on absolute accuracy w.r.t. the baselines.

Benjamin LeBrun, Alessandro Sordoni, Timothy J. O'Donnell

A fundamental characteristic of natural language is the high rate at which speakers produce novel expressions. Because of this novelty, a heavy-tail of rare events accounts for a significant amount of the total probability mass of distributions in language (Baayen, 2001). Standard language modeling metrics such as perplexity quantify the performance of language models (LM) in aggregate. As a result, we have relatively little understanding of whether neural LMs accurately estimate the probability of sequences in this heavy-tail of rare events. To address this gap, we develop a controlled evaluation scheme which uses generative models trained on natural data as artificial languages from which we can exactly compute sequence probabilities. Training LMs on generations from these artificial languages, we compare the sequence-level probability estimates given by LMs to the true probabilities in the target language. Our experiments reveal that LSTM and Transformer language models (i) systematically underestimate the probability of sequences drawn from the target language, and (ii) do so more severely for less-probable sequences. Investigating where this probability mass went, (iii) we find that LMs tend to overestimate the probability of ill formed (perturbed) sequences. In addition, we find that this underestimation behaviour (iv) is weakened, but not eliminated by greater amounts of training data, and (v) is exacerbated for target distributions with lower entropy.

Silviu Pitis, Ziang Xiao, Nicolas Le Roux et al.

While finetuning language models from pairwise preferences has proven remarkably effective, the underspecified nature of natural language presents critical challenges. Direct preference feedback is uninterpretable, difficult to provide where multidimensional criteria may apply, and often inconsistent, either because it is based on incomplete instructions or provided by diverse principals. To address these challenges, we consider the two-step preference modeling procedure that first resolves the under-specification by selecting a context, and then evaluates preference with respect to the chosen context. We decompose reward modeling error according to these two steps, which suggests that supervising context in addition to context-specific preference may be a viable approach to aligning models with diverse human preferences. For this to work, the ability of models to evaluate context-specific preference is critical. To this end, we contribute context-conditioned preference datasets and accompanying experiments that investigate the ability of language models to evaluate context-specific preference. We use our datasets to (1) show that existing preference models benefit from, but fail to fully consider, added context, (2) finetune a context-aware reward model with context-specific performance exceeding that of GPT-4 and Llama 3 70B on tested datasets, and (3) investigate the value of context-aware preference modeling.

Minseon Kim, Lucas Caccia, Zhengyan Shi et al.

User prompts to large language models (LLMs) are often ambiguous or under-specified, and subtle contextual cues shaped by user intentions, prior knowledge, and risk factors strongly influence what constitutes an appropriate response. Misinterpreting intent or risks may lead to unsafe outputs, while overly cautious interpretations can cause unnecessary refusal of benign requests. In this paper, we question the conventional framework in which LLMs generate immediate responses to requests without considering broader contextual factors. User requests are situated within broader contexts such as intentions, knowledge, and prior experience, which strongly influence what constitutes an appropriate answer. We propose a framework that extracts and leverages such contextual information from the user prompt itself. Specifically, a reinforcement learning based context generator, designed in an autoencoder-like fashion, is trained to infer contextual signals grounded in the prompt and use them to guide response generation. This approach is particularly important for safety tasks, where ambiguous requests may bypass safeguards while benign but confusing requests can trigger unnecessary refusals. Experiments show that our method reduces harmful responses by an average of 5.6% on the SafetyInstruct dataset across multiple foundation models and improves the harmonic mean of attack success rate and compliance on benign prompts by 6.2% on XSTest and WildJailbreak. These results demonstrate the effectiveness of context extraction for safer and more reliable LLM inferences.

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.

Sanae Lotfi, Lucas Caccia, Alessandro Sordoni et al.

While large language models (LLMs) fine-tuned with lightweight adapters achieve strong performance across diverse tasks, their performance on individual tasks depends on the fine-tuning strategy. Fusing independently trained models with different strengths has shown promise for multi-task learning through three main strategies: ensembling, which combines outputs from independent models; merging, which fuses model weights via parameter averaging; and routing, which integrates models in an input-dependent fashion. However, many design decisions in these approaches remain understudied, and the relative benefits of more sophisticated ensembling, merging and routing techniques are not fully understood. We empirically evaluate their trade-offs, addressing two key questions: What are the advantages of going beyond uniform ensembling or merging? And does the flexibility of routing justify its complexity? Our findings indicate that non-uniform ensembling and merging improve performance, but routing offers even greater gains. To mitigate the computational cost of routing, we analyze expert selection techniques, showing that clustering and greedy subset selection can maintain reasonable performance with minimal overhead. These insights advance our understanding of model fusion for multi-task learning.

Wanru Zhao, Lucas Caccia, Zhengyan Shi et al.

Curriculum learning is a class of training strategies that organizes the data being exposed to a model by difficulty, gradually from simpler to more complex examples. This research explores a reverse curriculum generation approach that recursively decomposes complex datasets into simpler, more learnable components. We propose a teacher-student framework where the teacher is equipped with the ability to reason step-by-step, which is used to recursively generate easier versions of examples, enabling the student model to progressively master difficult tasks. We propose a novel scoring system to measure data difficulty based on its structural complexity and conceptual depth, allowing curriculum construction over decomposed data. Experiments on math datasets (MATH and AIME) and code generation datasets demonstrate that models trained with curricula generated by our approach exhibit superior performance compared to standard training on original datasets.

Hyunji Lee, Minseon Kim, Chinmay Singh et al.

As coding agents are increasingly deployed in large codebases, the need to automatically design challenging, codebase-level evaluation is central. We propose Gistify, a task where a coding LLM must create a single, minimal, self-contained file that can reproduce a specific functionality of a codebase. The coding LLM is given full access to a codebase along with a specific entrypoint (e.g., a python command), and the generated file must replicate the output of the same command ran under the full codebase, while containing only the essential components necessary to execute the provided command. Success on Gistify requires both structural understanding of the codebase, accurate modeling of its execution flow as well as the ability to produce potentially large code patches. Our findings show that current state-of-the-art models struggle to reliably solve Gistify tasks, especially ones with long executions traces.

Yuchen Lu, Zhen Liu, Aristide Baratin et al.

We address the problem of evaluating the quality of self-supervised learning (SSL) models without access to supervised labels, while being agnostic to the architecture, learning algorithm or data manipulation used during training. We argue that representations can be evaluated through the lens of expressiveness and learnability. We propose to use the Intrinsic Dimension (ID) to assess expressiveness and introduce Cluster Learnability (CL) to assess learnability. CL is measured in terms of the performance of a KNN classifier trained to predict labels obtained by clustering the representations with K-means. We thus combine CL and ID into a single predictor -- CLID. Through a large-scale empirical study with a diverse family of SSL algorithms, we find that CLID better correlates with in-distribution model performance than other competing recent evaluation schemes. We also benchmark CLID on out-of-domain generalization, where CLID serves as a predictor of the transfer performance of SSL models on several visual classification tasks, yielding improvements with respect to the competing baselines.

Weijia Xu, Alessandro Sordoni, Chandan Singh et al.

We introduce EvoLib, a test-time learning framework that enables large language models to accumulate, reuse, and evolve knowledge across problem instances without parameter updates or external supervision. Instead of adapting model parameters, our approach maintains a shared library of knowledge abstractions, including modular skills and reflective insights, automatically extracted from the model's own inference trajectories. To support continual improvement, we introduce a principled weighting and consolidation mechanism that jointly optimizes for immediate utility and long-term value. This allows simple, instance-specific abstractions to evolve into more general and reusable ones over time. Across challenging benchmarks in mathematical reasoning, code generation, and multi-turn agentic environments, EvoLib improves substantially over the top test-time scaling and learning methods without ground-truth feedback.

Arian Hosseini, Xingdi Yuan, Nikolay Malkin et al. · deepmind

Common self-improvement approaches for large language models (LLMs), such as STaR, iteratively fine-tune LLMs on self-generated solutions to improve their problem-solving ability. However, these approaches discard the large amounts of incorrect solutions generated during this process, potentially neglecting valuable information in such solutions. To address this shortcoming, we propose V-STaR that utilizes both the correct and incorrect solutions generated during the self-improvement process to train a verifier using DPO that judges correctness of model-generated solutions. This verifier is used at inference time to select one solution among many candidate solutions. Running V-STaR for multiple iterations results in progressively better reasoners and verifiers, delivering a 4% to 17% test accuracy improvement over existing self-improvement and verification approaches on common code generation and math reasoning benchmarks with LLaMA2 models.

Anirudh Goyal, Alessandro Sordoni, Marc-Alexandre Côté et al.

Many efforts have been devoted to training generative latent variable models with autoregressive decoders, such as recurrent neural networks (RNN). Stochastic recurrent models have been successful in capturing the variability observed in natural sequential data such as speech. We unify successful ideas from recently proposed architectures into a stochastic recurrent model: each step in the sequence is associated with a latent variable that is used to condition the recurrent dynamics for future steps. Training is performed with amortized variational inference where the approximate posterior is augmented with a RNN that runs backward through the sequence. In addition to maximizing the variational lower bound, we ease training of the latent variables by adding an auxiliary cost which forces them to reconstruct the state of the backward recurrent network. This provides the latent variables with a task-independent objective that enhances the performance of the overall model. We found this strategy to perform better than alternative approaches such as KL annealing. Although being conceptually simple, our model achieves state-of-the-art results on standard speech benchmarks such as TIMIT and Blizzard and competitive performance on sequential MNIST. Finally, we apply our model to language modeling on the IMDB dataset where the auxiliary cost helps in learning interpretable latent variables. Source Code: \url{https://github.com/anirudh9119/zforcing_nips17}

Sophie Xhonneux, Alessandro Sordoni, Stephan Günnemann et al.

Large language models (LLMs) are vulnerable to adversarial attacks that can bypass their safety guardrails. In many domains, adversarial training has proven to be one of the most promising methods to reliably improve robustness against such attacks. Yet, in the context of LLMs, current methods for adversarial training are hindered by the high computational costs required to perform discrete adversarial attacks at each training iteration. We address this problem by instead calculating adversarial attacks in the continuous embedding space of the LLM, which is orders of magnitudes more efficient. We propose a fast adversarial training algorithm (C-AdvUL) composed of two losses: the first makes the model robust on continuous embedding attacks computed on an adversarial behaviour dataset; the second ensures the usefulness of the final model by fine-tuning on utility data. Moreover, we introduce C-AdvIPO, an adversarial variant of IPO that does not require utility data for adversarially robust alignment. Our empirical evaluation on five models from different families (Gemma, Phi3, Mistral, Zephyr, Llama2) and at different scales (2B, 3.8B, 7B) shows that both algorithms substantially enhance LLM robustness against discrete attacks (GCG, AutoDAN, PAIR), while maintaining utility. Our results demonstrate that robustness to continuous perturbations can extrapolate to discrete threat models. Thereby, we present a path toward scalable adversarial training algorithms for robustly aligning LLMs.

Oleksiy Ostapenko, Zhan Su, Edoardo Maria Ponti et al.

The growing number of parameter-efficient adaptations of a base large language model (LLM) calls for studying whether we can reuse such trained adapters to improve performance for new tasks. We study how to best build a library of adapters given multi-task data and devise techniques for both zero-shot and supervised task generalization through routing in such library. We benchmark existing approaches to build this library and introduce model-based clustering, MBC, a method that groups tasks based on the similarity of their adapter parameters, indirectly optimizing for transfer across the multi-task dataset. To re-use the library, we present a novel zero-shot routing mechanism, Arrow, which enables dynamic selection of the most relevant adapters for new inputs without the need for retraining. We experiment with several LLMs, such as Phi-2 and Mistral, on a wide array of held-out tasks, verifying that MBC-based adapters and Arrow routing lead to superior generalization to new tasks. We make steps towards creating modular, adaptable LLMs that can match or outperform traditional joint training.

Kusha Sareen, Morgane M Moss, Alessandro Sordoni et al.

Prevalent reinforcement learning~(RL) methods for fine-tuning LLM reasoners, such as GRPO or Leave-one-out PPO, abandon the learned value function in favor of empirically estimated returns. This hinders test-time compute scaling that relies on using the value-function for verification. In this work, we propose RL$^V$ that augments any ``value-free'' RL method by jointly training the LLM as both a reasoner and a generative verifier using RL-generated data, adding verification capabilities without significant overhead. Empirically, RL$^V$ boosts MATH accuracy by over 20\% with parallel sampling and enables $8-32\times$ efficient test-time compute scaling compared to the base RL method. RL$^V$ also exhibits strong generalization capabilities for both easy-to-hard and out-of-domain tasks. Furthermore, RL$^V$ achieves $1.2-1.6\times$ higher performance when jointly scaling parallel and sequential test-time compute with a long reasoning R1 model.

Jean-Philippe Corbeil, Amin Dada, Jean-Michel Attendu et al.

High computation costs and latency of large language models such as GPT-4 have limited their deployment in clinical settings. Small language models (SLMs) offer a cost-effective alternative, but their limited capacity requires biomedical domain adaptation, which remains challenging. An additional bottleneck is the unavailability and high sensitivity of clinical data. To address these challenges, we propose a novel framework for adapting SLMs into high-performing clinical models. We introduce the MediPhi collection of 3.8B-parameter SLMs developed with our novel framework: pre-instruction tuning of experts on relevant medical and clinical corpora (PMC, Medical Guideline, MedWiki, etc.), model merging, and clinical-tasks alignment. To cover most clinical tasks, we extended the CLUE benchmark to CLUE+, doubling its size. Our expert models deliver relative improvements on this benchmark over the base model without any task-specific fine-tuning: 64.3% on medical entities, 49.5% on radiology reports, and 44% on ICD-10 coding (outperforming GPT-4-0125 by 14%). We unify the expert models into MediPhi via model merging, preserving gains across benchmarks. Furthermore, we built the MediFlow collection, a synthetic dataset of 2.5 million high-quality instructions on 14 medical NLP tasks, 98 fine-grained document types, and JSON format support. Alignment of MediPhi using supervised fine-tuning and direct preference optimization achieves further gains of 18.9% on average.

Xingdi Yuan, Morgane M Moss, Charbel El Feghali et al. · microsoft-research

Large Language Models (LLMs) are increasingly relied upon for coding tasks, yet in most scenarios it is assumed that all relevant information can be either accessed in context or matches their training data. We posit that LLMs can benefit from the ability to interactively explore a codebase to gather the information relevant to their task. To achieve this, we present a textual environment, namely debug-gym, for developing LLM-based agents in an interactive coding setting. Our environment is lightweight and provides a preset of useful tools, such as a Python debugger (pdb), designed to facilitate an LLM-based agent's interactive debugging. Beyond coding and debugging tasks, this approach can be generalized to other tasks that would benefit from information-seeking behavior by an LLM agent.

Lucas Caccia, Alan Ansell, Edoardo Ponti et al.

Dynamically integrating new or rapidly evolving information after (Large) Language Model pre-training remains challenging, particularly in low-data scenarios or when dealing with private and specialized documents. In-context learning and retrieval-augmented generation (RAG) face limitations, including their high inference costs and their inability to capture global document information. In this paper, we propose a way of modularizing knowledge by training document-level Knowledge Modules (KMs). KMs are lightweight components implemented as parameter-efficient LoRA modules, which are trained to store information about new documents and can be easily plugged into models on demand. We show that next-token prediction performs poorly as the training objective for KMs. We instead propose Deep Context Distillation: we learn KMs parameters such as to simulate hidden states and logits of a teacher that takes the document in context. Our method outperforms standard next-token prediction and pre-instruction training techniques, across two datasets. Finally, we highlight synergies between KMs and RAG.

Atharv Sonwane, Isadora White, Hyunji Lee et al.

High quality bugs are key to training the next generation of language model based software engineering (SWE) agents. We introduce a novel method for synthetic generation of difficult and diverse bugs. Our method instructs SWE Agents to introduce a feature into the codebase whereby they may unintentionally break tests, resulting in bugs. Prior approaches often induce an out-of-distribution effect by generating bugs intentionally (e.g. by introducing local perturbation to existing code), which does not reflect realistic development processes. We perform qualitative analysis to demonstrate that our approach for generating bugs more closely reflects the patterns found in human-authored edits. Through extensive experiments, we demonstrate that our bugs provide more efficient training data for supervised fine-tuning, outperforming other bug datasets by 2% with half the training data (1.2k vs. 3k bugs). We train on our newly generated bugs in addition to existing bug datasets to get FrogBoss a state-of-the-art 32B parameter model on SWE-bench Verified with a pass@1 of 54.6% and FrogMini a state-of-the-art 14B model on SWE-bench Verified with a pass@1 of 45.3% on SWE-bench Verified averaged over three seeds.

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.

Jean-Philippe Corbeil, Minseon Kim, Alessandro Sordoni et al.

As the performance of large language models (LLMs) continues to advance, their adoption is expanding across a wide range of domains, including the medical field. The integration of LLMs into medical applications raises critical safety concerns, particularly due to their use by users with diverse roles, e.g. patients and clinicians, and the potential for model's outputs to directly affect human health. Despite the domain-specific capabilities of medical LLMs, prior safety evaluations have largely focused only on general safety benchmarks. In this paper, we introduce a safety evaluation protocol tailored to the medical domain in both patient user and clinician user perspectives, alongside general safety assessments and quantitatively analyze the safety of medical LLMs. We bridge a gap in the literature by building the PatientSafetyBench containing 466 samples over 5 critical categories to measure safety from the perspective of the patient. We apply our red-teaming protocols on the MediPhi model collection as a case study. To our knowledge, this is the first work to define safety evaluation criteria for medical LLMs through targeted red-teaming taking three different points of view - patient, clinician, and general user - establishing a foundation for safer deployment in medical domains.

Samin Yeasar Arnob, Zhan Su, Minseon Kim et al. · mila

Merging parameter-efficient task experts has recently gained growing attention as a way to build modular architectures that can be rapidly adapted on the fly for specific downstream tasks, without requiring additional fine-tuning. Typically, LoRA serves as the foundational building block of such parameter-efficient modular architectures, leveraging low-rank weight structures to reduce the number of trainable parameters. In this paper, we study the properties of sparse adapters, which train only a subset of weights in the base neural network, as potential building blocks of modular architectures. First, we propose a simple method for training highly effective sparse adapters, which is conceptually simpler than existing methods in the literature and surprisingly outperforms both LoRA and full fine-tuning in our setting. Next, we investigate the merging properties of these sparse adapters by merging adapters for up to 20 natural language processing tasks, thus scaling beyond what is usually studied in the literature. Our findings demonstrate that sparse adapters yield superior in-distribution performance post-merging compared to LoRA or full model merging. Achieving strong held-out performance remains a challenge for all methods considered.

Edoardo M. Ponti, Alessandro Sordoni, Yoshua Bengio et al.

A modular design encourages neural models to disentangle and recombine different facets of knowledge to generalise more systematically to new tasks. In this work, we assume that each task is associated with a subset of latent discrete skills from a (potentially small) inventory. In turn, skills correspond to parameter-efficient (sparse / low-rank) model parameterisations. By jointly learning these and a task-skill allocation matrix, the network for each task is instantiated as the average of the parameters of active skills. To favour non-trivial soft partitions of skills across tasks, we experiment with a series of inductive biases, such as an Indian Buffet Process prior and a two-speed learning rate. We evaluate our latent-skill model on two main settings: 1) multitask reinforcement learning for grounded instruction following on 8 levels of the BabyAI platform; and 2) few-shot adaptation of pre-trained text-to-text generative models on CrossFit, a benchmark comprising 160 NLP tasks. We find that the modular design of a network significantly increases sample efficiency in reinforcement learning and few-shot generalisation in supervised learning, compared to baselines with fully shared, task-specific, or conditionally generated parameters where knowledge is entangled across tasks. In addition, we show how discrete skills help interpretability, as they yield an explicit hierarchy of tasks.

Ian Porada, Alessandro Sordoni, Jackie Chi Kit Cheung

Transformer models pre-trained with a masked-language-modeling objective (e.g., BERT) encode commonsense knowledge as evidenced by behavioral probes; however, the extent to which this knowledge is acquired by systematic inference over the semantics of the pre-training corpora is an open question. To answer this question, we selectively inject verbalized knowledge into the minibatches of a BERT model during pre-training and evaluate how well the model generalizes to supported inferences. We find generalization does not improve over the course of pre-training, suggesting that commonsense knowledge is acquired from surface-level, co-occurrence patterns rather than induced, systematic reasoning.

Meghana Moorthy Bhat, Alessandro Sordoni, Subhabrata Mukherjee

While pre-trained language models have obtained state-of-the-art performance for several natural language understanding tasks, they are quite opaque in terms of their decision-making process. While some recent works focus on rationalizing neural predictions by highlighting salient concepts in the text as justifications or rationales, they rely on thousands of labeled training examples for both task labels as well as an-notated rationales for every instance. Such extensive large-scale annotations are infeasible to obtain for many tasks. To this end, we develop a multi-task teacher-student framework based on self-training language models with limited task-specific labels and rationales, and judicious sample selection to learn from informative pseudo-labeled examples1. We study several characteristics of what constitutes a good rationale and demonstrate that the neural model performance can be significantly improved by making it aware of its rationalized predictions, particularly in low-resource settings. Extensive experiments in several bench-mark datasets demonstrate the effectiveness of our approach.

Eva Portelance, Michael C. Frank, Dan Jurafsky et al.

By the age of two, children tend to assume that new word categories are based on objects' shape, rather than their color or texture; this assumption is called the shape bias. They are thought to learn this bias by observing that their caregiver's language is biased towards shape based categories. This presents a chicken and egg problem: if the shape bias must be present in the language in order for children to learn it, how did it arise in language in the first place? In this paper, we propose that communicative efficiency explains both how the shape bias emerged and why it persists across generations. We model this process with neural emergent language agents that learn to communicate about raw pixelated images. First, we show that the shape bias emerges as a result of efficient communication strategies employed by agents. Second, we show that pressure brought on by communicative need is also necessary for it to persist across generations; simply having a shape bias in an agent's input language is insufficient. These results suggest that, over and above the operation of other learning strategies, the shape bias in human learners may emerge and be sustained by communicative pressures.

Alessandro Sordoni, Nouha Dziri, Hannes Schulz et al.

Recent contrastive representation learning methods rely on estimating mutual information (MI) between multiple views of an underlying context. E.g., we can derive multiple views of a given image by applying data augmentation, or we can split a sequence into views comprising the past and future of some step in the sequence. Contrastive lower bounds on MI are easy to optimize, but have a strong underestimation bias when estimating large amounts of MI. We propose decomposing the full MI estimation problem into a sum of smaller estimation problems by splitting one of the views into progressively more informed subviews and by applying the chain rule on MI between the decomposed views. This expression contains a sum of unconditional and conditional MI terms, each measuring modest chunks of the total MI, which facilitates approximation via contrastive bounds. To maximize the sum, we formulate a contrastive lower bound on the conditional MI which can be approximated efficiently. We refer to our general approach as Decomposed Estimation of Mutual Information (DEMI). We show that DEMI can capture a larger amount of MI than standard non-decomposed contrastive bounds in a synthetic setting, and learns better representations in a vision domain and for dialogue generation.

Arian Hosseini, Siva Reddy, Dzmitry Bahdanau et al.

Negation is a core construction in natural language. Despite being very successful on many tasks, state-of-the-art pre-trained language models often handle negation incorrectly. To improve language models in this regard, we propose to augment the language modeling objective with an unlikelihood objective that is based on negated generic sentences from a raw text corpus. By training BERT with the resulting combined objective we reduce the mean top~1 error rate to 4% on the negated LAMA dataset. We also see some improvements on the negated NLI benchmarks.

Jacob Louis Hoover, Alessandro Sordoni, Wenyu Du et al.

Are pairs of words that tend to occur together also likely to stand in a linguistic dependency? This empirical question is motivated by a long history of literature in cognitive science, psycholinguistics, and NLP. In this work we contribute an extensive analysis of the relationship between linguistic dependencies and statistical dependence between words. Improving on previous work, we introduce the use of large pretrained language models to compute contextualized estimates of the pointwise mutual information between words (CPMI). For multiple models and languages, we extract dependency trees which maximize CPMI, and compare to gold standard linguistic dependencies. Overall, we find that CPMI dependencies achieve an unlabelled undirected attachment score of at most $\approx 0.5$. While far above chance, and consistently above a non-contextualized PMI baseline, this score is generally comparable to a simple baseline formed by connecting adjacent words. We analyze which kinds of linguistic dependencies are best captured in CPMI dependencies, and also find marked differences between the estimates of the large pretrained language models, illustrating how their different training schemes affect the type of dependencies they capture.

Yikang Shen, Shawn Tan, Alessandro Sordoni et al.

Syntax is fundamental to our thinking about language. Failing to capture the structure of input language could lead to generalization problems and over-parametrization. In the present work, we propose a new syntax-aware language model: Syntactic Ordered Memory (SOM). The model explicitly models the structure with an incremental parser and maintains the conditional probability setting of a standard language model (left-to-right). To train the incremental parser and avoid exposure bias, we also propose a novel dynamic oracle, so that SOM is more robust to wrong parsing decisions. Experiments show that SOM can achieve strong results in language modeling, incremental parsing and syntactic generalization tests, while using fewer parameters than other models.

Shawn Tan, Yikang Shen, Timothy J. O'Donnell et al.

We model the recursive production property of context-free grammars for natural and synthetic languages. To this end, we present a dynamic programming algorithm that marginalises over latent binary tree structures with $N$ leaves, allowing us to compute the likelihood of a sequence of $N$ tokens under a latent tree model, which we maximise to train a recursive neural function. We demonstrate performance on two synthetic tasks: SCAN (Lake and Baroni, 2017), where it outperforms previous models on the LENGTH split, and English question formation (McCoy et al., 2020), where it performs comparably to decoders with the ground-truth tree structure. We also present experimental results on German-English translation on the Multi30k dataset (Elliott et al., 2016), and qualitatively analyse the induced tree structures our model learns for the SCAN tasks and the German-English translation task.

Tu Vu, Tong Wang, Tsendsuren Munkhdalai et al.

Recent advances in NLP demonstrate the effectiveness of training large-scale language models and transferring them to downstream tasks. Can fine-tuning these models on tasks other than language modeling further improve performance? In this paper, we conduct an extensive study of the transferability between 33 NLP tasks across three broad classes of problems (text classification, question answering, and sequence labeling). Our results show that transfer learning is more beneficial than previously thought, especially when target task data is scarce, and can improve performance even when the source task is small or differs substantially from the target task (e.g., part-of-speech tagging transfers well to the DROP QA dataset). We also develop task embeddings that can be used to predict the most transferable source tasks for a given target task, and we validate their effectiveness in experiments controlled for source and target data size. Overall, our experiments reveal that factors such as source data size, task and domain similarity, and task complexity all play a role in determining transferability.

Igor Shalyminov, Alessandro Sordoni, Adam Atkinson et al.

Domain adaptation has recently become a key problem in dialogue systems research. Deep learning, while being the preferred technique for modeling such systems, works best given massive training data. However, in the real-world scenario, such resources aren't available for every new domain, so the ability to train with a few dialogue examples can be considered essential. Pre-training on large data sources and adapting to the target data has become the standard method for few-shot problems within the deep learning framework. In this paper, we present the winning entry at the fast domain adaptation task of DSTC8, a hybrid generative-retrieval model based on GPT-2 fine-tuned to the multi-domain MetaLWOz dataset. Robust and diverse in response generation, our model uses retrieval logic as a fallback, being SoTA on MetaLWOz in human evaluation (>4% improvement over the 2nd place system) and attaining competitive generalization performance in adaptation to the unseen MultiWOZ dataset.

Laura Dietz, Bhaskar Mitra, Jeremy Pickens et al.

The vision of HIPstIR is that early stage information retrieval (IR) researchers get together to develop a future for non-mainstream ideas and research agendas in IR. The first iteration of this vision materialized in the form of a three day workshop in Portsmouth, New Hampshire attended by 24 researchers across academia and industry. Attendees pre-submitted one or more topics that they want to pitch at the meeting. Then over the three days during the workshop, we self-organized into groups and worked on six specific proposals of common interest. In this report, we present an overview of the workshop and brief summaries of the six proposals that resulted from the workshop.

Yadollah Yaghoobzadeh, Soroush Mehri, Remi Tachet et al.

Neural NLP models tend to rely on spurious correlations between labels and input features to perform their tasks. Minority examples, i.e., examples that contradict the spurious correlations present in the majority of data points, have been shown to increase the out-of-distribution generalization of pre-trained language models. In this paper, we first propose using example forgetting to find minority examples without prior knowledge of the spurious correlations present in the dataset. Forgettable examples are instances either learned and then forgotten during training or never learned. We empirically show how these examples are related to minorities in our training sets. Then, we introduce a new approach to robustify models by fine-tuning our models twice, first on the full training data and second on the minorities only. We obtain substantial improvements in out-of-distribution generalization when applying our approach to the MNLI, QQP, and FEVER datasets.

Yikang Shen, Shawn Tan, Arian Hosseini et al.

Stack-augmented recurrent neural networks (RNNs) have been of interest to the deep learning community for some time. However, the difficulty of training memory models remains a problem obstructing the widespread use of such models. In this paper, we propose the Ordered Memory architecture. Inspired by Ordered Neurons (Shen et al., 2018), we introduce a new attention-based mechanism and use its cumulative probability to control the writing and erasing operation of the memory. We also introduce a new Gated Recursive Cell to compose lower-level representations into higher-level representation. We demonstrate that our model achieves strong performance on the logical inference task (Bowman et al., 2015)and the ListOps (Nangia and Bowman, 2018) task. We can also interpret the model to retrieve the induced tree structure, and find that these induced structures align with the ground truth. Finally, we evaluate our model on the Stanford SentimentTreebank tasks (Socher et al., 2013), and find that it performs comparatively with the state-of-the-art methods in the literature.

Tsendsuren Munkhdalai, Alessandro Sordoni, Tong Wang et al.

We augment recurrent neural networks with an external memory mechanism that builds upon recent progress in metalearning. We conceptualize this memory as a rapidly adaptable function that we parameterize as a deep neural network. Reading from the neural memory function amounts to pushing an input (the key vector) through the function to produce an output (the value vector). Writing to memory means changing the function; specifically, updating the parameters of the neural network to encode desired information. We leverage training and algorithmic techniques from metalearning to update the neural memory function in one shot. The proposed memory-augmented model achieves strong performance on a variety of learning problems, from supervised question answering to reinforcement learning.

Mariya Toneva, Alessandro Sordoni, Remi Tachet des Combes et al.

Inspired by the phenomenon of catastrophic forgetting, we investigate the learning dynamics of neural networks as they train on single classification tasks. Our goal is to understand whether a related phenomenon occurs when data does not undergo a clear distributional shift. We define a `forgetting event' to have occurred when an individual training example transitions from being classified correctly to incorrectly over the course of learning. Across several benchmark data sets, we find that: (i) certain examples are forgotten with high frequency, and some not at all; (ii) a data set's (un)forgettable examples generalize across neural architectures; and (iii) based on forgetting dynamics, a significant fraction of examples can be omitted from the training data set while still maintaining state-of-the-art generalization performance.

Yikang Shen, Shawn Tan, Alessandro Sordoni et al.

Natural language is hierarchically structured: smaller units (e.g., phrases) are nested within larger units (e.g., clauses). When a larger constituent ends, all of the smaller constituents that are nested within it must also be closed. While the standard LSTM architecture allows different neurons to track information at different time scales, it does not have an explicit bias towards modeling a hierarchy of constituents. This paper proposes to add such an inductive bias by ordering the neurons; a vector of master input and forget gates ensures that when a given neuron is updated, all the neurons that follow it in the ordering are also updated. Our novel recurrent architecture, ordered neurons LSTM (ON-LSTM), achieves good performance on four different tasks: language modeling, unsupervised parsing, targeted syntactic evaluation, and logical inference.

Philip Bachman, Riashat Islam, Alessandro Sordoni et al.

We introduce a deep generative model for functions. Our model provides a joint distribution p(f, z) over functions f and latent variables z which lets us efficiently sample from the marginal p(f) and maximize a variational lower bound on the entropy H(f). We can thus maximize objectives of the form E_{f~p(f)}[R(f)] + c*H(f), where R(f) denotes, e.g., a data log-likelihood term or an expected reward. Such objectives encompass Bayesian deep learning in function space, rather than parameter space, and Bayesian deep RL with representations of uncertainty that offer benefits over bootstrapping and parameter noise. In this short paper we describe our model, situate it in the context of prior work, and present proof-of-concept experiments for regression and RL.

Xingdi Yuan, Marc-Alexandre Côté, Alessandro Sordoni et al.

We propose a recurrent RL agent with an episodic exploration mechanism that helps discovering good policies in text-based game environments. We show promising results on a set of generated text-based games of varying difficulty where the goal is to collect a coin located at the end of a chain of rooms. In contrast to previous text-based RL approaches, we observe that our agent learns policies that generalize to unseen games of greater difficulty.

Nan Rosemary Ke, Konrad Zolna, Alessandro Sordoni et al.

Recurrent Neural Networks (RNNs) with attention mechanisms have obtained state-of-the-art results for many sequence processing tasks. Most of these models use a simple form of encoder with attention that looks over the entire sequence and assigns a weight to each token independently. We present a mechanism for focusing RNN encoders for sequence modelling tasks which allows them to attend to key parts of the input as needed. We formulate this using a multi-layer conditional sequence encoder that reads in one token at a time and makes a discrete decision on whether the token is relevant to the context or question being asked. The discrete gating mechanism takes in the context embedding and the current hidden state as inputs and controls information flow into the layer above. We train it using policy gradient methods. We evaluate this method on several types of tasks with different attributes. First, we evaluate the method on synthetic tasks which allow us to evaluate the model for its generalization ability and probe the behavior of the gates in more controlled settings. We then evaluate this approach on large scale Question Answering tasks including the challenging MS MARCO and SearchQA tasks. Our models shows consistent improvements for both tasks over prior work and our baselines. It has also shown to generalize significantly better on synthetic tasks as compared to the baselines.

Yikang Shen, Zhouhan Lin, Athul Paul Jacob et al.

In this work, we propose a novel constituency parsing scheme. The model predicts a vector of real-valued scalars, named syntactic distances, for each split position in the input sentence. The syntactic distances specify the order in which the split points will be selected, recursively partitioning the input, in a top-down fashion. Compared to traditional shift-reduce parsing schemes, our approach is free from the potential problem of compounding errors, while being faster and easier to parallelize. Our model achieves competitive performance amongst single model, discriminative parsers in the PTB dataset and outperforms previous models in the CTB dataset.

Amjad Almahairi, Sai Rajeswar, Alessandro Sordoni et al.

Learning inter-domain mappings from unpaired data can improve performance in structured prediction tasks, such as image segmentation, by reducing the need for paired data. CycleGAN was recently proposed for this problem, but critically assumes the underlying inter-domain mapping is approximately deterministic and one-to-one. This assumption renders the model ineffective for tasks requiring flexible, many-to-many mappings. We propose a new model, called Augmented CycleGAN, which learns many-to-many mappings between domains. We examine Augmented CycleGAN qualitatively and quantitatively on several image datasets.

Dmitriy Serdyuk, Nan Rosemary Ke, Alessandro Sordoni et al.

We propose a simple technique for encouraging generative RNNs to plan ahead. We train a "backward" recurrent network to generate a given sequence in reverse order, and we encourage states of the forward model to predict cotemporal states of the backward model. The backward network is used only during training, and plays no role during sampling or inference. We hypothesize that our approach eases modeling of long-term dependencies by implicitly forcing the forward states to hold information about the longer-term future (as contained in the backward states). We show empirically that our approach achieves 9% relative improvement for a speech recognition task, and achieves significant improvement on a COCO caption generation task.