分享

For visual recognition, knowledge distillation typically involves transferring knowledge from a large, well-trained teacher model to a smaller student model. In this paper, we introduce an effective method to distill knowledge from an off-the-shelf vision-language model (VLM), demonstrating that it provides novel supervision in addition to those from a conventional vision-only teacher model. Our key technical contribution is the development of a framework that generates novel text supervision and distills free-form text into a vision encoder. We showcase the effectiveness of our approach, termed VLM-KD, across various benchmark datasets, showing that it surpasses several state-of-the-art long-tail visual classifiers. To our knowledge, this work is the first to utilize knowledge distillation with text supervision generated by an off-the-shelf VLM and apply it to vanilla randomly initialized vision encoders.

Teaching new information to pre-trained large language models (PLM) is a crucial but challenging task. Model adaptation techniques, such as fine-tuning and parameter-efficient training have been shown to store new facts at a slow rate; continual learning is an option but is costly and prone to catastrophic forgetting. This work studies and quantifies how PLM may learn and remember new world knowledge facts that do not occur in their pre-training corpus, which only contains world knowledge up to a certain date. To that purpose, we first propose Novel-WD, a new dataset consisting of sentences containing novel facts extracted from recent Wikidata updates, along with two evaluation tasks in the form of causal language modeling and multiple choice questions (MCQ). We make this dataset freely available to the community, and release a procedure to later build new versions of similar datasets with up-to-date information. We also explore the use of prefix-tuning for novel information learning, and analyze how much information can be stored within a given prefix. We show that a single fact can reliably be encoded within a single prefix, and that the prefix capacity increases with its length and with the base model size.

High-resource language models often fall short in the African context, where there is a critical need for models that are efficient, accessible, and locally relevant, even amidst significant computing and data constraints. This paper introduces InkubaLM, a small language model with 0.4 billion parameters, which achieves performance comparable to models with significantly larger parameter counts and more extensive training data on tasks such as machine translation, question-answering, AfriMMLU, and the AfriXnli task. Notably, InkubaLM outperforms many larger models in sentiment analysis and demonstrates remarkable consistency across multiple languages. This work represents a pivotal advancement in challenging the conventional paradigm that effective language models must rely on substantial resources. Our model and datasets are publicly available at https://huggingface.co/lelapa to encourage research and development on low-resource languages.

This paper proposes a method for extracting speaker embedding for each speaker from a variable-length recording containing multiple speakers. Speaker embeddings are crucial not only for speaker recognition but also for various multi-speaker speech applications such as speaker diarization and target-speaker speech processing. Despite the challenges of obtaining a single speaker's speech without pre-registration in multi-speaker scenarios, most studies on speaker embedding extraction focus on extracting embeddings only from single-speaker recordings. Some methods have been proposed for extracting speaker embeddings directly from multi-speaker recordings, but they typically require preparing a model for each possible number of speakers or involve complicated training procedures. The proposed method computes the embeddings of multiple speakers by focusing on different parts of the frame-wise embeddings extracted from the input multi-speaker audio. This is achieved by recursively computing attention weights for pooling the frame-wise embeddings. Additionally, we propose using the calculated attention weights to estimate the number of speakers in the recording, which allows the same model to be applied to various numbers of speakers. Experimental evaluations demonstrate the effectiveness of the proposed method in speaker verification and diarization tasks.

The task of multi-modal emotion recognition in conversation (MERC) aims to analyze the genuine emotional state of each utterance based on the multi-modal information in the conversation, which is crucial for conversation understanding. Existing methods focus on using graph neural networks (GNN) to model conversational relationships and capture contextual latent semantic relationships. However, due to the complexity of GNN, existing methods cannot efficiently capture the potential dependencies between long-distance utterances, which limits the performance of MERC. In this paper, we propose an Efficient Long-distance Latent Relation-aware Graph Neural Network (ELR-GNN) for multi-modal emotion recognition in conversations. Specifically, we first use pre-extracted text, video and audio features as input to Bi-LSTM to capture contextual semantic information and obtain low-level utterance features. Then, we use low-level utterance features to construct a conversational emotion interaction graph. To efficiently capture the potential dependencies between long-distance utterances, we use the dilated generalized forward push algorithm to precompute the emotional propagation between global utterances and design an emotional relation-aware operator to capture the potential semantic associations between different utterances. Furthermore, we combine early fusion and adaptive late fusion mechanisms to fuse latent dependency information between speaker relationship information and context. Finally, we obtain high-level discourse features and feed them into MLP for emotion prediction. Extensive experimental results show that ELR-GNN achieves state-of-the-art performance on the benchmark datasets IEMOCAP and MELD, with running times reduced by 52\% and 35\%, respectively.

This paper explores the pitfalls in evaluating multilingual automatic speech recognition (ASR) models, with a particular focus on Indic language scripts. We investigate the text normalization routine employed by leading ASR models, including OpenAI Whisper, Meta's MMS, Seamless, and Assembly AI's Conformer, and their unintended consequences on performance metrics. Our research reveals that current text normalization practices, while aiming to standardize ASR outputs for fair comparison, by removing inconsistencies such as variations in spelling, punctuation, and special characters, are fundamentally flawed when applied to Indic scripts. Through empirical analysis using text similarity scores and in-depth linguistic examination, we demonstrate that these flaws lead to artificially improved performance metrics for Indic languages. We conclude by proposing a shift towards developing text normalization routines that leverage native linguistic expertise, ensuring more robust and accurate evaluations of multilingual ASR models.

Generating diverse responses from large language models (LLMs) is crucial for applications such as planning/search and synthetic data generation, where diversity provides distinct answers across generations. Prior approaches rely on increasing temperature to increase diversity. However, contrary to popular belief, we show not only does this approach produce lower quality individual generations as temperature increases, but it depends on model's next-token probabilities being similar to the true distribution of answers. We propose SimpleStrat, an alternative approach that uses the language model itself to partition the space into strata. At inference, a random stratum is selected and a sample drawn from within the strata. To measure diversity, we introduce CoverageQA, a dataset of underspecified questions with multiple equally plausible answers, and assess diversity by measuring KL Divergence between the output distribution and uniform distribution over valid ground truth answers. As computing probability per response/solution for proprietary models is infeasible, we measure recall on ground truth solutions. Our evaluation show using SimpleStrat achieves higher recall by 0.05 compared to GPT-4o and 0.36 average reduction in KL Divergence compared to Llama 3.

Social conventions are the foundation for social and economic life. As legions of AI agents increasingly interact with each other and with humans, their ability to form shared conventions will determine how effectively they will coordinate behaviors, integrate into society and influence it. Here, we investigate the dynamics of conventions within populations of Large Language Model (LLM) agents using simulated interactions. First, we show that globally accepted social conventions can spontaneously arise from local interactions between communicating LLMs. Second, we demonstrate how strong collective biases can emerge during this process, even when individual agents appear to be unbiased. Third, we examine how minority groups of committed LLMs can drive social change by establishing new social conventions. We show that once these minority groups reach a critical size, they can consistently overturn established behaviors. In all cases, contrasting the experimental results with predictions from a minimal multi-agent model allows us to isolate the specific role of LLM agents. Our results clarify how AI systems can autonomously develop norms without explicit programming and have implications for designing AI systems that align with human values and societal goals.

This paper presents $\forall$uto$\exists$$\lor\!\land$L, a novel benchmark for scaling Large Language Model (LLM) assessment in formal tasks with clear notions of correctness, such as truth maintenance in translation and logical reasoning. $\forall$uto$\exists$$\lor\!\land$L is the first benchmarking paradigm that offers several key advantages necessary for scaling objective evaluation of LLMs without human labeling: (a) ability to evaluate LLMs of increasing sophistication by auto-generating tasks at different levels of difficulty; (b) auto-generation of ground truth that eliminates dependence on expensive and time-consuming human annotation; (c) the use of automatically generated, randomized datasets that mitigate the ability of successive LLMs to overfit to static datasets used in many contemporary benchmarks. Empirical analysis shows that an LLM's performance on $\forall$uto$\exists$$\lor\!\land$L is highly indicative of its performance on a diverse array of other benchmarks focusing on translation and reasoning tasks, making it a valuable autonomous evaluation paradigm in settings where hand-curated datasets can be hard to obtain and/or update.

Logical errors in programming don't raise compiler alerts, making them hard to detect. These silent errors can disrupt a program's function or cause run-time issues. Their correction requires deep insight into the program's logic, highlighting the importance of automated detection and repair. In this paper, we introduce LecPrompt to localize and repair logical errors, an prompt-based approach that harnesses the capabilities of CodeBERT, a transformer-based large language model trained on code. First, LecPrompt leverages a large language model to calculate perplexity and log probability metrics, pinpointing logical errors at both token and line levels. Through statistical analysis, it identifies tokens and lines that deviate significantly from the expected patterns recognized by large language models, marking them as potential error sources. Second, by framing the logical error correction challenge as a Masked Language Modeling (MLM) task, LecPrompt employs CodeBERT to autoregressively repair the identified error tokens. Finally, the soft-prompt method provides a novel solution in low-cost scenarios, ensuring that the model can be fine-tuned to the specific nuances of the logical error correction task without incurring high computational costs. To evaluate LecPrompt's performance, we created a method to introduce logical errors into correct code and applying this on QuixBugs to produce the QuixBugs-LE dataset. Our evaluations on the QuixBugs-LE dataset for both Python and Java highlight the impressive capabilities of our method, LecPrompt. For Python, LecPrompt achieves a noteworthy 74.58% top-1 token-level repair accuracy and 27.4% program-level repair accuracy. In Java, LecPrompt delivers a 69.23\% top-1 token-level repair accuracy and 24.7% full program-level repair accuracy.

Within a perception framework for autonomous mobile and robotic systems, semantic analysis of 3D point clouds typically generated by LiDARs is key to numerous applications, such as object detection and recognition, and scene reconstruction. Scene semantic segmentation can be achieved by directly integrating 3D spatial data with specialized deep neural networks. Although this type of data provides rich geometric information regarding the surrounding environment, it also presents numerous challenges: its unstructured and sparse nature, its unpredictable size, and its demanding computational requirements. These characteristics hinder the real-time semantic analysis, particularly on resource-constrained hardware architectures that constitute the main computational components of numerous robotic applications. Therefore, in this paper, we investigate various 3D semantic segmentation methodologies and analyze their performance and capabilities for resource-constrained inference on embedded NVIDIA Jetson platforms. We evaluate them for a fair comparison through a standardized training protocol and data augmentations, providing benchmark results on the Jetson AGX Orin and AGX Xavier series for two large-scale outdoor datasets: SemanticKITTI and nuScenes.

Eliciting a high-dimensional probability distribution from an expert via noisy judgments is notoriously challenging, yet useful for many applications, such as prior elicitation and reward modeling. We introduce a method for eliciting the expert's belief density as a normalizing flow based solely on preferential questions such as comparing or ranking alternatives. This allows eliciting in principle arbitrarily flexible densities, but flow estimation is susceptible to the challenge of collapsing or diverging probability mass that makes it difficult in practice. We tackle this problem by introducing a novel functional prior for the flow, motivated by a decision-theoretic argument, and show empirically that the belief density can be inferred as the function-space maximum a posteriori estimate. We demonstrate our method by eliciting multivariate belief densities of simulated experts, including the prior belief of a general-purpose large language model over a real-world dataset.