In this work, we present and evaluate SELMA, a Speech-Enabled Language Model for virtual Assistant interactions that integrates audio and text as inputs to a Large Language Model (LLM). SELMA is designed to handle three primary and two auxiliary tasks related to interactions with virtual assistants simultaneously within a single end-to-end model. We employ low-rank adaptation modules for parameter-efficient training of both the audio encoder and the LLM. Additionally, we implement a feature pooling strategy enabling the system to recognize global patterns and improve accuracy on tasks less reliant on individual sequence elements. Experimental results on Voice Trigger (VT) detection, Device-Directed Speech Detection (DDSD), and Automatic Speech Recognition (ASR), demonstrate that our approach both simplifies the typical input processing pipeline of virtual assistants significantly and also improves performance compared to dedicated models for each individual task. SELMA yields relative Equal-Error Rate improvements of 64% on the VT detection task, and 22% on DDSD, while also achieving word error rates close to the baseline.

Generating semantically coherent text requires a robust internal representation of linguistic structures, which traditional embedding techniques often fail to capture adequately. A novel approach, Latent Lexical Projection (LLP), is introduced to refine lexical representations through a structured transformation into a latent space, thereby enhancing the alignment between input embeddings and their contextual meanings. The method integrates an optimized projection mechanism within an existing language model architecture, enabling more accurate token selection while maintaining syntactic integrity. Evaluations across multiple benchmarks indicate a reduction in perplexity and an increase in BLEU scores, suggesting improvements in predictive accuracy and fluency. The analysis of lexical diversity reveals a more varied vocabulary in generated text, addressing common issues of redundancy and repetitive phrase structures. Further assessments of entropy distributions demonstrate a decline in uncertainty during decoding, reflecting enhanced confidence in word selection. Additionally, long-range dependency retention exhibits measurable gains, with increased classification accuracy at extended token distances. Computational efficiency remains within manageable constraints, despite the added projection mechanism, highlighting the practicality of LLP for integration into existing architectures.

It is difficult to anticipate the myriad challenges that a predictive model will encounter once deployed. Common practice entails a reactive, cyclical approach: model deployment, data mining, and retraining. We instead develop a proactive longtail discovery process by imagining additional data during training. In particular, we develop general model-based longtail signals, including a differentiable, single forward pass formulation of epistemic uncertainty that does not impact model parameters or predictive performance but can flag rare or hard inputs. We leverage these signals as guidance to generate additional training data from a latent diffusion model in a process we call Longtail Guidance (LTG). Crucially, we can perform LTG without retraining the diffusion model or the predictive model, and we do not need to expose the predictive model to intermediate diffusion states. Data generated by LTG exhibit semantically meaningful variation, yield significant generalization improvements on image classification benchmarks, and can be analyzed to proactively discover, explain, and address conceptual gaps in a predictive model.

In the realm of large-scale language models, a significant challenge arises when extrapolating sequences beyond the maximum allowable length. This is because the model's position embedding mechanisms are limited to positions encountered during training, thus preventing effective representation of positions in longer sequences. We analyzed conventional position encoding methods for long contexts and found the following characteristics. However, because it uses only a fixed scale parameter, it does not fully exploit the advantages of wavelet transforms, which capture the fine movements of non-stationary signals using multiple scales (window sizes). This limitation could explain why RoPE performs poorly in extrapolation. However, it has limitations in capturing deep dependencies because it restricts the receptive field of the model. From these insights, we propose a new position representation method that captures multiple scales (i.e., window sizes) by leveraging wavelet transforms without limiting the model's attention field. Experimental results show that this new method improves the performance of the model in both short and long contexts. In particular, our method allows extrapolation of position information without limiting the model's attention field. Several pre-trained large language models based on Transformer architecture (Vaswani et al., 2017) have demonstrated robust capabilities in various generative tasks (Devlin et al., 2019; Raffel et al., 2020; Brown et al., 2020; Touvron et al., 2023a; Jiang et al., 2023). However, limitations on the input sequence length arise due to the computational resource constraints encountered during the pre-training phase. This weakness is primarily attributed to the positional encoding's ineffectiveness in handling sequences that exceed the length of those encountered during the model's training phase (Devlin et al., 2019; Press et al., 2022). Rotary Position Embedding (RoPE) (Su et al., 2021) has become a common approach in many language models that handle long contexts, and it employs a rotation matrix to encode positional information and facilitate the processing of long sequences. To manage sequences longer than those encountered during training, various scaling strategies (Chen et al., 2023; bloc97, 2023; Peng et al., 2024; Liu et al., 2024) have been applied to RoPE, although these often require additional finetuning and incur further learning costs in addition to those of pre-training. In contrast, Attention with Linear Biases (ALiBi) (Press et al., 2022) is able to sequence length estimation beyond the limits of pre-training without requiring additional fine-tuning.

Many-shot jailbreaking circumvents the safety alignment of large language models by exploiting their ability to process long input sequences. To achieve this, the malicious target prompt is prefixed with hundreds of fabricated conversational turns between the user and the model. These fabricated exchanges are randomly sampled from a pool of malicious questions and responses, making it appear as though the model has already complied with harmful instructions. In this paper, we present PANDAS: a hybrid technique that improves many-shot jailbreaking by modifying these fabricated dialogues with positive affirmations, negative demonstrations, and an optimized adaptive sampling method tailored to the target prompt's topic. Extensive experiments on AdvBench and HarmBench, using state-of-the-art LLMs, demonstrate that PANDAS significantly outperforms baseline methods in long-context scenarios. Through an attention analysis, we provide insights on how long-context vulnerabilities are exploited and show how PANDAS further improves upon many-shot jailbreaking.

This paper revisits the implementation of $\textbf{L}$oad-$\textbf{b}$alancing $\textbf{L}$oss (LBL) when training Mixture-of-Experts (MoEs) models. Specifically, LBL for MoEs is defined as $N_E \sum_{i=1}^{N_E} f_i p_i$, where $N_E$ is the total number of experts, $f_i$ represents the frequency of expert $i$ being selected, and $p_i$ denotes the average gating score of the expert $i$. Existing MoE training frameworks usually employ the parallel training strategy so that $f_i$ and the LBL are calculated within a $\textbf{micro-batch}$ and then averaged across parallel groups. In essence, a micro-batch for training billion-scale LLMs normally contains very few sequences. So, the micro-batch LBL is almost at the sequence level, and the router is pushed to distribute the token evenly within each sequence. Under this strict constraint, even tokens from a domain-specific sequence ($\textit{e.g.}$, code) are uniformly routed to all experts, thereby inhibiting expert specialization. In this work, we propose calculating LBL using a $\textbf{global-batch}$ to loose this constraint. Because a global-batch contains much more diverse sequences than a micro-batch, which will encourage load balance at the corpus level. Specifically, we introduce an extra communication step to synchronize $f_i$ across micro-batches and then use it to calculate the LBL. Through experiments on training MoEs-based LLMs (up to $\textbf{42.8B}$ total parameters and $\textbf{400B}$ tokens), we surprisingly find that the global-batch LBL strategy yields excellent performance gains in both pre-training perplexity and downstream tasks. Our analysis reveals that the global-batch LBL also greatly improves the domain specialization of MoE experts.

Cass Sunstein's essay 'On Bob Dylan' describes Dylan's 'dishabituating' style -- a constant refusal to conform to expectation and a penchant for reinventing his musical and lyrical identity. In this paper, I extend Sunstein's observations through a large-scale computational analysis of Dylan's lyrics from 1962 to 2012. Using o3-mini-high (a large language model), I extract concept-to-concept relationships from the lyrics and construct directed knowledge graphs that capture Dylan's thematic structure. I then quantify shifts in sentiment, metaphorical expression, thematic diversity, and network complexity over time. The results indicate that Dylan's lyrics increasingly rely on metaphor, display an evolving sentiment profile, and exhibit heightened dishabituation -- measured here as a growing variance in the network centrality of key concepts. I also find that references to movement, protest, and mythic imagery fluctuate in ways that align with well-known phases of Dylan's career, reflecting the dynamic and unpredictable quality of his art. These findings not only deepen our empirical understanding of Sunstein's thesis but also introduce a novel computational method for analyzing an artist's evolution-offering broader applicability to the study of cultural and creative change.

Training Large Language Models (LLMs) demand significant time and computational resources due to their large model sizes and optimizer states. To overcome these challenges, recent methods, such as BAdam, employ partial weight updates to enhance time and memory efficiency, though sometimes at the cost of performance. Others, like GaLore, focus on maintaining performance while optimizing memory usage through full parameter training, but may incur higher time complexity. By leveraging the low-rank structure of the gradient and the Grassmannian geometry, we propose SubTrack-Grad, a subspace tracking-based optimization method that efficiently tracks the evolving gradient subspace by incorporating estimation errors and previously identified subspaces. SubTrack-Grad delivers better or on-par results compared to GaLore, while significantly outperforming BAdam, which, despite being time-efficient, compromises performance. SubTrack-Grad reduces wall-time by up to 20.57% on GLUE tasks (15% average reduction) and up to 65% on SuperGLUE tasks (22% average reduction) compared to GaLore. Notably, for a 3B parameter model, GaLore incurred a substantial 157% increase in wall-time compared to full-rank training, whereas SubTrack-Grad exhibited a 31% increase, representing a 49% reduction in wall-time, while enjoying the same memory reductions as GaLore.

State of the art large language models rely on randomization to respond to a prompt. As an immediate consequence, a model may respond differently to the same prompt if asked multiple times. In this work, we argue that the evaluation and ranking of large language models should control for the randomization underpinning their functioning. Our starting point is the development of a causal model for coupled autoregressive generation, which allows different large language models to sample responses with the same source of randomness. Building upon our causal model, we first show that, on evaluations based on benchmark datasets, coupled autoregressive generation leads to the same conclusions as vanilla autoregressive generation but using provably fewer samples. However, we further show that, on evaluations based on (human) pairwise comparisons, coupled and vanilla autoregressive generation can surprisingly lead to different rankings when comparing more than two models, even with an infinite amount of samples. This suggests that the apparent advantage of a model over others in existing evaluation protocols may not be genuine but rather confounded by the randomness inherent to the generation process. To illustrate and complement our theoretical results, we conduct experiments with several large language models from the Llama family. We find that, across multiple knowledge areas from the popular MMLU benchmark dataset, coupled autoregressive generation requires up to 40% fewer samples to reach the same conclusions as vanilla autoregressive generation. Further, using data from the LMSYS Chatbot Arena platform, we find that the win-rates derived from pairwise comparisons by a strong large language model to prompts differ under coupled and vanilla autoregressive generation.

Large Language Models (LLMs) have shown remarkable potential in reasoning while they still suffer from severe factual hallucinations due to timeliness, accuracy, and coverage of parametric knowledge. Meanwhile, integrating reasoning with retrieval-augmented generation (RAG) remains challenging due to ineffective task decomposition and redundant retrieval, which can introduce noise and degrade response quality. In this paper, we propose DeepRAG, a framework that models retrieval-augmented reasoning as a Markov Decision Process (MDP), enabling strategic and adaptive retrieval. By iteratively decomposing queries, DeepRAG dynamically determines whether to retrieve external knowledge or rely on parametric reasoning at each step. Experiments show that DeepRAG improves retrieval efficiency while improving answer accuracy by 21.99%, demonstrating its effectiveness in optimizing retrieval-augmented reasoning.