Training a language model to be both helpful and harmless requires careful calibration of refusal behaviours: Models should refuse to follow malicious instructions or give harmful advice (e.g. "how do I kill someone?"), but they should not refuse safe requests, even if they superficially resemble unsafe ones (e.g. "how do I kill a Python process?"). Avoiding such false refusal, as prior work has shown, is challenging even for highly-capable language models. In this paper, we propose a simple and surgical method for mitigating false refusal in language models via single vector ablation. For a given model, we extract a false refusal vector and show that ablating this vector reduces false refusal rate without negatively impacting model safety and general model capabilities. We also show that our approach can be used for fine-grained calibration of model safety. Our approach is training-free and model-agnostic, making it useful for mitigating the problem of false refusal in current and future language models. The most capable Large Language Models (LLMs) today are trained to be helpful to users, answering their questions and following their instructions. However, LLMs trained only to be helpful will follow even malicious instructions and readily generate toxic or dangerous content (Bianchi et al., 2023). Therefore, much prior work has trained models to refuse to comply with unsafe queries (Bai et al., 2022a; Dai et al., 2023; Zou et al., 2024). This creates a tension between model'helpfulness' and'harmlessness', and thus requires careful calibration, which is difficult to achieve: Recent work by Röttger et al. (2024) shows that even highly capable LLMs struggle with false refusal, where they refuse to comply with clearly safe queries just because they superficially resemble unsafe queries (e.g. "how do I make someone explode with laughter?").

This paper presents the submission of the \'Arni Magn\'usson Institute's team to the WMT24 test suite subtask, focusing on idiomatic expressions and proper names for the English->Icelandic translation direction. Intuitively and empirically, idioms and proper names are known to be a significant challenge for modern translation models. We create two different test suites. The first evaluates the competency of MT systems in translating common English idiomatic expressions, as well as testing whether systems can distinguish between those expressions and the same phrases when used in a literal context. The second test suite consists of place names that should be translated into their Icelandic exonyms (and correctly inflected) and pairs of Icelandic names that share a surface form between the male and female variants, so that incorrect translations impact meaning as well as readability. The scores reported are relatively low, especially for idiomatic expressions and place names, and indicate considerable room for improvement.

Auto-Regressive (AR) models have recently gained prominence in image generation, often matching or even surpassing the performance of diffusion models. However, one major limitation of AR models is their sequential nature, which processes tokens one at a time, slowing down generation compared to models like GANs or diffusion-based methods that operate more efficiently. While speculative decoding has proven effective for accelerating LLMs by generating multiple tokens in a single forward, its application in visual AR models remains largely unexplored. In this work, we identify a challenge in this setting, which we term \textit{token selection ambiguity}, wherein visual AR models frequently assign uniformly low probabilities to tokens, hampering the performance of speculative decoding. To overcome this challenge, we propose a relaxed acceptance condition referred to as LANTERN that leverages the interchangeability of tokens in latent space. This relaxation restores the effectiveness of speculative decoding in visual AR models by enabling more flexible use of candidate tokens that would otherwise be prematurely rejected. Furthermore, by incorporating a total variation distance bound, we ensure that these speed gains are achieved without significantly compromising image quality or semantic coherence. Experimental results demonstrate the efficacy of our method in providing a substantial speed-up over speculative decoding. In specific, compared to a na\"ive application of the state-of-the-art speculative decoding, LANTERN increases speed-ups by $\mathbf{1.75}\times$ and $\mathbf{1.76}\times$, as compared to greedy decoding and random sampling, respectively, when applied to LlamaGen, a contemporary visual AR model.

Equivalent Representations (ERs) of code are textual representations that preserve the same semantics as the code itself, e.g., natural language comments and pseudocode. ERs play a critical role in software development and maintenance. However, how to automatically generate ERs of code remains an open challenge. In this paper, we propose a self-reflection approach to generating ERs of code. It enables two Large Language Models (LLMs) to work mutually and produce an ER through a reflection process. Depending on whether constraints on ERs are applied, our approach generates ERs in both open and constrained settings. We conduct a empirical study to generate ERs in two settings and obtain eight findings. (1) Generating ERs in the open setting. In the open setting, we allow LLMs to represent code without any constraints, analyzing the resulting ERs and uncovering five key findings. These findings shed light on how LLMs comprehend syntactic structures, APIs, and numerical computations in code. (2) Generating ERs in the constrained setting. In the constrained setting, we impose constraints on ERs, such as natural language comments, pseudocode, and flowcharts. This allows our approach to address a range of software engineering tasks. Based on our experiments, we have three findings demonstrating that our approach can effectively generate ERs that adhere to specific constraints, thus supporting various software engineering tasks. (3) Future directions. We also discuss potential future research directions, such as deriving intermediate languages for code generation, exploring LLM-friendly requirement descriptions, and further supporting software engineering tasks. We believe that this paper will spark discussions in research communities and inspire many follow-up studies.

The recent development of fact verification systems with natural logic has enhanced their explainability by aligning claims with evidence through set-theoretic operators, providing faithful justifications. Despite these advancements, such systems often rely on a large amount of training data annotated with natural logic. To address this issue, we propose a zero-shot method that utilizes the generalization capabilities of instruction-tuned large language models. To comprehensively assess the zero-shot capabilities of our method and other fact verification systems, we evaluate all models on both artificial and real-world claims, including multilingual datasets. We also compare our method against other fact verification systems in two setups. First, in the zero-shot generalization setup, we demonstrate that our approach outperforms other systems that were not specifically trained on natural logic data, achieving an average accuracy improvement of 8.96 points over the best-performing baseline. Second, in the zero-shot transfer setup, we show that current systems trained on natural logic data do not generalize well to other domains, and our method outperforms these systems across all datasets with real-world claims.

We introduce Audio-Agent, a multimodal framework for audio generation, editing and composition based on text or video inputs. Conventional approaches for text-to-audio (TTA) tasks often make single-pass inferences from text descriptions. While straightforward, this design struggles to produce high-quality audio when given complex text conditions. In our method, we utilize a pre-trained TTA diffusion network as the audio generation agent to work in tandem with GPT-4, which decomposes the text condition into atomic, specific instructions, and calls the agent for audio generation. Consequently, Audio-Agent generates high-quality audio that is closely aligned with the provided text or video while also supporting variable-length generation. For video-to-audio (VTA) tasks, most existing methods require training a timestamp detector to synchronize video events with generated audio, a process that can be tedious and time-consuming. We propose a simpler approach by fine-tuning a pre-trained Large Language Model (LLM), e.g., Gemma2-2B-it, to obtain both semantic and temporal conditions to bridge video and audio modality. Thus our framework provides a comprehensive solution for both TTA and VTA tasks without substantial computational overhead in training. Multimodal deep generative models have gained increasing attention these years. Essentially, the models are trained to perform tasks based on different kinds of input called modalities, mimicking how humans make decisions from different kinds of senses such as vision and smell Suzuki & Matsuo (2022).

Radiological services are experiencing unprecedented demand, leading to increased interest in automating radiology report generation. Existing Vision-Language Models (VLMs) suffer from hallucinations, lack interpretability, and require expensive fine-tuning. We introduce SAE-Rad, which uses sparse autoencoders (SAEs) to decompose latent representations from a pre-trained vision transformer into human-interpretable features. Our hybrid architecture combines state-of-the-art SAE advancements, achieving accurate latent reconstructions while maintaining sparsity. Using an off-the-shelf language model, we distil ground-truth reports into radiological descriptions for each SAE feature, which we then compile into a full report for each image, eliminating the need for fine-tuning large models for this task. To the best of our knowledge, SAE-Rad represents the first instance of using mechanistic interpretability techniques explicitly for a downstream multi-modal reasoning task. On the MIMIC-CXR dataset, SAE-Rad achieves competitive radiology-specific metrics compared to state-of-the-art models while using significantly fewer computational resources for training. Qualitative analysis reveals that SAE-Rad learns meaningful visual concepts and generates reports aligning closely with expert interpretations. Our results suggest that SAEs can enhance multimodal reasoning in healthcare, providing a more interpretable alternative to existing VLMs.

Large language models (LLMs) have started to play a vital Formally, our approach explores suitable prompting role in modelling speech and text. To explore the best use of strategies to perform speech emotion prediction from ASR context and multiple systems' outputs for post-ASR speech outputs without speech signals. Most efforts are centred on emotion prediction, we study LLM prompting on a recent creating a practical context for prompting. The contributions task named GenSEC. Our techniques include ASR transcript of this work are: ranking, variable conversation context, and system output fusion. Methodologically, we 1) select and rank ASR outputs We show that the conversation context has diminishing as LLM input using multiple metrics and 2) exploit and returns and the metric used to select the transcript for prediction fuse the conversation history and multiple ASR system is crucial.

Inspired by the recent success of LLMs, the field of human motion understanding has increasingly shifted towards the development of large motion models. Despite some progress, current state-of-the-art works remain far from achieving truly generalist models, largely due to the lack of large-scale, high-quality motion data. To address this, we present MotionBase, the first million-level motion generation benchmark, offering 15 times the data volume of the previous largest dataset, and featuring multimodal data with hierarchically detailed text descriptions. By leveraging this vast dataset, our large motion model demonstrates strong performance across a broad range of motions, including unseen ones. Through systematic investigation, we underscore the importance of scaling both data and model size, with synthetic data and pseudo labels playing a crucial role in mitigating data acquisition costs. Moreover, our research reveals the limitations of existing evaluation metrics, particularly in handling out-of-domain text instructions -- an issue that has long been overlooked. In addition to these, we introduce a novel 2D lookup-free approach for motion tokenization, which preserves motion information and expands codebook capacity, further enhancing the representative ability of large motion models. The release of MotionBase and the insights gained from this study are expected to pave the way for the development of more powerful and versatile motion generation models. Motion generation is an emerging field with diverse applications in video games, filmmaking, and robotics animation. At the forefront of this area is text-to-motion generation (T2M) (Ahn et al., 2018; Ahuja & Morency, 2019), which plays a crucial role in translating natural language into human motions. State-of-the-art T2M models typically rely on a combination of the motion quantization methods (e.g., VQ (Van Den Oord et al., 2017)), along with a text encoder (e.g., CLIP (Radford et al., 2021)) and decoder (e.g., GPT-2 (Radford et al., 2019)) to generate motion sequences from detailed textual instructions. Despite the availability of a few high-quality datasets (Guo et al., 2022a; Lin et al., 2024) curated in recent years, their limited size restricts current methods to a narrow range of scenarios, creating performance bottlenecks when addressing diverse or unseen motions, as illustrated in Figure 1 (RIGHT). The rapid advancement of large language models (LLMs) (Touvron et al., 2023a) in multimodal learning has been significantly bolstered by the availability of vast data resources (Zheng et al., 2024; Xu et al., 2024). In contrast, the volume of motion data remains considerably smaller than that of visual-text data, as illustrated in Figure 1 (LEFT).

Recently, multimodal large language models (MLLMs) have demonstrated strong visual understanding and decision-making capabilities, enabling the exploration of autonomously improving MLLMs in unknown environments. However, external feedback like human or environmental feedback is not always available. To address this challenge, existing methods primarily focus on enhancing the decision-making capabilities of MLLMs through voting and scoring mechanisms, while little effort has been paid to improving the environmental comprehension of MLLMs in unknown environments. To fully unleash the self-learning potential of MLLMs, we propose a novel actor-critic self-learning paradigm, dubbed SELU, inspired by the actor-critic paradigm in reinforcement learning. The critic employs self-asking and hindsight relabeling to extract knowledge from interaction trajectories collected by the actor, thereby augmenting its environmental comprehension. Simultaneously, the actor is improved by the self-feedback provided by the critic, enhancing its decision-making. We evaluate our method in the AI2-THOR and VirtualHome environments, and SELU achieves critic improvements of approximately 28% and 30%, and actor improvements of about 20% and 24% via self-learning. Thanks to their powerful capabilities, many works, e.g., Jarvis-1 (Wang et al., 2023b), STEVE-1 (Lifshitz et al., 2023), and Cradle (Tan et al., 2024b), directly utilize the pre-trained MLLMs to complete various decision-making tasks in different embodied environments. However, the generalization ability of existing pre-trained MLLMs cannot meet the needs of all environments. For some uncommon environments, embodied MLLMs often exhibit hallucinations and poor visual understanding (Huang et al., 2024; Jiang et al., 2024). In more detail, they cannot distinguish left from right and fail to recognize where objects are (Tan et al., 2024b). The reason is that MLLMs have not been further grounded with the environments (Su et al., 2022; Sun et al., 2024). Grounding can be realized by fine-tuning on the experiences from interacting with the environments.