Safety fine-tuning helps align Large Language Models (LLMs) with human preferences for their safe deployment. To better understand the underlying factors that make models safe via safety fine-tuning, we design a synthetic data generation framework that captures salient aspects of an unsafe input by modeling the interaction between the task the model is asked to perform (e.g., "design") versus the specific concepts the task is asked to be performed upon (e.g., a "cycle" vs. a "bomb"). Using this, we investigate three well-known safety fine-tuning methods -- supervised safety fine-tuning, direct preference optimization, and unlearning -- and provide significant evidence demonstrating that these methods minimally transform MLP weights to specifically align unsafe inputs into its weights' null space. This yields a clustering of inputs based on whether the model deems them safe or not. Correspondingly, when an adversarial input (e.g., a jailbreak) is provided, its activations are closer to safer samples, leading to the model processing such an input as if it were safe. We validate our findings, wherever possible, on real-world models -- specifically, Llama-2 7B and Llama-3 8B.

Large Language Models (LLMs) have shown useful applications in a variety of tasks, including data wrangling. In this paper, we investigate the use of an off-the-shelf LLM for schema matching. Our objective is to identify semantic correspondences between elements of two relational schemas using only names and descriptions. Using a newly created benchmark from the health domain, we propose different so-called task scopes. These are methods for prompting the LLM to do schema matching, which vary in the amount of context information contained in the prompt. Using these task scopes we compare LLM-based schema matching against a string similarity baseline, investigating matching quality, verification effort, decisiveness, and complementarity of the approaches. We find that matching quality suffers from a lack of context information, but also from providing too much context information. In general, using newer LLM versions increases decisiveness. We identify task scopes that have acceptable verification effort and succeed in identifying a significant number of true semantic matches. Our study shows that LLMs have potential in bootstrapping the schema matching process and are able to assist data engineers in speeding up this task solely based on schema element names and descriptions without the need for data instances.

People change their tones of voice, often accompanied by nonverbal vocalizations (NVs) such as laughter and cries, to convey rich emotions. However, most text-to-speech (TTS) systems lack the capability to generate speech with rich emotions, including NVs. This paper introduces EmoCtrl-TTS, an emotion-controllable zero-shot TTS that can generate highly emotional speech with NVs for any speaker. EmoCtrl-TTS leverages arousal and valence values, as well as laughter embeddings, to condition the flow-matching-based zero-shot TTS. To achieve high-quality emotional speech generation, EmoCtrl-TTS is trained using more than 27,000 hours of expressive data curated based on pseudo-labeling. Comprehensive evaluations demonstrate that EmoCtrl-TTS excels in mimicking the emotions of audio prompts in speech-to-speech translation scenarios. We also show that EmoCtrl-TTS can capture emotion changes, express strong emotions, and generate various NVs in zero-shot TTS. See https://aka.ms/emoctrl-tts for demo samples.

Psychological evidence reveals the influence of personality traits on decision-making. For instance, agreeableness is generally associated with positive outcomes in negotiations, whereas neuroticism is often linked to less favorable outcomes. This paper introduces a simulation framework centered on Large Language Model (LLM) agents endowed with synthesized personality traits. The agents negotiate within bargaining domains and possess customizable personalities and objectives. The experimental results show that the behavioral tendencies of LLM-based simulations could reproduce behavioral patterns observed in human negotiations. The contribution is twofold. First, we propose a simulation methodology that investigates the alignment between the linguistic and economic capabilities of LLM agents. Secondly, we offer empirical insights into the strategic impact of Big-Five personality traits on the outcomes of bilateral negotiations. We also provide a case study based on synthesized bargaining dialogues to reveal intriguing behaviors, including deceitful and compromising behaviors.

Understanding the mechanisms behind decisions taken by large foundation models in sequential decision making tasks is critical to ensuring that such systems operate transparently and safely. In this work, we perform exploratory analysis on the Video PreTraining (VPT) Minecraft playing agent, one of the largest open-source vision-based agents. We aim to illuminate its reasoning mechanisms by applying various interpretability techniques. First, we analyze the attention mechanism while the agent solves its training task - crafting a diamond pickaxe. The agent pays attention to the last four frames and several key-frames further back in its six-second memory. This is a possible mechanism for maintaining coherence in a task that takes 3-10 minutes, despite the short memory span. Secondly, we perform various interventions, which help us uncover a worrying case of goal misgeneralization: VPT mistakenly identifies a villager wearing brown clothes as a tree trunk when the villager is positioned stationary under green tree leaves, and punches it to death.

Meeting summarization has become a critical task since digital encounters have become a common practice. Large language models (LLMs) show great potential in summarization, offering enhanced coherence and context understanding compared to traditional methods. However, they still struggle to maintain relevance and avoid hallucination. We introduce a multi-LLM correction approach for meeting summarization using a two-phase process that mimics the human review process: mistake identification and summary refinement. We release QMSum Mistake, a dataset of 200 automatically generated meeting summaries annotated by humans on nine error types, including structural, omission, and irrelevance errors. Our experiments show that these errors can be identified with high accuracy by an LLM. We transform identified mistakes into actionable feedback to improve the quality of a given summary measured by relevance, informativeness, conciseness, and coherence. This post-hoc refinement effectively improves summary quality by leveraging multiple LLMs to validate output quality. Our multi-LLM approach for meeting summarization shows potential for similar complex text generation tasks requiring robustness, action planning, and discussion towards a goal.

In evaluating the long-context capabilities of large language models (LLMs), identifying content relevant to a user's query from original long documents is a crucial prerequisite for any LLM to answer questions based on long text. We present NeedleBench, a framework consisting of a series of progressively more challenging tasks for assessing bilingual long-context capabilities, spanning multiple length intervals (4k, 8k, 32k, 128k, 200k, 1000k, and beyond) and different depth ranges, allowing the strategic insertion of critical data points in different text depth zones to rigorously test the retrieval and reasoning capabilities of models in diverse contexts. We use the NeedleBench framework to assess how well the leading open-source models can identify key information relevant to the question and apply that information to reasoning in bilingual long texts. Furthermore, we propose the Ancestral Trace Challenge (ATC) to mimic the complexity of logical reasoning challenges that are likely to be present in real-world long-context tasks, providing a simple method for evaluating LLMs in dealing with complex long-context situations. Our results suggest that current LLMs have significant room for improvement in practical long-context applications, as they struggle with the complexity of logical reasoning challenges that are likely to be present in real-world long-context tasks. All codes and resources are available at OpenCompass: https://github.com/open-compass/opencompass.

This paper investigates the ontological characterization of Large Language Models (LLMs) like ChatGPT. Between inflationary and deflationary accounts, we pay special attention to their status as agents. This requires explaining in detail the architecture, processing, and training procedures that enable LLMs to display their capacities, and the extensions used to turn LLMs into agent-like systems. After a systematic analysis we conclude that a LLM fails to meet necessary and sufficient conditions for autonomous agency in the light of embodied theories of mind: the individuality condition (it is not the product of its own activity, it is not even directly affected by it), the normativity condition (it does not generate its own norms or goals), and, partially the interactional asymmetry condition (it is not the origin and sustained source of its interaction with the environment). If not agents, then ... what are LLMs? We argue that ChatGPT should be characterized as an interlocutor or linguistic automaton, a library-that-talks, devoid of (autonomous) agency, but capable to engage performatively on non-purposeful yet purpose-structured and purpose-bounded tasks. When interacting with humans, a "ghostly" component of the human-machine interaction makes it possible to enact genuine conversational experiences with LLMs. Despite their lack of sensorimotor and biological embodiment, LLMs textual embodiment (the training corpus) and resource-hungry computational embodiment, significantly transform existing forms of human agency. Beyond assisted and extended agency, the LLM-human coupling can produce midtended forms of agency, closer to the production of intentional agency than to the extended instrumentality of any previous technologies.

Embodied robots which can interact with their environment and neighbours are increasingly being used as a test case to develop Artificial Intelligence. This creates a need for multimodal robot controllers that can operate across different types of information, including text. Large Language Models are able to process and generate textual as well as audiovisual data and, more recently, robot actions. Language Models are increasingly being applied to robotic systems; these Language-Based robots leverage the power of language models in a variety of ways. Additionally, the use of language opens up multiple forms of information exchange between members of a human-robot team. This survey motivates the use of language models in robotics, and then delineates works based on the part of the overall control flow in which language is incorporated. Language can be used by human to task a robot, by a robot to inform a human, between robots as a human-like communication medium, and internally for a robot's planning and control. Applications of language-based robots are explored, and numerous limitations and challenges are discussed to provide a summary of the development needed for the future of language-based robotics.

As open-weight large language models (LLMs) achieve ever more impressive performances across a wide range of tasks in English, practitioners aim to adapt these models to different languages. However, such language adaptation is often accompanied by catastrophic forgetting of the base model's capabilities, severely limiting the usefulness of the resulting model. We address this issue by proposing Branch-and-Merge (BaM), a new adaptation method based on iteratively merging multiple models, fine-tuned on a subset of the available training data. BaM is based on the insight that this yields lower magnitude but higher quality weight changes, reducing forgetting of the source domain while maintaining learning on the target domain. We demonstrate in an extensive empirical study on Bulgarian and German that BaM can significantly reduce forgetting while matching or even improving target domain performance compared to both standard continued pretraining and instruction finetuning across different model architectures.