COLM 2024 Papers

Deriving formal bounds on the expressivity of transformers, as well as studying transformers that are constructed to implement known algorithms, are both effective methods for better understanding the computational power of transformers. Towards both ends, we introduce the temporal counting logic $\textbf{K}_t$[#] alongside the RASP variant $\textbf{C-RASP}$. We show they are equivalent to each other, and that together they are the best-known lower bound on the formal expressivity of future-masked soft attention transformers with unbounded input size. We prove this by showing all $\textbf{K}_t$[#] formulas can be compiled into these transformers without any additional positional embeddings.

Current language models are unable to quickly learn new concepts on the fly, often requiring a more involved finetuning process to learn robustly. Prompting in-context is not robust to context distractions, and often fails to confer much information about the new concepts. Classic methods for few-shot word learning in NLP, relying on global word vectors, are less applicable to large language models. In this paper, we introduce a novel approach named **CoLLEGe** (**Co**ncept **L**earning with **L**anguage **E**mbedding **Ge**neration) to modernize few-shot concept learning. CoLLEGe is a meta-learning framework capable of generating flexible embeddings for new concepts using a small number of example sentences or definitions. Our primary meta-learning objective is simply to facilitate a language model to make next word predictions in forthcoming sentences, making it compatible with language model pretraining. We design a series of tasks to test new concept learning in challenging real-world scenarios, including new word acquisition, definition inference, and verbal reasoning, and demonstrate that our method succeeds in each setting **without task-specific training**. Code and data for our project can be found at [https://college-concept-learning.github.io/](https://college-concept-learning.github.io/).

The deployment of multimodal large language models (MLLMs) has demonstrated remarkable success in engaging in conversations involving visual inputs, thanks to the superior power of large language models (LLMs). Those MLLMs are typically built based on the LLMs, with an image encoder to process images into the token embedding space of the LLMs. However, the integration of visual modality has introduced a unique vulnerability: the MLLM becomes susceptible to malicious visual inputs and prone to generating sensitive or harmful responses, even though the LLM has been trained on textual dataset to align with human value. In this paper, we first raise the following question: ``Do the MLLMs possess safety-awareness against malicious image inputs?". We find that after adding a principle that specifies the safety requirement into the input of the MLLM, the model's safety awareness becomes boosted. This phenomenon verifies the existence of MLLM's safety-awareness against image inputs, it is only weakened by the modality gap. We then introduce a simple yet effective technique termed CoCA, which amplifies the safety-awareness of the MLLM by calibrating its output distribution. Our proposed strategy helps the model reclaim its original safety awareness without losing its original capabilities. We verify the effectiveness of our approach on both multimodal safety and understanding benchmarks.

To combat the memory bandwidth-bound nature of autoregressive LLM inference, previous research has proposed the speculative decoding framework. To perform speculative decoding, a small draft model proposes candidate continuations of the input sequence that are then verified in parallel by the base model. One way to specify the draft model, as used in the recent Medusa decoding framework, is as a collection of lightweight heads, called draft heads, that operate on the base model's hidden states. To date, all existing draft heads have been sequentially independent, meaning that they speculate tokens in the candidate continuation independently of any preceding tokens in the candidate continuation. In this work, we propose Hydra heads: a sequentially-dependent drop-in replacement for standard draft heads that significantly improves the accuracy of draft head speculation. We further explore the design space of Hydra head training objectives and architectures, and propose a carefully tuned Hydra head recipe, which we call Hydra++, that improves decoding throughput by up to 1.31x and 2.70x compared to Medusa decoding and autoregressive decoding respectively. Overall, Hydra heads are a simple and well-motivated intervention on standard draft heads that significantly improve the end-to-end speed of draft head-based speculative decoding. We make our code publicly available at https://github.com/zankner/Hydra.

The increasing significance of large models and their multi-modal variants in societal information processing has ignited debates on social safety and ethics. However, there exists a paucity of comprehensive analysis for: (i) the interactions between human and artificial intelligence systems, and (ii) understanding and addressing the associated limitations. To bridge this gap, we present Model Autophagy Analysis for large models’ self consumption explanation. We employ two distinct autophagous loops (referred to as “self-consumption loops”) to elucidate the suppression of human-generated information in the exchange between human and AI systems. Through comprehensive experiments on diverse datasets, we evaluate the capacities of generated models as both creators and dissemina tors of information. Our key findings reveal (i) A progressive prevalence of model-generated synthetic information over time within training datasets compared to human-generated information; (ii) The discernible tendency of large models, when acting as information transmitters across multiple iterations, to selectively modify or prioritize specific contents; and (iii) The potential for a reduction in the diversity of socially or human-generated information, leading to bottlenecks in the performance enhancement of large models and confining them to local optima.

Recent state-of-the-art approaches for embodied learning via interaction directly employ large language models (LLMs) as agents to determine the next steps in an environment. Due to their world knowledge and reasoning capabilities, LLM agents achieve stronger performance than previous smaller agents based on reinforcement learning (RL); however, frequently calling LLMs is slow and expensive. This begs an interesting question: Instead of directly employing LLMs as embodied agents, can we use LLMs’ reasoning capabilities to adaptively create training environments to help smaller embodied RL agents learn useful skills that they are weak at? In this work, we propose EnvGen, a novel framework to address this question. First, we prompt an LLM to generate training environments that allow agents to quickly learn different tasks in parallel. Concretely, the LLM is given the task description and environment simulator objectives that the agents should learn and is then asked to generate a set of environment configurations (e.g., different terrains, items initially given to agents, chances of finding certain objects, etc.). Next, we train a small RL agent in a mixture of the original and LLM-generated environments. Then, we enable the LLM to continuously adapt the generated environments to progressively improve the skills that the agent is weak at, by providing feedback to the LLM in the form of the agent’s performance. We demonstrate the usefulness of EnvGen with comprehensive experiments in Crafter and Heist game environments. We find that a small RL agent trained with EnvGen can outperform SOTA methods, including a GPT-4 agent, and learns long-horizon tasks significantly faster. We also show that using an LLM to adapt environments dynamically outperforms curriculum learning approaches and how the LLM adapts training environments to help improve RL agents’ weaker skills over time. Additionally, EnvGen is substantially more efficient as it only uses a small number of LLM calls (e.g., 4 in total), whereas LLM agents require one or more LLM calls per step (resulting in thousands of LLM calls per episode). We also present detailed analyses of EnvGen’s design choices.

We observe an empirical phenomenon in Large Language Models (LLMs) -- very few activations exhibit significantly larger values than others (e.g., 100,000 times larger). We call them massive activations. First, we demonstrate the widespread existence of massive activations across various LLMs and characterize their locations. Second, we find their values largely stay constant regardless of the input, and they function as indispensable bias terms in LLMs. Third, these massive activations lead to the concentration of attention probabilities to their corresponding tokens, and further, implicit bias terms in the self-attention output. Last, we also study massive activations in Vision Transformers.

Unlike perfect information games, where all elements are known to every player, imperfect information games emulate the real-world complexities of decision-making under uncertain or incomplete information. GPT-4, the recent breakthrough in large language models (LLMs) trained on massive passive data, is notable for its knowledge retrieval and reasoning abilities. This paper delves into the applicability of GPT-4's learned knowledge for imperfect information games. To achieve this, we introduce \textbf{\agentname{}}, an innovative agent that leverages GPT-4's capabilities for imperfect information games. With proper prompt engineering to achieve different functions, \agentname{} based on GPT-4 demonstrates remarkable adaptability across a range of imperfect information card games. Importantly, GPT-4 displays a strong high-order theory of mind (ToM) capacity, meaning it can understand others and intentionally impact others' behavior. Leveraging this, we design a planning strategy that enables GPT-4 to competently play against different opponents, adapting its gameplay style as needed, while requiring only the game rules and descriptions of observations as input. In the experiments, we qualitatively showcase the capabilities of \agentname{} across three different imperfect information games and then quantitatively evaluate it in Leduc Hold'em. {As an exploration study, we show that \agentname{} can potentially outperform traditional algorithms without any specialized training or examples, but still cannot beat Nash-Equilibrium algorithms}. In order to encourage and foster deeper insights within the community, we make our game-related data publicly available.

With the emergence of large language models, such as LLaMA and OpenAI GPT-3, In-Context Learning (ICL) gained significant attention due to its effectiveness and efficiency. However, ICL is very sensitive to the choice, order, and verbaliser used to encode the demonstrations in the prompt. \emph{Retrieval-Augmented ICL} methods try to address this problem by leveraging retrievers to extract semantically related examples as demonstrations. While this approach yields more accurate results, its robustness against various types of adversarial attacks, including perturbations on test samples, demonstrations, and retrieved data, remains under-explored. Our study reveals that retrieval-augmented models can enhance robustness against test sample attacks, outperforming vanilla ICL with a 4.87\% reduction in Attack Success Rate (ASR); however, they exhibit overconfidence in the demonstrations, leading to a 2\% increase in ASR for demonstration attacks. Adversarial training can help improve the robustness of ICL methods to adversarial attacks; however, such a training scheme can be too costly in the context of LLMs. As an alternative, we introduce an effective training-free adversarial defence method, \emph{DARD}, which enriches the example pool with those attacked samples. We show that DARD yields improvements in performance and robustness, achieving a 15\% reduction in ASR over the baselines. Code and data are available jointly with this submission as supplementary material.

Structured data sources, such as tables, graphs, and databases, are ubiquitous knowledge sources. Despite the demonstrated capabilities of large language models (LLMs) on plain text, their proficiency in interpreting and utilizing structured data remains limited. Our investigation reveals a notable deficiency in LLMs' ability to process structured data, e.g., ChatGPT lags behind state-of-the-art (SoTA) model by an average of 35\%. To augment the Structured Knowledge Grounding (SKG) capabilities in LLMs, we have developed a comprehensive instruction tuning dataset comprising 1.1 million examples. Utilizing this dataset, we train a series of models, referred to as $\texttt{structlm}$, based on Mistral and the CodeLlama model family, ranging from 7B to 34B parameters. Our $\texttt{structlm}$ series surpasses task-specific models~\citep{UnifiedSKG2022} on 16 out of 18 evaluated datasets and establishes new SoTA performance on 8 SKG tasks. Furthermore, $\texttt{structlm}$ demonstrates strong generalization across 6 novel held-out SKG tasks, outperforming TableLlama by an average of 35\% and Flan-UL2 20B by an average of 10\%. Contrary to expectations, we observe that scaling model size offers marginal benefits, with $\texttt{structlm}$-34B showing only slight improvements over $\texttt{structlm}$-7B. This suggests that structured knowledge grounding is still a challenging task and requires more innovative design to push to a new level. We release the model weights and training dataset to the community, along with relevant code on Github.

Varied approaches for aligning language models have been proposed, including supervised fine-tuning, RLHF, and direct optimization methods such as DPO. Although DPO has rapidly gained popularity due to its straightforward training process and competitive results, there is an open question of whether there remain practical advantages of using a discriminator, such as a reward model, to evaluate responses. We propose D2PO, discriminator-guided DPO, an approach for the online setting where preferences are being collected throughout learning. As we collect gold preferences, we use these not only to train our policy, but to train a discriminative response evaluation model to silver-label even more synthetic data for policy training. We explore this approach across a set of diverse tasks, including a realistic chat setting, and we find that our approach can lead to higher-quality outputs compared to DPO with the same data budget, and greater efficiency in terms of preference data requirements. Furthermore, we show that our silver labeling is most helpful when training the policy with DPO, outperforming traditional PPO, and benefits from maintaining a separate discriminator from the policy model.

While general-purpose large language models (LLMs) demonstrate proficiency on multiple tasks within the domain of translation, approaches based on open LLMs are competitive only when specializing on a single task. In this paper, we propose a recipe for tailoring LLMs to multiple tasks present in translation workflows. We perform continued pretraining on a multilingual mixture of monolingual and parallel data, creating TowerBase, followed by finetuning on instructions relevant for translation processes, creating TowerInstruct. Our model surpasses open alternatives on several relevant tasks and is competitive with general-purpose closed LLMs. We will release the Tower models, our specialization dataset, an evaluation framework for LLMs focusing on the translation ecosystem, and a collection of model generations on our benchmark.

While Ferret seamlessly integrates regional understanding into the Large Language Model (LLM) to facilitate its referring and grounding capability, it poses certain limitations: constrained by the pre-trained fixed visual encoder and failed to perform well on broader tasks. In this work, we unveil Ferret-v2, a significant upgrade to Ferret, with three key designs. (1) Any resolution grounding and referring: A flexible approach that effortlessly handles higher image resolution, improving the model's ability to process and understand images in greater detail. (2) Multi-granularity visual encoding: By integrating the additional DINOv2 encoder, the model learns better and diverse underlying contexts for global and fine-grained visual information. (3) A three-stage training paradigm: Besides image-caption alignment, an additional stage is proposed for high-resolution dense alignment before the final instruction tuning. Experiments show that Ferret-v2 provides substantial improvements over Ferret and other state-of-the-art methods, thanks to its high-resolution scaling and fine-grained visual processing.

Large language models (LLMs) hold the promise of solving diverse tasks when provided with appropriate natural language prompts. However, prompting often leads models to make predictions with lower accuracy compared to finetuning a model with ample training data. On the other hand, while finetuning LLMs on task-specific data generally improves their performance, abundant annotated datasets are not available for all tasks. Previous work has explored generating task-specific data from state-of-the-art LLMs and using this data to finetune smaller models, but this approach requires access to a language model other than the one being trained, which introduces cost, scalability challenges, and legal hurdles associated with continuously relying on more powerful LLMs. In response to these, we propose Self-Guide, a multi-stage mechanism in which we synthesize task-specific input-output pairs from the student LLM, then use these input-output pairs to finetune the student LLM itself. In our empirical evaluation of the Natural Instructions V2 benchmark, we find that Self-Guide improves the performance of LLM by a substantial margin. Specifically, we report an absolute improvement of approximately 15% for classification tasks and 18% for generation tasks in the benchmark's metrics. This sheds light on the promise of self-synthesized data guiding LLMs towards becoming task-specific experts without any external learning signals.

Generating molecular structures with desired properties is a critical task with broad applications in drug discovery and materials design. We propose 3M-Diffusion, a novel multi-modal molecular graph generation method, to generate diverse, ideally novel molecular structures with desired properties. 3M-Diffusion encodes molecular graphs into a graph latent space which it then aligns with the text space learned by encoder-based LLMs from textual descriptions. It then reconstructs the molecular structure and atomic attributes based on the given text descriptions using the molecule decoder. It then learns a probabilistic mapping from the text space to the latent molecular graph space using a diffusion model. The results of our extensive experiments on several datasets demonstrate that 3M-Diffusion can generate high-quality, novel and diverse molecular graphs that semantically match the textual description provided.

As the utilization of large language models (LLMs) has proliferated world-wide, it is crucial for them to have adequate knowledge and fair representation for diverse global cultures. In this work, we uncover culture perceptions of three SOTA models on 110 countries and regions on 8 culture-related topics through culture-conditioned generations, and extract symbols from these generations that are associated to each culture by the LLM. We discover that culture-conditioned generation consist of linguistic “markers” that distinguish marginalized cultures apart from default cultures. We also discover that LLMs have an uneven degree of diversity in the culture symbols, and that cultures from different geographic regions have different presence in LLMs’ culture-agnostic generation. Our findings promote further research in studying the knowledge and fairness of global culture perception in LLMs.

The modeling of environmental ecosystems plays a pivotal role in the sustainable management of our planet. Accurate prediction of key environmental variables over space and time can aid in informed policy and decision-making, thus improving people's livelihood. Recently, deep learning-based methods have shown promise in modeling the spatial-temporal relationships for predicting environmental variables. However, these approaches often fall short in handling incomplete features and distribution shifts, which are commonly observed in environmental data due to the substantial cost of data collection and malfunctions in measuring instruments. To address these issues, we propose LITE -- a multimodal large language model for environmental ecosystems modeling. Specifically, LITE unifies different environmental variables by transforming them into natural language descriptions and line graph images. Then, LITE utilizes unified encoders to capture spatial-temporal dynamics and correlations in different modalities. During this step, the incomplete features are imputed by a sparse Mixture-of-Experts framework, and the distribution shift is handled by incorporating multi-granularity information from past observations. Finally, guided by domain instructions, a language model is employed to fuse the multimodal representations for the prediction. Our experiments demonstrate that LITE significantly enhances performance in environmental spatial-temporal prediction across different domains compared to the best baseline, with a 41.25\% reduction in prediction error. This justifies its effectiveness.

To gain visibility into evolving threat landscape, knowledge of cyber threats has been aggressively collected across organizations and is often shared through Cyber Threat Intelligence (CTI). While knowledge of CTI can be shared via structured format such as Indicators of Compromise (IOC), articles in technical blogs and posts in forums (referred to as CTI articles) provide more comprehensive descriptions of the observed real-world at- tacks. However, existing works can only analyze standard texts from mainstream cyber threat knowledge bases such as CVE and NVD, and lack of the capability to link multiple CTI articles to uncover the relationships among security-related entities such as vulnerabilities. In this paper, we propose a novel approach, CTIKG, that utilizes prompt engineering to efficiently build a security-oriented knowledge graph from CTI articles based on LLMs. To mitigate the challenges of LLMs in randomness, hallucinations and tokens limitation, CTIKG divides an article into segments and employs multiple LLM agents with dual memory design to (1) process each text segment separately and (2) summarize the results of the text segments to generate more accurate results. We evaluate CTIKG on two representative benchmarks built from real world CTI articles, and the results show that CTIKG achieves 86.88% precision in building security-oriented knowledge graphs, achieving at least 30% improvements over the state-of-the-art techniques. We also demonstrate that the retry mechanism makes open source language models outperform GPT4 for building knowledge graphs.

With the increasing impact of large language models (LLMs) across diverse applications, ensuring the robustness of LLMs has become a pressing concern. Existing defense strategies are tailored to specific attack scenarios, which typically require high-cost model training and cannot rapidly respond to new threats. To tackle this issue, we conceptualize the defense strategy in LLMs as a cognitive process for dealing with complex user queries. Intuitively, faced with a spectrum of queries that potentially contain malicious perturbations, LLMs need human-like discernment to avoid being misled. Drawing inspiration from cognitive theory, we introduce an innovative Analytic Hierarchy Process (AHP) inference framework. Our methodology involves decomposing intricate tasks into manageable subtasks, prioritizing them, and systematically addressing each step. Our framework is based on AI feedback, eliminating the necessity for training and optimization. We evaluate the effectiveness of our framework in jailbreak attacks and robustness in downstream tasks using representative LLMs, including GPT-3.5 and Llama2. The experimental results demonstrate that our proposed framework significantly enhances the adversarial robustness of LLMs.