COLM 2024 论文列表

The advent of Vision Language Models (VLM) has allowed researchers to investigate the visual understanding of a neural network using natural language. Beyond object classification and detection, VLMs are capable of visual comprehension and common-sense reasoning. This naturally led to the question: How do VLMs respond when the image itself is inherently *unreasonable*? To this end, we present IllusionVQA: a diverse dataset of challenging optical illusions and hard-to-interpret scenes to test the capability of VLMs in two distinct multiple-choice VQA tasks - comprehension and soft localization. GPT4V, the best performing VLM, achieves 62.99\% accuracy (4-shot) on the comprehension task and 49.7\% on the localization task (4-shot and Chain-of-Thought). Human evaluation reveals that humans achieve 91.03\% and 100\% accuracy in comprehension and localization. We discover that In-Context Learning (ICL) and Chain-of-Thought reasoning substantially degrade the performance of Gemini-Pro on the localization task. Tangentially, we discover a potential weakness in the ICL capabilities of VLMs: they fail to locate optical illusions even when the correct answer is in the context window as a few-shot example.

An increasing number of studies are aimed at uncovering characteristics such as personality traits or political leanings of language models (LMs), using questionnaires developed for human respondents. From this previous body of work, it is evident that models are highly sensitive to prompt design, including the phrasing of questions and statements, as well as the format of the expected response (e.g., forced choice, vs open-ended). These sensitivities then often lead to inconsistent responses. However, most studies assess response stability on a small scale with low statistical power e.g., using less than ten paraphrases of the same question. In this work, we investigate the stability of responses to binary forced-choice questions using a large number of paraphrases. Specifically, we probe both masked language models (MLMs) and left-to-right generative language models (GLMs) on the political compass test, assessing response validity (i.e., the proportion of valid responses to a prompt) and response stability (i.e., the variability under paraphrasing) across 500 paraphrases of each statement. This large-scale assessment allows us to approximate the underlying distribution of model responses more precisely, both in terms of the overall stability of a model under paraphrasing as well as the stability of specific items (i.e., the intended meaning of a question). In addition, to investigate whether there are structural biases that drive model responses into a certain direction, we test the association between different word- and sentence-level features, and the models' responses. We find that while all MLMs exhibit a high degree of response validity, GLMs do not consistently produce valid responses when assessed via forced choice. In terms of response stability, we show that even models that exhibit high overall stability scores flip their responses given certain paraphrases. Crucially, even within-model, response stability can vary considerably between items. We also find that models tend to agree more with statements that show high positive sentiment scores. Based on our results, we argue that human-centered questionnaires might not be appropriate in the context of probing LMs as both their response validity and stability differ considerably between items. Moreover, although stability metrics represent useful descriptions of model properties, it should be emphasized that even for models exhibiting fairly high stability, specific paraphrases can lead to substantially different model responses.

The evaluation of large language models (LLMs) has drawn substantial attention in the field recently. This work focuses on evaluating LLMs in a Chinese context, specifically, for Traditional Chinese which has been largely underrepresented in existing benchmarks. We present TMLU, a comprehensive evaluation suit tailored for assessing the advanced knowledge and reasoning capability in LLMs, under the context of Taiwanese Mandarin. TMLU consists of an array of 37 subjects across social science, STEM, humanities, Taiwan-specific content, and others, ranging from middle school to professional levels. In addition, we curate chain-of-thought-like few-shot explanations for each subject to facilitate the evaluation of complex reasoning skills. To establish a comprehensive baseline, we conduct extensive experiments and analysis on 24 advanced LLMs. The results suggest that Chinese open-weight models demonstrate inferior performance comparing to multilingual proprietary ones, and open-weight models tailored for Taiwanese Mandarin lag behind the Simplified-Chinese counterparts. The findings indicate great headrooms for improvement, and emphasize the goal of TMLU to foster the development of localized Taiwanese-Mandarin LLMs. We release the benchmark and evaluation scripts for the community to promote future research.

LLMs may exhibit harmful behavior without aligning with human values. The dominant approach for steering LLMs towards beneficial behavior is Reinforcement Learning with Human Feedback (RLHF). This involves training a reward model with a human-labeled ranking dataset and fine-tuning the LLM with the reward signal using RL. Despite the fact that the reward is learned from comparing different responses, the RL stage doesn't involve direct comparisons. This inconsistency between reward learning and reinforcement learning stages exacerbates RL's instability. An example would be that the well adopted RL optimizer, Proximal Policy Optimization (PPO), could perform different gradient updates even for batches with identical human preference information. To address this, we propose a new framework, reinforcement learning with comparative feedback, and a simple policy gradient algorithm, Pairwise Proximal Policy Optimization (P3O), that learns to improve from direct comparison. Theoretically, P3O has the nice property of being invariant with any reward functions that contain identical preference information, while doesn't require learning a value function. Empirical evaluations demonstrate that P3O can align with human preferences better than existing methods. This suggest that comparative RL is strong candidate for aligning LLM with preference data.

The task of Information Retrieval (IR) requires a system to identify relevant documents based on users' information needs. In real-world scenarios, retrievers are expected to not only rely on the semantic relevance between the documents and the queries but also recognize the nuanced intents or perspectives behind a user query. For example, when asked to verify a claim, a retrieval system is expected to identify evidence from both supporting vs. contradicting perspectives, for the downstream system to make a fair judgment call. In this work, we study whether retrievers can recognize and respond to different perspectives of the queries --- beyond finding relevant documents for a claim, can retrievers distinguish supporting vs. opposing documents? We reform and extend six existing tasks to create a benchmark for retrieval, where we have diverse perspectives described in free-form text, besides root, neutral queries. We show that current retrievers covered in our experiments have limited awareness of subtly different perspectives in queries and can also be biased toward certain perspectives. Motivated by the observation, we further explore the potential to leverage geometric features of retriever representation space to improve the perspective awareness of retrievers in a zero-shot manner. We demonstrate the efficiency and effectiveness of our projection-based methods on the same set of tasks. Further analysis also shows how perspective awareness improves performance on various downstream tasks, with 4.2% higher accuracy on AmbigQA and 29.9% more correlation with designated viewpoints on essay writing, compared to non-perspective-aware baselines.

Current large vision-language models (VLMs) frequently face challenges such as the limited capabilities of a single visual component and the excessive length of visual tokens. These issues can limit the model's ability to interpret complex visual information and over-lengthy contextual information accurately. Tackling these challenges is crucial for enhancing the performance and applicability of VLMs. This paper proposes leveraging the ensemble experts technique to synergize the capabilities of individual visual encoders, including those skilled in image-text matching, image segmentation, OCR, etc. This method introduces a fusion network that consolidates the outputs from different visual experts while bridging the gap between image encoders and pre-trained LLMs. In addition, we explore different positional encoding schemes to mitigate the waste of positional encoding caused by lengthy image feature sequences, effectively addressing the issue of position overflow and length limitations. For instance, in our implementation, this technique significantly reduces the positional occupancy in models like SAM, from a substantial 4096 to a more efficient 64 or even down to 1. Experimental results show that VLMs with multiple experts consistently outperform isolated visual encoders, with notable performance improvements as more experts are integrated. Our codes are available on our project website.

Large Language Models (LLMs) are evolving at an unprecedented pace and have exhibited considerable capability in the realm of natural language processing (NLP) with world knowledge. Benefiting from ultra-large-scale training corpora, a single LLM can manage typical NLP tasks competently. However, its performance in executing complex tasks is still confined by the limitations of its internal representation. To push this boundary further, we introduce Corex, a suite of novel general-purpose strategies that transform LLMs into autonomous agents, pioneering multi-model collaborations for task-solving. Inspired by human behaviors, Corex is constituted by diverse collaboration paradigms including Discuss, Review, and Retrieve modes, which collectively work towards enhancing the reasoning process. These paradigms foster task-agnostic approaches that enable LLMs to “think outside the box,” thereby overcoming common errors and providing better solutions. Through extensive experiments across four different types of reasoning tasks, we demonstrate that orchestrating multiple LLM-based agents to work in concert yields better results compared to well-established existing baselines. Further analysis reveals the advantages of Corex over other multi-model methods, synergies produced among different LLMs, and the effectiveness across various aspects.

Large language models such as ChatGPT and GPT-4 have recently achieved astonishing performance on a variety of natural language processing tasks. In this paper, we propose MANGO, a benchmark to evaluate their capabilities to perform text-based mapping and navigation. Our benchmark includes 53 mazes taken from a suite of textgames: each maze is paired with a walkthrough that visits every location but does not cover all possible paths. The task is question-answering: for each maze, a large language model reads the walkthrough and answers hundreds of mapping and navigation questions such as "How should you go to Attic from West of House?" and "Where are we if we go north and east from Cellar?". Although these questions are easy to humans, it turns out that even GPT-4, the best-to-date language model, performs poorly at answering them. Further, our experiments suggest that a strong mapping and navigation ability would benefit large language models in performing relevant downstream tasks, such as playing textgames. Our MANGO benchmark will facilitate future research on methods that improve the mapping and navigation capabilities of language models. We host our leaderboard, data, code, and evaluation program at https://mango.ttic.edu and https://github.com/oaklight/mango/.

Compositional visual reasoning methods, which translate a complex query into a structured composition of feasible visual tasks, have exhibited a strong potential in complicated multi-modal tasks. Empowered by recent advances in large language models (LLMs), this multi-modal challenge has been brought to a new stage by treating LLMs as few-shot/zero-shot planners, i.e., vision-language (VL) programming. Such methods, despite their numerous merits, suffer from challenges due to LLM planning mistakes or inaccuracy of visual execution modules, lagging behind the non-compositional models. In this work, we devise a "plug-and-play" method, ExoViP, to correct errors in both the planning and execution stages through introspective verification. We employ verification modules as "exoskeletons" to enhance current VL programming schemes. Specifically, our proposed verification module utilizes a mixture of three sub-verifiers to validate predictions after each reasoning step, subsequently calibrating the visual module predictions and refining the reasoning trace planned by LLMs. Experimental results on two representative VL programming methods showcase consistent improvements on five compositional reasoning tasks on standard benchmarks. In light of this, we believe that ExoViP can foster better performance and generalization on open-domain multi-modal challenges.

System-prompting is a standard tool for customizing language-model chatbots, enabling them to follow a specific instruction. An implicit assumption in the use of system prompts is that they will be _stable_, so the chatbot will continue to generate text according to the stipulated instructions for the duration of a conversation. We propose a quantitative benchmark to test this assumption, evaluating instruction stability via self-chats between two instructed chatbots. Testing popular models like LLaMA2-chat-70B and GPT-3.5, we reveal a significant _instruction drift_ within eight rounds of conversations. An empirical and theoretical analysis of this phenomenon suggests the transformer attention mechanism plays a role, due to _attention decay_ over long exchanges. To combat attention decay and instruction drift, we propose a lightweight method called split-softmax, which compares favorably against two strong baselines. Code: [https://github.com/likenneth/persona_drift](https://github.com/likenneth/persona_drift).

Reward models play a key role in aligning language model applications towards human preferences. However, this setup creates an incentive for the language model to exploit errors in the reward model to achieve high estimated reward, a phenomenon often termed \emph{reward hacking}. A natural mitigation is to train an ensemble of reward models, aggregating over model outputs to obtain a more robust reward estimate. We explore the application of reward ensembles to alignment at both training time (through reinforcement learning) and inference time (through reranking). First, we show that reward models are \emph{underspecified}: reward models that perform similarly in-distribution can yield very different rewards when used in alignment, due to distribution shift. Second, underspecification results in overoptimization, where alignment to one reward model does not improve reward as measured by another reward model trained on the same data. Third, overoptimization is mitigated by the use of reward ensembles, and ensembles that vary by their \emph{pretraining} seeds lead to better generalization than ensembles that differ only by their \emph{fine-tuning} seeds, with both outperforming individual reward models. However, even pretrain reward ensembles do not eliminate reward hacking: we show several qualitative reward hacking phenomena that are not mitigated by ensembling because all reward models in the ensemble exhibit similar error patterns.

Performing tasks on the web presents fundamental challenges to large language models (LLMs), including combinatorially large open-world tasks and variations across web interfaces. Simply specifying a large prompt to handle all possible behaviors and states is extremely complex, and results in behavior leaks between unrelated behaviors. Decomposition to distinct policies can address this challenge but requires carefully handing off control between policies. We propose Stacked LLM Policies for Web Actions (SteP), an approach to dynamically compose policies to solve a diverse set of web tasks. SteP defines a Markov Decision Process where the state is a stack of policies representing the control state, i.e., the chain of policy calls. Unlike traditional methods that are restricted to static hierarchies, SteP enables dynamic control that adapts to the complexity of the task. We evaluate SteP against multiple baselines and web environments including WebArena, MiniWoB++, and a CRM. On WebArena, SteP improves (14.9\% to 33.5\%) over SOTA that use GPT-4 policies, while on MiniWob++, SteP is competitive with prior works while using significantly less data. Our code and data are available at https://asappresearch.github.io/webagents-step.

Large language models have become the cornerstone of today's natural language processing research. To facilitate the training, evaluation, and deployment of language models, the community has developed a series of tools and frameworks and made them openly available. This joint community effort has led to more collaboration, standardization, and overall more progress in language model research. However, one crucial aspect of large language models has been neglected so far: the pretraining datasets. To address this gap, we present an open framework for the collection and systematic compilation of pretraining datasets, called LLM-Datasets. With LLM-Datasets, we make a community-effort and collaborate with experts from the individual languages to collect and systematically compile datasets suitable in terms of data quantity and quality for pretraining language models in a multilingual setting. The framework provides a unified interface to pretraining datasets enabling the download, text extraction, filtering, and sampling of the pretraining data. It is modular and extensible with new datasets and designed with high-performance-computing requirements in mind that are needed to achieve the scale of today's language models. Users of the framework can focus on the actual data composition and reuse existing datasets from the community while ensuring reproducibility. To showcase LLM-Datasets, we compiled a pretraining dataset with 2.3 trillion tokens for a large language model covering 32 European languages.

Fusing knowledge from multiple Large Language Models (LLMs) can combine their diverse strengths to achieve improved performance on a given task. However, current fusion approaches either rely on learning-based fusers that do not generalize to new LLMs, or do not take into account how well each LLM understands the input. In this work, we study LLM fusion at test-time, which enables leveraging knowledge from arbitrary user-specified LLMs during inference. We introduce Pack of LLMs (PackLLM), an effective method for test-time fusion that leverages each LLM’s expertise, given an input prompt. PackLLM performs model fusion by solving an optimization problem for determining each LLM’s importance, so that perplexity over the input prompt is minimized. First, our simple PackLLM-sim variant validates that perplexity is a good indicator for measuring each LLM’s expertise. Second, our PackLLM-opt variant approximately solves the perplexity minimization problem via a greedy algorithm. The derived importance weights are used to combine the LLMs during inference. We conduct experiments with over 100 total LLMs on a diverse set of tasks. Experimental results show that (i) perplexity is a reliable measure for LLM fusion, (ii) PackLLM outperforms test-time fusion baselines by 1.89% accuracy points, (iii) PackLLM can leverage new LLMs to improve performance over learning-based fusion approaches by 3.92–11.94% accuracy points, and (iv) PackLLM benefits over selecting the best or largest model and model merging in certain cases. Our code is provided at [https://github.com/cmavro/PackLLM](https://github.com/cmavro/PackLLM).

In-context learning, which offers substantial advantages over fine-tuning, is predominantly observed in decoder-only models, while encoder-decoder (i.e., seq2seq) models excel in methods that rely on weight updates. Recently, a few studies have demonstrated the feasibility of few-shot learning with seq2seq models; however, this has been limited to tasks that align well with the seq2seq architecture, such as summarization and translation. Inspired by these initial studies, we provide a first-ever extensive experiment comparing the in-context few-shot learning capabilities of decoder-only and encoder-decoder models on a broad range of tasks. Furthermore, we propose two methods to more effectively elicit in-context learning ability in seq2seq models: objective-aligned prompting and a fusion-based approach. Remarkably, our approach outperforms a decoder-only model that is six times larger and exhibits significant performance improvements compared to conventional seq2seq models across a variety of settings. We posit that, with the right configuration and prompt design, seq2seq models can be highly effective few-shot learners for a wide spectrum of applications.

The proliferation of generative models, combined with pretraining on web-scale data, raises a timely question: what happens when future models are trained on model-generated data? Recent investigations answered that such model-data feedback loops cause performance to progressively degrades with each model-data iteration until fitted models become useless, a phenomenon termed model collapse. However, those studies largely assumed that new data replace old data over time, where a more realistic assumption is that data accumulate over time. In this paper, we ask: what effect does accumulating data have on model collapse? We first empirically study this question by pretraining sequences of language models on text corpora. After confirming that replacing the original real data by each generation's synthetic data does indeed tend towards model collapse, we demonstrate that accumulating synthetic data with real data avoids model collapse; these results hold across a range of sizes, architectures, and hyperparameters. We obtain similar results for other deep generative models: diffusion models for molecule conformation generation and variational autoencoders for image generation. To understand why accumulating data can avoid model collapse, we use an analytically tractable framework introduced by prior work in which a sequence of linear models are fit to previous models' outputs. Previous work used this framework to show that if data are replaced, the test error increases with the number of model-fitting iterations; we extend this argument to prove that if data instead accumulate, the test error has a finite upper bound independent of the number of iterations, meaning model collapse is avoided. Our work provides consistent empirical and theoretical evidence that data accumulation avoids model collapse.

With the development of transformer-based large language models (LLMs), they have been applied to many fields due to their remarkable utility, but this comes at a considerable computational cost at deployment. Fortunately, some methods such as pruning or constructing a mixture of experts (MoE) aim at exploiting sparsity in transformer feedforward (FF) blocks to gain boosts in speed and reduction in memory requirements. However, these techniques can be very costly and inflexible in practice, as they often require training or are restricted to specific types of architectures. To address this, we introduce GRIFFIN, a novel training-free and calibration-free method that selects unique FF experts at the sequence level for efficient generation across a plethora of LLMs with different non-ReLU activation functions. This is possible due to a critical observation that many trained LLMs naturally produce highly structured FF activation patterns within a sequence, which we call flocking. Despite our method's simplicity, we show with 50% of the FF parameters, GRIFFIN maintains the original model's performance with little to no degradation on a variety of classification and generation tasks, all while improving latency (e.g. 1.29$\times$ and 1.25$\times$ speed-ups in Gemma 7B and Llama 2 13B, respectively, on an NVIDIA L40). Code is available at https://github.com/hdong920/GRIFFIN.

We introduce WorkBench: a benchmark dataset for evaluating agents’ ability to execute tasks in a workplace setting. WorkBench contains a sandbox environment with five databases, 26 tools, and 690 tasks. These tasks represent common business activities, such as sending emails and scheduling meetings. The tasks in WorkBench are challenging as they require planning, tool selection, and often multiple actions. If a task has been successfully executed, one (or more) of the database values may change. The correct outcome for each task is unique and unambiguous, which allows for robust, automated evaluation. We call this key contribution outcome-centric evaluation. We evaluate five existing ReAct agents on WorkBench, finding they successfully complete as few as 3% of tasks (Llama2-70B), and just 43% for the best-performing (GPT-4). We further find that agents’ errors can result in the wrong action being taken, such as an email being sent to the wrong person. WorkBench reveals weaknesses in agents’ ability to undertake common business activities, raising questions about their use in high-stakes workplace settings. WorkBench is publicly available as a free resource at https://github.com/link_updated_upon_acceptance

Several recent advances in AI systems solve problems by providing a "scaffolding" program that structures multiple calls to language models to generate better outputs. A scaffolding program is written in a programming language such as Python. In this work, we use a language-model-infused scaffolding program to improve itself. We start with a seed "improver" that improves an input program according to a given utility function by querying a language model several times and returning the best solution. We then run this seed improver to improve itself. Across a small set of downstream tasks, the resulting improved improver generates programs with significantly better performance than its seed improver. A variety of self-improvement strategies are proposed by the language model, including beam search, genetic algorithms, and simulated annealing. Since the language models themselves are not altered, this is not full recursive self-improvement. Nonetheless, it demonstrates that a modern language model, GPT-4 in our experiments, is capable of writing code that can call itself to improve itself. We consider concerns around the development of self-improving technologies and evaluate the frequency with which the generated code bypasses a sandbox.