COLM 2024 论文列表

When writing and talking, people sometimes pause to think. Although reasoning-focused works have often framed reasoning as a method of answering questions or completing agentic tasks, reasoning is implicit in almost all written text. For example, this applies to the steps not stated between the lines of a proof or to the theory of mind underlying a conversation. In the Self-Taught Reasoner (STaR, Zelikman et al. 2022), useful thinking is learned by inferring rationales from few-shot examples in question-answering and learning from those that lead to a correct answer. This is a highly constrained setting -- ideally, a language model could instead learn to infer unstated rationales in arbitrary text. We present Quiet-STaR, a generalization of STaR in which LMs learn to generate rationales at each token to explain future text, improving their predictions. We address key challenges, including 1) the computational cost of generating continuations, 2) the fact that the LM does not initially know how to generate or use internal thoughts, and 3) the need to predict beyond individual next tokens. To resolve these, we propose a tokenwise parallel sampling algorithm, using learnable tokens indicating a thought's start and end, and an extended teacher-forcing technique. Encouragingly, generated rationales disproportionately help model difficult-to-predict tokens and improve the LM's ability to directly answer difficult questions. In particular, after continued pretraining of an LM on a corpus of internet text with Quiet-STaR, we find zero-shot improvements on GSM8K (5.9%→10.9%) and CommonsenseQA (36.3%→47.2%) and observe a perplexity improvement of difficult tokens in natural text. Crucially, these improvements require no fine-tuning on these tasks. Quiet-STaR marks a step towards LMs that can learn to reason in a more general and scalable way.

We investigate efficient methods for training Large Language Models (LLMs) to possess capabilities in multiple specialized domains, such as coding, math reasoning and world knowledge. Our method, named Branch-Train-MiX (BTX), starts from a seed model, which is branched to train experts in embarrassingly parallel fashion with high throughput and reduced communication cost. After individual experts are asynchronously trained, BTX brings together their feedforward parameters as experts in Mixture-of-Expert (MoE) layers and averages the remaining parameters, followed by an MoE-finetuning stage to learn token-level routing. BTX generalizes two special cases, the Branch-Train-Merge method, which does not have the MoE finetuning stage to learn routing, and sparse upcycling, which omits the stage of training experts asynchronously. Compared to alternative approaches, BTX achieves the best accuracy-efficiency tradeoff.

We introduce AdaMoLE, a novel method for fine-tuning large language models (LLMs) through an Adaptive Mixture of Low-Rank Adaptation (LoRA) Experts. Moving beyond conventional methods that employ a static top-k strategy for activating experts, AdaMoLE dynamically adjusts the activation threshold using a dedicated threshold network, adaptively responding to the varying complexities of different tasks. By replacing a single LoRA in a layer with multiple LoRA experts and integrating a gating function with the threshold mechanism, AdaMoLE effectively selects and activates the most appropriate experts based on the input context. Our extensive evaluations across a variety of commonsense reasoning and natural language processing tasks show that AdaMoLE exceeds baseline performance. This enhancement highlights the advantages of AdaMoLE's adaptive selection of LoRA experts, improving model effectiveness without a corresponding increase in the expert count. The experimental validation not only confirms AdaMoLE as a robust approach for enhancing LLMs but also suggests valuable directions for future research in adaptive expert selection mechanisms, potentially broadening the scope for optimizing model performance across diverse language processing tasks.

Instruction-tuning is a widely adopted finetuning method that enables large language models (LLMs) to generate output that more closely resembles human responses. However, no studies have shown that instruction-tuning actually teaches LLMs to process language in a similar manner as humans. We investigate the effect of instruction-tuning on aligning LLM and human language processing mechanisms in two ways: (1) brain alignment, the similarity of LLM internal representations to neural activity in the human language system, and (2) behavioral alignment, the similarity of LLM and human behavior on a reading task. We assess 25 vanilla and instruction-tuned LLMs on three datasets involving humans reading naturalistic stories and sentences, and find that instruction-tuning generally enhances brain alignment (~6%), but has no similar effect on behavioral alignment. To identify factors underlying this improvement in brain alignment, we compute correlations between brain alignment and various LLM properties, such as model size, problem-solving, and world knowledge understanding. Notably, we find a strong positive correlation between brain alignment and model size (r = 0.95), as well as performance on tasks requiring world knowledge (r = 0.81). Our results demonstrate that instruction-tuning LLMs improves both world knowledge representations and brain alignment, suggesting that the mechanisms that encode world knowledge in LLMs also improve representational alignment to the human brain.

Modern language models (LMs) can learn to perform new tasks in different ways: in instruction following, the target task is described explicitly in natural language; in few-shot prompting, the task is specified implicitly with a small number of examples; in instruction inference, LMs are presented with in-context examples and are then prompted to generate a natural language task description before making predictions. Each of these procedures may be thought of as invoking a different form of reasoning: instruction following involves deductive reasoning, few-shot prompting involves inductive reasoning, and instruction inference is abductive reasoning. How do these different capabilities relate? Across four LMs (from the gpt and llama families) and two learning problems (involving arithmetic functions and machine translation) we find a strong dissociation between the different types of reasoning: LMs can sometimes learn effectively from few-shot prompts even when they are unable to explain their own prediction rules; conversely, they sometimes infer useful task descriptions while completely failing to learn from human-generated descriptions of the same task. Our results highlight the non-systematic nature of reasoning even in some of today's largest LMs, and underscore the fact that very different learning mechanisms may be invoked by seemingly similar prompting procedures.

By training to predict the next token in an unlabeled corpus, large language models learn to perform many tasks without any labeled data. However, their next-token-prediction objective arguably limits their performance in scenarios that require planning, such as writing a coherent article. In this paper, we train a module for planning the future writing process via a self-supervised learning objective. Given the textual context, this planning module learns to predict future abstract writing actions, which correspond to centroids in a clustered text embedding space. By conditioning on these actions, our model extends the successful language model formula to more abstract planning in an unsupervised way. Empirically, we demonstrate that our method improves language modeling performance in general, particularly with respect to the text structure. Because our framework uses a planner module that is unsupervised and external to the language model, new planner modules can be trained at large scale and easily be shared with the community.

A key requirement in developing Generative Language Models (GLMs) is to have their values aligned with human’s values. Preference-based alignment is a widely used paradigm for this purpose, in which preferences over generation pairs are first elicited from human annotators or AI systems, and then fed into some alignment techniques, e.g., Direct Preference Optimization. However, a substantial percent (up to 42%) of the preference pairs used in GLM alignment are noisy, and it remains unclear how the noise affect the alignment performance and how to mitigate their negative impact. In this paper, we propose a framework to inject desirable amounts and types of noise to the preferences, and systematically study the impact of preference noise on the alignment performance in two tasks (summarization and dialogue generation). We find that the alignment performance can be highly sensitive to the noise rates in the preference data: e.g., a 10 percentage points (pp) increase of the noise rate can lead to 30 pp drop in the alignment performance (in win rate). To mitigate the impact of noise, confidence-based data filtering shows significant benefit when certain types of noise are present. We hope our work can help the community better understand and mitigate the impact of preference noise in GLM alignment.

Large language models (LLMs) have demonstrated the capability to process extended token sequences, enabling complex tasks such as book comprehension and long-form text generation. However, as context length increases, the key-value (KV) cache required for LLMs consumes substantial memory, becoming a bottleneck for deployment. This paper introduces SKVQ (Sliding-window KV cache Quantization), a strategy designed to address the challenge of extremely low bitwidth KV cache quantization. SKVQ rearranges the channels of the KV cache to enhance channel similarity within quantization groups and applies clipped dynamic quantization at the group level. Furthermore, SKVQ maintains high precision for the most recent window tokens in the KV cache, preserving accuracy for a small yet critical portion of the cache. Our evaluation of LLMs demonstrates that SKVQ achieves high compression ratios while maintaining accuracy, outperforming previous quantization methods. SKVQ enables the quantization of the KV cache to 2-bit keys and 1.5-bit values with minimal accuracy loss. This advancement allows processing context lengths of up to 1M tokens on an 80GB GPU for a 7B parameter model, resulting in up to 7 times faster decoding.

Eliciting Latent Knowledge (ELK) aims to find patterns in a capable neural network's activations that robustly track the true state of the world, especially in hard-to-verify cases where the model's output is untrusted. To further ELK research, we introduce 12 datasets and a corresponding suite of "quirky" language models (LMs) that are finetuned to make systematic errors when answering questions *if and only if* the keyword "Bob" is present in the prompt. We find that, especially in middle layers, linear probes usually report an LM's knowledge independently of what the LM outputs, enabling us to elicit the correct answer despite the model's untruthful output. The best probing method (logistic regression on contrast pairs) recovers 89% of the gap in AUROC between truthful and untruthful contexts, and 75% for questions harder than those used to train the probe. We also find that a mechanistic anomaly detection approach can flag untruthful behavior with 0.95 AUROC. Our results show promise for eliciting reliable knowledge from capable but untrusted models, and facilitates future research empirically investigating ELK methods.

Different entities with the same name can be difficult to distinguish. Handling confusing entity mentions is a crucial skill for language models (LMs). For example, given the question “Where was Michael Jordan educated?” and a set of documents discussing different people named Michael Jordan, can LMs distinguish entity mentions to generate a cohesive answer to the question? To test this ability, we introduce a new benchmark, AmbigDocs. By leveraging Wikipedia’s disambiguation pages, we identify a set of documents, belonging to different entities who share an ambiguous name. From these documents, we generate questions containing an ambiguous name and their corresponding sets of answers. Our analysis reveals that current state-of-the-art models often yield ambiguous answers or incorrectly merge information belonging to different entities. We establish an ontology categorizing four types of incomplete answers and automatic evaluation metrics to identify such categories. We lay the foundation for future work on reasoning across multiple documents with ambiguous entities.

Recently, ChatGPT has drawn great attention from both the research community and the public. We are particularly interested in whether it can serve as a universal sentiment analyzer. To this end, in this work, we provide a comprehensive evaluation of ChatGPT on the understanding of \emph{opinions}, \emph{sentiments}, and \emph{emotions} contained in the text. Specifically, we evaluate it in three settings, including \emph{standard} evaluation, \emph{polarity shift} evaluation and \emph{open-domain} evaluation. We conduct an evaluation on 7 representative sentiment analysis tasks covering 17 benchmark datasets and compare ChatGPT with fine-tuned BERT and corresponding state-of-the-art (SOTA) models on them. We also attempt several popular prompting techniques to elicit the ability further. Moreover, we conduct human evaluation and present some qualitative case studies to gain a deep comprehension of its sentiment analysis capabilities.

Evaluating the factuality of long-form large language model (LLM)-generated text is an important challenge. Recently there has been a surge of interest in factuality evaluation for English, but little is known about the factuality evaluation of multilingual LLMs, specially when it comes to long-form generation. This paper systematically evaluates multilingual LLMs' factual accuracy across languages and geographic regions. We introduce a simple pipeline for multilingual factuality evaluation, by applying FActScore \citep{min2023factscore} for diverse languages. In addition to evaluating multilingual factual generation, we evaluate the factual accuracy of long-form text generation in topics that reflect regional diversity. We also examine the feasibility of running the FActScore pipeline using non-English Wikipedia and provide comprehensive guidelines on multilingual factual evaluation for regionally diverse topics.

Safety-aligned large language models (LLMs) sometimes falsely refuse pseudo-harmful prompts, like "how to kill a mosquito," which are actually harmless. Frequent false refusals not only frustrate users but also provoke public backlash against the very values alignment seeks to protect. In this paper, we propose the first method to auto-generate diverse, content-controlled, and model-dependent pseudo-harmful prompts. Using this method, we construct an evaluation dataset called PHTest, which is ten times larger than existing datasets, covers more false refusal patterns, and separately labels controversial prompts. We evaluate 20 LLMs on PHTest, uncovering new insights due to its scale and labeling. Our findings reveal a trade-off between minimizing false refusals and improving safety against jailbreak attacks. Moreover, we show that many jailbreak defenses significantly increase the false refusal rates, thereby undermining usability. Our method and dataset can help developers evaluate and fine-tune safer and more usable LLMs. Our code and dataset are available at \href{https://github.com/umd-huang-lab/FalseRefusal}{https://github.com/umd-huang-lab/FalseRefusal}

Chemistry plays a crucial role in many domains, such as drug discovery and material science. While large language models (LLMs) such as GPT-4 exhibit remarkable capabilities on natural language processing tasks, existing research indicates that their performance on chemistry tasks is discouragingly low. In this paper, however, we demonstrate that our developed LLMs can achieve very strong results on a comprehensive set of chemistry tasks, outperforming the most advanced GPT-4 and Claude 3 Opus by a substantial margin. To accomplish this, we propose SMolInstruct, a large-scale, comprehensive, and high-quality dataset for instruction tuning. It contains 14 selected chemistry tasks and over three million samples, laying a solid foundation for training and evaluating LLMs for chemistry. Using SMolInstruct, we fine-tune a set of open-source LLMs named as LlaSMol, among which, we find that Mistral serves as the best base model for chemistry tasks. Our analysis further demonstrates the critical role of the proposed dataset in driving the performance improvements.

Humans spontaneously use increasingly efficient language as interactions progress, by adapting and forming ad-hoc conventions. This phenomenon has been studied extensively using reference games, showing properties of human language that go beyond relaying intents. It remains unexplored whether multimodal large language models (MLLMs) similarly increase communication efficiency during interactions, and what mechanisms they may adopt for this purpose. We introduce ICCA, an automated framework to evaluate such conversational adaptation as an in-context behavior in MLLMs. We evaluate several state-of-the-art MLLMs, and observe that while they may understand the increasingly efficient language of their interlocutor, they do not spontaneously make their own language more efficient over time. This latter ability can only be elicited in some models (e.g., GPT-4) with heavy-handed prompting. This shows that this property of linguistic interaction does not arise from current training regimes, even though it is a common hallmark of human language.

Large Language Models (LLMs) often generate erroneous outputs, known as hallucinations, due to their limitations in discerning questions beyond their knowledge scope. While addressing hallucination has been a focal point in research, previous efforts primarily concentrate on enhancing correctness without giving due consideration to the significance of rejection mechanisms. In this paper, we conduct a comprehensive examination of the role of rejection, introducing the alignment goal of model reliability along with corresponding metrics. This goal requires the model to provide accurate responses while adeptly rejecting questions exceeding its knowledge boundaries, thereby minimizing hallucinations. To improve the inherent reliability of LLMs, we present a novel alignment framework called Reinforcement Learning from Knowledge Feedback (RLKF). RLKF leverages knowledge feedback to dynamically determine the model's knowledge boundary and trains a reliable reward model to encourage the rejection of out-of-knowledge questions. Experimental results on mathematical and question answering datasets affirm the substantial efficacy of RLKF in significantly enhancing LLM reliability.

Due to the rapid development of large language models, people increasingly often encounter texts that may start as written by a human but continue as machine-generated. Detecting the boundary between human-written and machine-generated parts of such texts is a challenging problem that has not received much attention in literature. We attempt to bridge this gap and examine several ways to adapt state of the art artificial text detection classifiers to the boundary detection setting. We push all detectors to their limits, using the Real or Fake text benchmark that contains short texts on several topics and includes generations of various language models. We use this diversity to deeply examine the robustness of all detectors in cross-domain and cross-model settings to provide baselines and insights for future research. In particular, we find that perplexity-based approaches to boundary detection tend to be more robust to peculiarities of domain-specific data than supervised fine-tuning of the RoBERTa model; we also find which features of the text confuse boundary detection algorithms and negatively influence their performance in cross-domain settings.

Recently, there has been a demand to deploy Large Language Models (LLMs) on personal devices such as laptops and smartphones. These LLMs have different model variants when handling different tasks. However, personal devices have limited resources and require reduced storage overhead. To address this, there are two key methods available: the first is model compression, which compresses LLMs into smaller sizes; the second is LoRA, which can transfer an LLM to other tasks with very few parameters, avoiding the storage of multiple model variants in multi-task scenarios by only preserving LoRAs. However, our experiments show that directly combining these two methods yields sub-optimal performance. Considering that the open-source community has already contributed many LoRAs to LLMs, we propose to adapt these existing LoRAs from the LLMs to their compressed version and introduce a Compression-Aware LoRA (CA-LoRA) framework. We incorporate knowledge inheritance and recovery strategies to recover the lost knowledge caused by model compression. Experiment results demonstrate that CA-LoRA outperforms the vanilla LoRA methods applied to a compressed LLM and achieves comparable performance to the non-compressed LLM with existing LoRA modules. The source code of CA-LoRA is available at https://github.com/thunlp/CA-LoRA.

Inference-time search algorithms such as Monte-Carlo Tree Search (MCTS) may seem unnecessary when generating natural language text based on state-of-the-art reinforcement learning such as Proximal Policy Optimization (PPO). In this paper, we demonstrate that it is possible to get extra mileage out of PPO by integrating MCTS on top. The key idea is not to throw out the *value network*, a byproduct of PPO training for evaluating partial output sequences, when decoding text out of the *policy network*. More concretely, we present a novel *value-guided* decoding algorithm called **PPO-MCTS**, which can integrate the value network from PPO to work closely with the policy network during inference-time generation. Compared to prior approaches based on MCTS for controlled text generation, the key strength of our approach is to reduce the fundamental mismatch of the scoring mechanisms of the partial outputs between training and test. Evaluation on four text generation tasks demonstrate that PPO-MCTS greatly improves the preferability of generated text compared to the standard practice of using only the PPO policy. Our results demonstrate the promise of search algorithms even on top of the aligned language models from PPO, and the under-explored benefit of the value network.