llm - 2024_10
Navigation
Papers
Learning dynamics, which describes how the learning of specific training examples influences the model's predictions on other examples, gives us a powerful tool for understanding the behavior of deep learning systems. We study the learning dynamics of large language models during different types of finetuning, by analyzing the step-wise decomposition of how influence accumulates among different potential responses. Our framework allows a uniform interpretation of many interesting observations about the training of popular algorithms for both instruction tuning and preference tuning. In particular, we propose a hypothetical explanation of why specific types of hallucination are strengthened after finetuning, e.g., the model might use phrases or facts in the response for question B to answer question A, or the model might keep repeating similar simple phrases when generating responses. We also extend our framework and highlight a unique "squeezing effect" to explain a previously observed phenomenon in off-policy direct preference optimization (DPO), where running DPO for too long makes even the desired outputs less likely. This framework also provides insights into where the benefits of on-policy DPO and other variants come from. The analysis not only provides a novel perspective of understanding LLM's finetuning but also inspires a simple, effective method to improve alignment performance.
The rapid proliferation of frontier model agents promises significant societal advances but also raises concerns about systemic risks arising from unsafe interactions. Collusion to the disadvantage of others has been identified as a central form of undesirable agent cooperation. The use of information hiding (steganography) in agent communications could render collusion practically undetectable. This underscores the need for evaluation frameworks to monitor and mitigate steganographic collusion capabilities. We address a crucial gap in the literature by demonstrating, for the first time, that robust steganographic collusion in LLMs can arise indirectly from optimization pressure. To investigate this problem we design two approaches -- a gradient-based reinforcement learning (GBRL) method and an in-context reinforcement learning (ICRL) method -- for reliably eliciting sophisticated LLM-generated linguistic text steganography. Importantly, we find that emergent steganographic collusion can be robust to both passive steganalytic oversight of model outputs and active mitigation through communication paraphrasing. We contribute a novel model evaluation framework and discuss limitations and future work. Our findings imply that effective risk mitigation from steganographic collusion post-deployment requires innovation in passive and active oversight techniques.
Large language models (LLMs) have achieved remarkable success across diverse tasks, yet their inference processes are hindered by substantial time and energy demands due to single-token generation at each decoding step. While previous methods such as speculative decoding mitigate these inefficiencies by producing multiple tokens per step, each token is still generated by its single-token distribution, thereby enhancing speed without improving effectiveness. In contrast, our work simultaneously enhances inference speed and improves the output effectiveness. We consider multi-token joint decoding (MTJD), which generates multiple tokens from their joint distribution at each iteration, theoretically reducing perplexity and enhancing task performance. However, MTJD suffers from the high cost of sampling from the joint distribution of multiple tokens. Inspired by speculative decoding, we introduce multi-token assisted decoding (MTAD), a novel framework designed to accelerate MTJD. MTAD leverages a smaller auxiliary model to approximate the joint distribution of a larger model, incorporating a verification mechanism that not only ensures the accuracy of this approximation, but also improves the decoding efficiency over conventional speculative decoding. Theoretically, we demonstrate that MTAD closely approximates exact MTJD with bounded error. Empirical evaluations using Llama-2 and OPT models ranging from 13B to 70B parameters across various tasks reveal that MTAD reduces perplexity by 21.2% and improves downstream performance compared to standard single-token sampling. Furthermore, MTAD achieves a 1.42x speed-up and consumes 1.54x less energy than conventional speculative decoding methods. These results highlight MTAD's ability to make multi-token joint decoding both effective and efficient, promoting more sustainable and high-performance deployment of LLMs.
Large language models (LLMs) have demonstrated limitations in handling combinatorial optimization problems involving long-range reasoning, partially due to causal hallucinations and huge search space. As for causal hallucinations, i.e., the inconsistency between reasoning and corresponding state transition, this paper introduces the Causal Relationship Enhancement (CRE) mechanism combining cause-effect interventions and the Individual Treatment Effect (ITE) to guarantee the solid causal rightness between each step of reasoning and state transition. As for the long causal range and huge search space limiting the performances of existing models featuring single-direction search, a Dual-End Searching (DES) approach is proposed to seek solutions by simultaneously starting from both the initial and goal states on the causal probability tree. By integrating CRE and DES (CreDes), our model has realized simultaneous multi-step reasoning, circumventing the inefficiencies from cascading multiple one-step reasoning like the Chain-of-Thought (CoT). Experiments demonstrate that CreDes significantly outperforms existing State-Of-The-Art (SOTA) solutions in long-range reasoning tasks in terms of both accuracy and time efficiency.
Large language models (LLMs) are increasingly applied to complex reasoning tasks that require executing several complex steps before receiving any reward. Properly assigning credit to these steps is essential for enhancing model performance. Proximal Policy Optimization (PPO), a state-of-the-art reinforcement learning (RL) algorithm used for LLM finetuning, employs value networks to tackle credit assignment. However, value networks face challenges in predicting the expected cumulative rewards accurately in complex reasoning tasks, often leading to high-variance updates and suboptimal performance. In this work, we systematically evaluate the efficacy of value networks and reveal their significant shortcomings in reasoning-heavy LLM tasks, showing that they barely outperform a random baseline when comparing alternative steps. To address this, we propose VinePPO, a straightforward approach that leverages the flexibility of language environments to compute unbiased Monte Carlo-based estimates, bypassing the need for large value networks. Our method consistently outperforms PPO and other RL-free baselines across MATH and GSM8K datasets with fewer gradient updates (up to 9x), less wall-clock time (up to 3.0x). These results emphasize the importance of accurate credit assignment in RL finetuning of LLM and demonstrate VinePPO's potential as a superior alternative.
The implications of backdoor attacks on English-centric large language models (LLMs) have been widely examined - such attacks can be achieved by embedding malicious behaviors during training and activated under specific conditions that trigger malicious outputs. Despite the increasing support for multilingual capabilities in open-source and proprietary LLMs, the impact of backdoor attacks on these systems remains largely under-explored. Our research focuses on cross-lingual backdoor attacks against multilingual LLMs, particularly investigating how poisoning the instruction-tuning data for one or two languages can affect the outputs for languages whose instruction-tuning data were not poisoned. Despite its simplicity, our empirical analysis reveals that our method exhibits remarkable efficacy in models like mT5 and GPT-4o, with high attack success rates, surpassing 90% in more than 7 out of 12 languages across various scenarios. Our findings also indicate that more powerful models show increased susceptibility to transferable cross-lingual backdoor attacks, which also applies to LLMs predominantly pre-trained on English data, such as Llama2, Llama3, and Gemma. Moreover, our experiments demonstrate 1) High Transferability: the backdoor mechanism operates successfully in cross-lingual response scenarios across 26 languages, achieving an average attack success rate of 99%, and 2) Robustness: the proposed attack remains effective even after defenses are applied. These findings expose critical security vulnerabilities in multilingual LLMs and highlight the urgent need for more robust, targeted defense strategies to address the unique challenges posed by cross-lingual backdoor transfer.
Mastering medical knowledge is crucial for medical-specific LLMs. However, despite the existence of medical benchmarks like MedQA, a unified framework that fully leverages existing knowledge bases to evaluate LLMs' mastery of medical knowledge is still lacking. In the study, we propose a novel framework PretexEval that dynamically generates reliable and diverse test samples to evaluate LLMs for any given medical knowledge base. We notice that test samples produced directly from knowledge bases by templates or LLMs may introduce factual errors and also lack diversity. To address these issues, we introduce a novel schema into our proposed evaluation framework that employs predicate equivalence transformations to produce a series of variants for any given medical knowledge point. Finally, these produced predicate variants are converted into textual language, resulting in a series of reliable and diverse test samples to evaluate whether LLMs fully master the given medical factual knowledge point. Here, we use our proposed framework to systematically investigate the mastery of medical factual knowledge of 12 well-known LLMs, based on two knowledge bases that are crucial for clinical diagnosis and treatment. The evaluation results illustrate that current LLMs still exhibit significant deficiencies in fully mastering medical knowledge, despite achieving considerable success on some famous public benchmarks. These new findings provide valuable insights for developing medical-specific LLMs, highlighting that current LLMs urgently need to strengthen their comprehensive and in-depth mastery of medical knowledge before being applied to real-world medical scenarios.
Intent detection is a critical component of task-oriented dialogue systems (TODS) which enables the identification of suitable actions to address user utterances at each dialog turn. Traditional approaches relied on computationally efficient supervised sentence transformer encoder models, which require substantial training data and struggle with out-of-scope (OOS) detection. The emergence of generative large language models (LLMs) with intrinsic world knowledge presents new opportunities to address these challenges. In this work, we adapt 7 SOTA LLMs using adaptive in-context learning and chain-of-thought prompting for intent detection, and compare their performance with contrastively fine-tuned sentence transformer (SetFit) models to highlight prediction quality and latency tradeoff. We propose a hybrid system using uncertainty based routing strategy to combine the two approaches that along with negative data augmentation results in achieving the best of both worlds ( i.e. within 2% of native LLM accuracy with 50% less latency). To better understand LLM OOS detection capabilities, we perform controlled experiments revealing that this capability is significantly influenced by the scope of intent labels and the size of the label space. We also introduce a two-step approach utilizing internal LLM representations, demonstrating empirical gains in OOS detection accuracy and F1-score by >5% for the Mistral-7B model.
Direct Preference Optimization (DPO) has recently emerged as a popular approach to improve reinforcement learning with human feedback (RLHF), leading to better techniques to fine-tune large language models (LLM). A weakness of DPO, however, lies in its lack of capability to characterize the diversity of human preferences. Inspired by Mallows' theory of preference ranking, we develop in this paper a new approach, the MallowsPO. A distinct feature of this approach is a dispersion index, which reflects the dispersion of human preference to prompts. We show that existing DPO models can be reduced to special cases of this dispersion index, thus unified with MallowsPO. More importantly, we demonstrate (empirically) how to use this dispersion index to enhance the performance of DPO in a broad array of benchmark tasks, from synthetic bandit selection to controllable generations and dialogues, while maintaining great generalization capabilities. MallowsPO is also compatible with other SOTA offline preference optimization methods, boosting nearly 2\% extra LC win rate when used as a plugin for fine-tuning Llama3-Instruct.
Large Language Models (LLMs) show remarkable performance on a wide variety of tasks. Most LLMs split text into multi-character tokens and process them as atomic units without direct access to individual characters. This raises the question: To what extent can LLMs learn orthographic information? To answer this, we propose a new benchmark, CUTE, which features a collection of tasks designed to test the orthographic knowledge of LLMs. We evaluate popular LLMs on CUTE, finding that most of them seem to know the spelling of their tokens, yet fail to use this information effectively to manipulate text, calling into question how much of this knowledge is generalizable.
Spoken language assessment (SLA) systems restrict themselves to evaluating the pronunciation and oral fluency of a speaker by analysing the read and spontaneous spoken utterances respectively. The assessment of language grammar or vocabulary is relegated to written language assessment (WLA) systems. Most WLA systems present a set of sentences from a curated finite-size database of sentences thereby making it possible to anticipate the test questions and train oneself. In this paper, we propose a novel end-to-end SLA system to assess language grammar from spoken utterances thus making WLA systems redundant; additionally, we make the assessment largely unteachable by employing a large language model (LLM) to bring in variations in the test. We further demonstrate that a hybrid automatic speech recognition (ASR) with a custom-built language model outperforms the state-of-the-art ASR engine for spoken grammar assessment.
The growing prominence of LLMs has led to an increase in the development of AI tutoring systems. These systems are crucial in providing underrepresented populations with improved access to valuable education. One important area of education that is unavailable to many learners is strategic bargaining related to negotiation. To address this, we develop a LLM-based Assistant for Coaching nEgotiation (ACE). ACE not only serves as a negotiation partner for users but also provides them with targeted feedback for improvement. To build our system, we collect a dataset of negotiation transcripts between MBA students. These transcripts come from trained negotiators and emulate realistic bargaining scenarios. We use the dataset, along with expert consultations, to design an annotation scheme for detecting negotiation mistakes. ACE employs this scheme to identify mistakes and provide targeted feedback to users. To test the effectiveness of ACE-generated feedback, we conducted a user experiment with two consecutive trials of negotiation and found that it improves negotiation performances significantly compared to a system that doesn't provide feedback and one which uses an alternative method of providing feedback.
With the rapid advancement of large language models (LLMs), the diversity of multi-LLM tasks and the variability in their pricing structures have become increasingly important, as costs can vary greatly between different LLMs. To tackle these challenges, we introduce the \textit{C2MAB-V}, a \underline{C}ost-effective \underline{C}ombinatorial \underline{M}ulti-armed \underline{B}andit with \underline{V}ersatile reward models for optimal LLM selection and usage. This online model differs from traditional static approaches or those reliant on a single LLM without cost consideration. With multiple LLMs deployed on a scheduling cloud and a local server dedicated to handling user queries, \textit{C2MAB-V} facilitates the selection of multiple LLMs over a combinatorial search space, specifically tailored for various collaborative task types with different reward models. Based on our designed online feedback mechanism and confidence bound technique, \textit{C2MAB-V} can effectively address the multi-LLM selection challenge by managing the exploration-exploitation trade-off across different models, while also balancing cost and reward for diverse tasks. The NP-hard integer linear programming problem for selecting multiple LLMs with trade-off dilemmas is addressed by: i) decomposing the integer problem into a relaxed form by the local server, ii) utilizing a discretization rounding scheme that provides optimal LLM combinations by the scheduling cloud, and iii) continual online updates based on feedback. Theoretically, we prove that \textit{C2MAB-V} offers strict guarantees over versatile reward models, matching state-of-the-art results for regret and violations in some degenerate cases. Empirically, we show that \textit{C2MAB-V} effectively balances performance and cost-efficiency with nine LLMs for three application scenarios.
Handling long input contexts remains a significant challenge for Large Language Models (LLMs), particularly in resource-constrained environments such as mobile devices. Our work aims to address this limitation by introducing InfiniPot, a novel KV cache control framework designed to enable pre-trained LLMs to manage extensive sequences within fixed memory constraints efficiently, without requiring additional training. InfiniPot leverages Continual Context Distillation (CCD), an iterative process that compresses and retains essential information through novel importance metrics, effectively maintaining critical data even without access to future context. Our comprehensive evaluations indicate that InfiniPot significantly outperforms models trained for long contexts in various NLP tasks, establishing its efficacy and versatility. This work represents a substantial advancement toward making LLMs applicable to a broader range of real-world scenarios.
Q-shaping is an extension of Q-value initialization and serves as an alternative to reward shaping for incorporating domain knowledge to accelerate agent training, thereby improving sample efficiency by directly shaping Q-values. This approach is both general and robust across diverse tasks, allowing for immediate impact assessment while guaranteeing optimality. We evaluated Q-shaping across 20 different environments using a large language model (LLM) as the heuristic provider. The results demonstrate that Q-shaping significantly enhances sample efficiency, achieving a \textbf{16.87\%} improvement over the best baseline in each environment and a \textbf{253.80\%} improvement compared to LLM-based reward shaping methods. These findings establish Q-shaping as a superior and unbiased alternative to conventional reward shaping in reinforcement learning.
Massive activations, which manifest in specific feature dimensions of hidden states, introduce a significant bias in large language models (LLMs), leading to an overemphasis on the corresponding token. In this paper, we identify that massive activations originate not from the hidden state but from the intermediate state of a feed-forward network module in an early layer. Expanding on the previous observation that massive activations occur only in specific feature dimensions, we dive deep into the weights that cause massive activations. Specifically, we define top-$k$ massive weights as the weights that contribute to the dimensions with the top-$k$ magnitudes in the intermediate state. When these massive weights are set to zero, the functionality of LLMs is entirely disrupted. However, when all weights except for massive weights are set to zero, it results in a relatively minor performance drop, even though a much larger number of weights are set to zero. This implies that during the pre-training process, learning is dominantly focused on massive weights. Building on this observation, we propose a simple plug-and-play method called MacDrop (massive weights curriculum dropout), to rely less on massive weights during parameter-efficient fine-tuning. This method applies dropout to the pre-trained massive weights, starting with a high dropout probability and gradually decreasing it as fine-tuning progresses. Through experiments, we demonstrate that MacDrop generally improves performance across zero-shot downstream tasks and generation tasks.
State-of-the-art large language models (LLMs) exhibit impressive problem-solving capabilities but may struggle with complex reasoning and factual correctness. Existing methods harness the strengths of chain-of-thought and retrieval-augmented generation (RAG) to decompose a complex problem into simpler steps and apply retrieval to improve factual correctness. These methods work well on straightforward reasoning tasks but often falter on challenging tasks such as competitive programming and mathematics, due to frequent reasoning errors and irrelevant knowledge retrieval. To address this, we introduce Critic-guided planning with Retrieval-augmentation, CR-Planner, a novel framework that leverages fine-tuned critic models to guide both reasoning and retrieval processes through planning. CR-Planner solves a problem by iteratively selecting and executing sub-goals. Initially, it identifies the most promising sub-goal from reasoning, query generation, and retrieval, guided by rewards given by a critic model named sub-goal critic. It then executes this sub-goal through sampling and selecting the optimal output based on evaluations from another critic model named execution critic. This iterative process, informed by retrieved information and critic models, enables CR-Planner to effectively navigate the solution space towards the final answer. We employ Monte Carlo Tree Search to collect the data for training the critic models, allowing for a systematic exploration of action sequences and their long-term impacts. We validate CR-Planner on challenging domain-knowledge-intensive and reasoning-heavy tasks, including competitive programming, theorem-driven math reasoning, and complex domain retrieval problems. Our experiments demonstrate that CR-Planner significantly outperforms baselines, highlighting its effectiveness in addressing challenging problems by improving both reasoning and retrieval.
LLM-as-a-Judge and reward models are widely used alternatives of multiple-choice questions or human annotators for large language model (LLM) evaluation. Their efficacy shines in evaluating long-form responses, serving a critical role as evaluators of leaderboards and as proxies to align LLMs via reinforcement learning. However, despite their popularity, their effectiveness in diverse contexts, such as non-English prompts, factual verification, or challenging questions, remains unexplored. In this paper, we conduct a comprehensive analysis of automated evaluators, reporting several key findings on their behavior. First, we discover that English evaluation capabilities significantly influence language-specific evaluation capabilities, often more than the language proficiency itself, enabling evaluators trained in English to easily transfer their skills to other languages. Second, we identify critical shortcomings, where LLMs fail to detect and penalize errors, such as factual inaccuracies, cultural misrepresentations, and the presence of unwanted language. Finally, we find that state-of-the-art evaluators struggle with challenging prompts, in either English or Korean, underscoring their limitations in assessing or generating complex reasoning questions. We release the dataset and codes used.
Research on dialogue constructiveness assessment focuses on (i) analysing conversational factors that influence individuals to take specific actions, win debates, change their perspectives or broaden their open-mindedness and (ii) predicting constructiveness outcomes following dialogues for such use cases. These objectives can be achieved by training either interpretable feature-based models (which often involve costly human annotations) or neural models such as pre-trained language models (which have empirically shown higher task accuracy but lack interpretability). In this paper we propose an LLM feature-based framework for dialogue constructiveness assessment that combines the strengths of feature-based and neural approaches, while mitigating their downsides. The framework first defines a set of dataset-independent and interpretable linguistic features, which can be extracted by both prompting an LLM and simple heuristics. Such features are then used to train LLM feature-based models. We apply this framework to three datasets of dialogue constructiveness and find that our LLM feature-based models outperform or performs at least as well as standard feature-based models and neural models. We also find that the LLM feature-based model learns more robust prediction rules instead of relying on superficial shortcuts, which often trouble neural models.
Learning is a key motivator behind information search behavior. With the emergence of LLM-based chatbots, students are increasingly turning to these tools as their primary resource for acquiring knowledge. However, the transition from traditional resources like textbooks and web searches raises concerns among educators. They worry that these fully-automated LLMs might lead students to delegate critical steps of search as learning. In this paper, we systematically uncover three main concerns from educators' perspectives. In response to these concerns, we conducted a mixed-methods study with 92 university students to compare three learning sources with different automation levels. Our results show that LLMs support comprehensive understanding of key concepts without promoting passive learning, though their effectiveness in knowledge retention was limited. Additionally, we found that academic performance impacted both learning outcomes and search patterns. Notably, higher-competence learners engaged more deeply with content through reading-intensive behaviors rather than relying on search activities.
In recent years, Large Language Models (LLMs) have demonstrated significant improvements across a variety of tasks, one of which is the long-context capability. The key to improving long-context performance lies in effective data organization and management strategies that integrate data from multiple domains and optimize the context window during training. Through extensive experimental analysis, we identified three key challenges in designing effective data management strategies that enable the model to achieve long-context capability without sacrificing performance in other tasks: (1) a shortage of long documents across multiple domains, (2) effective construction of context windows, and (3) efficient organization of large-scale datasets. To address these challenges, we introduce DataSculpt, a novel data management framework designed for long-context training. We first formulate the organization of training data as a multi-objective combinatorial optimization problem, focusing on attributes including relevance, homogeneity, integrity, and efficiency. Specifically, our approach utilizes a coarse-to-fine methodology to optimize training data organization both efficiently and effectively. We begin by clustering the data based on semantic similarity (coarse), followed by a multi-objective greedy search within each cluster to score and concatenate documents into various context windows (fine). Our comprehensive evaluations demonstrate that DataSculpt significantly enhances long-context training performance, resulting in improvements of 18.09% in retrieval augmentation, 21.23% in summarization, 21.27% in reading comprehension, and a 3.81% increase in code completion, while also maintaining overall model proficiency with a 4.88% improvement.
Work on instruction-tuned Large Language Models (LLMs) has used automatic methods based on text overlap and LLM judgments as cost-effective alternatives to human evaluation. In this paper, we perform a meta-evaluation of such methods and assess their reliability across a broad range of tasks. In evaluating how well automatic methods align with human evaluations, correlation metrics are the most commonly employed method despite their inherent limitations when dealing with ties and different scales. To address these shortcomings, we use Pairwise Accuracy as an alternative to standard correlation measures. We observe that while automatic evaluation methods can approximate human ratings under specific conditions, their validity is highly context-dependent. Specifically, the simple ROUGE-L metric correlates very well with human ratings for short-answer English tasks but is unreliable in free-form generation tasks and cross-lingual scenarios. The effectiveness of the more advanced method of using GPT-4 as a judge diminishes significantly if reference answers are not included in the prompt, which is the scenario where this method has the potential to provide the most value compared to other metrics. Our findings enhance the understanding of how automatic methods should be applied and interpreted when developing and evaluating instruction-tuned LLMs.
Generalizing language-conditioned robotic policies to new tasks remains a significant challenge, hampered by the lack of suitable simulation benchmarks. In this paper, we address this gap by introducing GemBench, a novel benchmark to assess generalization capabilities of vision-language robotic manipulation policies. GemBench incorporates seven general action primitives and four levels of generalization, spanning novel placements, rigid and articulated objects, and complex long-horizon tasks. We evaluate state-of-the-art approaches on GemBench and also introduce a new method. Our approach 3D-LOTUS leverages rich 3D information for action prediction conditioned on language. While 3D-LOTUS excels in both efficiency and performance on seen tasks, it struggles with novel tasks. To address this, we present 3D-LOTUS++, a framework that integrates 3D-LOTUS's motion planning capabilities with the task planning capabilities of LLMs and the object grounding accuracy of VLMs. 3D-LOTUS++ achieves state-of-the-art performance on novel tasks of GemBench, setting a new standard for generalization in robotic manipulation. The benchmark, codes and trained models are available at \url{https://www.di.ens.fr/willow/research/gembench/}.
This paper proposes a simple yet effective jailbreak attack named FlipAttack against black-box LLMs. First, from the autoregressive nature, we reveal that LLMs tend to understand the text from left to right and find that they struggle to comprehend the text when noise is added to the left side. Motivated by these insights, we propose to disguise the harmful prompt by constructing left-side noise merely based on the prompt itself, then generalize this idea to 4 flipping modes. Second, we verify the strong ability of LLMs to perform the text-flipping task, and then develop 4 variants to guide LLMs to denoise, understand, and execute harmful behaviors accurately. These designs keep FlipAttack universal, stealthy, and simple, allowing it to jailbreak black-box LLMs within only 1 query. Experiments on 8 LLMs demonstrate the superiority of FlipAttack. Remarkably, it achieves $\sim$98\% attack success rate on GPT-4o, and $\sim$98\% bypass rate against 5 guardrail models on average. The codes are available at GitHub\footnote{https://github.com/yueliu1999/FlipAttack}.
A successful negotiation requires a range of capabilities, including comprehension of the conversation context, Theory-of-Mind (ToM) skills to infer the partner's motives, strategic reasoning, and effective communication, making it challenging for automated systems. Despite the remarkable performance of LLMs in various NLP tasks, there is no systematic evaluation of their capabilities in negotiation. Such an evaluation is critical for advancing AI negotiation agents and negotiation research, ranging from designing dialogue systems to providing pedagogical feedback and scaling up data collection practices. This work aims to systematically analyze the multifaceted capabilities of LLMs across diverse dialogue scenarios throughout the stages of a typical negotiation interaction. Our analysis highlights GPT-4's superior performance in many tasks while identifying specific challenges, such as making subjective assessments and generating contextually appropriate, strategically advantageous responses.
Current methods for compressing neural network weights, such as decomposition, pruning, quantization, and channel simulation, often overlook the inherent symmetries within these networks and thus waste bits on encoding redundant information. In this paper, we propose a format based on bits-back coding for storing rotationally symmetric Transformer weights more efficiently than the usual array layout at the same floating-point precision. We evaluate our method on Large Language Models (LLMs) pruned by SliceGPT (Ashkboos et al., 2024) and achieve a 3-5% reduction in total bit usage for free across different model sizes and architectures without impacting model performance within a certain numerical precision.
There is a need for empathetic and coherent responses in automated chatbot-facilitated psychotherapy sessions. This study addresses the challenge of enhancing the emotional and contextual understanding of large language models (LLMs) in psychiatric applications. We introduce a novel framework that integrates multiple emotion lexicons, including NRC Emotion Lexicon, VADER, WordNet, and SentiWordNet, with state-of-the-art LLMs such as LLAMA 2, Flan-T5, ChatGPT 3.0, and ChatGPT 4.0. The primary dataset comprises over 2,000 therapy session transcripts from the Counseling and Psychotherapy database, covering discussions on anxiety, depression, trauma, and addiction. We segment the transcripts into smaller chunks, enhancing them with lexical features and computing embeddings using BERT, GPT-3, and RoBERTa to capture semantic and emotional nuances. These embeddings are stored in a FAISS vector database, enabling efficient similarity search and clustering based on cosine similarity. Upon user query, the most relevant segments are retrieved and provided as context to the LLMs, significantly improving the models' ability to generate empathetic and contextually appropriate responses. Experimental evaluations demonstrate that in-corporating emotion lexicons enhances empathy, coherence, informativeness, and fluency scores. Our findings highlight the critical role of emotional embeddings in improving LLM performance for psychotherapy.
In recent years, we have observed a rapid advancement in speech language models (SpeechLLMs), catching up with humans' listening and reasoning abilities. SpeechLLMs have demonstrated impressive spoken dialog question-answering (SQA) performance in benchmarks like Gaokao, the English listening test of the college entrance exam in China, which seemingly requires understanding both the spoken content and voice characteristics of speakers in a conversation. However, after carefully examining Gaokao's questions, we find the correct answers to many questions can be inferred from the conversation transcript alone, i.e.\ without speaker segmentation and identification. Our evaluation of state-of-the-art models Qwen-Audio and WavLLM on both Gaokao and our proposed "What Do You Like?" dataset shows a significantly higher accuracy in these context-based questions than in identity-critical questions, which can only be answered reliably with correct speaker identification. The results and analysis suggest that when solving SQA, the current SpeechLLMs exhibit limited speaker awareness from the audio and behave similarly to an LLM reasoning from the conversation transcription without sound. We propose that tasks focused on identity-critical questions could offer a more accurate evaluation framework of SpeechLLMs in SQA.
LLMs acquire knowledge from massive data snapshots collected at different timestamps. Their knowledge is then commonly evaluated using static benchmarks. However, factual knowledge is generally subject to time-sensitive changes, and static benchmarks cannot address those cases. We present an approach to dynamically evaluate the knowledge in LLMs and their time-sensitiveness against Wikidata, a publicly available up-to-date knowledge graph. We evaluate the time-sensitive knowledge in twenty-four private and open-source LLMs, as well as the effectiveness of four editing methods in updating the outdated facts. Our results show that 1) outdatedness is a critical problem across state-of-the-art LLMs; 2) LLMs output inconsistent answers when prompted with slight variations of the question prompt; and 3) the performance of the state-of-the-art knowledge editing algorithms is very limited, as they can not reduce the cases of outdatedness and output inconsistency.
In this work, we are interested in automated methods for knowledge graph creation (KGC) from input text. Progress on large language models (LLMs) has prompted a series of recent works applying them to KGC, e.g., via zero/few-shot prompting. Despite successes on small domain-specific datasets, these models face difficulties scaling up to text common in many real-world applications. A principal issue is that, in prior methods, the KG schema has to be included in the LLM prompt to generate valid triplets; larger and more complex schemas easily exceed the LLMs' context window length. Furthermore, there are scenarios where a fixed pre-defined schema is not available and we would like the method to construct a high-quality KG with a succinct self-generated schema. To address these problems, we propose a three-phase framework named Extract-Define-Canonicalize (EDC): open information extraction followed by schema definition and post-hoc canonicalization. EDC is flexible in that it can be applied to settings where a pre-defined target schema is available and when it is not; in the latter case, it constructs a schema automatically and applies self-canonicalization. To further improve performance, we introduce a trained component that retrieves schema elements relevant to the input text; this improves the LLMs' extraction performance in a retrieval-augmented generation-like manner. We demonstrate on three KGC benchmarks that EDC is able to extract high-quality triplets without any parameter tuning and with significantly larger schemas compared to prior works. Code for EDC is available at https://github.com/clear-nus/edc.
Question answering (QA) tasks have been extensively studied in the field of natural language processing (NLP). Answers to open-ended questions are highly diverse and difficult to quantify, and cannot be simply evaluated as correct or incorrect, unlike close-ended questions with definitive answers. While large language models (LLMs) have demonstrated strong capabilities across various tasks, they exhibit relatively weaker performance in evaluating answers to open-ended questions. In this study, we propose a method that leverages LLMs and the analytic hierarchy process (AHP) to assess answers to open-ended questions. We utilized LLMs to generate multiple evaluation criteria for a question. Subsequently, answers were subjected to pairwise comparisons under each criterion with LLMs, and scores for each answer were calculated in the AHP. We conducted experiments on four datasets using both ChatGPT-3.5-turbo and GPT-4. Our results indicate that our approach more closely aligns with human judgment compared to the four baselines. Additionally, we explored the impact of the number of criteria, variations in models, and differences in datasets on the results.
Despite the recent advancements in Large Language Models (LLMs), which have significantly enhanced the generative capabilities for various NLP tasks, LLMs still face limitations in directly handling retrieval tasks. However, many practical applications demand the seamless integration of both retrieval and generation. This paper introduces a novel and efficient One-pass Generation and retrieval framework (OneGen), designed to improve LLMs' performance on tasks that require both generation and retrieval. The proposed framework bridges the traditionally separate training approaches for generation and retrieval by incorporating retrieval tokens generated autoregressively. This enables a single LLM to handle both tasks simultaneously in a unified forward pass. We conduct experiments on two distinct types of composite tasks, RAG and Entity Linking, to validate the pluggability, effectiveness, and efficiency of OneGen in training and inference. Furthermore, our results show that integrating generation and retrieval within the same context preserves the generative capabilities of LLMs while improving retrieval performance. To the best of our knowledge, OneGen is the first to enable LLMs to conduct vector retrieval during the generation.
The scaling laws have become the de facto guidelines for designing large language models (LLMs), but they were studied under the assumption of unlimited computing resources for both training and inference. As LLMs are increasingly used as personalized intelligent assistants, their customization (i.e., learning through fine-tuning) and deployment onto resource-constrained edge devices will become more and more prevalent. An urging but open question is how a resource-constrained computing environment would affect the design choices for a personalized LLM. We study this problem empirically in this work. In particular, we consider the tradeoffs among a number of key design factors and their intertwined impacts on learning efficiency and accuracy. The factors include the learning methods for LLM customization, the amount of personalized data used for learning customization, the types and sizes of LLMs, the compression methods of LLMs, the amount of time afforded to learn, and the difficulty levels of the target use cases. Through extensive experimentation and benchmarking, we draw a number of surprisingly insightful guidelines for deploying LLMs onto resource-constrained devices. For example, an optimal choice between parameter learning and RAG may vary depending on the difficulty of the downstream task, the longer fine-tuning time does not necessarily help the model, and a compressed LLM may be a better choice than an uncompressed LLM to learn from limited personalized data.
While Large Language Models (LLMs) demonstrate impressive generation abilities, they frequently struggle when it comes to specialized domains due to their limited domain-specific knowledge. Studies on domain-specific LLMs resort to expanding the vocabulary before fine-tuning on domain-specific corpus, aiming to decrease the sequence length and enhance efficiency during decoding, without thoroughly investigating the results of vocabulary expansion to LLMs over different domains. Our pilot study reveals that expansion with only a subset of the entire vocabulary may lead to superior performance. Guided by the discovery, this paper explores how to identify a vocabulary subset to achieve the optimal results. We introduce VEGAD, an adaptive method that automatically identifies valuable words from a given domain vocabulary. Our method has been validated through experiments on three Chinese datasets, demonstrating its effectiveness. Additionally, we have undertaken comprehensive analyses of the method. The selection of a optimal subset for expansion has shown to enhance performance on both domain-specific tasks and general tasks, showcasing the potential of VEGAD.
The performance on general tasks decreases after Large Language Models (LLMs) are fine-tuned on domain-specific tasks, the phenomenon is known as Catastrophic Forgetting (CF). However, this paper presents a further challenge for real application of domain-specific LLMs beyond CF, called General Capabilities Integration (GCI), which necessitates the integration of both the general capabilities and domain knowledge within a single instance. The objective of GCI is not merely to retain previously acquired general capabilities alongside new domain knowledge, but to harmonize and utilize both sets of skills in a cohesive manner to enhance performance on domain-specific tasks. Taking legal domain as an example, we carefully design three groups of training and testing tasks without lacking practicability, and construct the corresponding datasets. To better incorporate general capabilities across domain-specific scenarios, we introduce ALoRA, which utilizes a multi-head attention module upon LoRA, facilitating direct information transfer from preceding tokens to the current one. This enhancement permits the representation to dynamically switch between domain-specific knowledge and general competencies according to the attention. Extensive experiments are conducted on the proposed tasks. The results exhibit the significance of our setting, and the effectiveness of our method.
Multilingual large language models (LLMs) seem to generalize somewhat across languages. We hypothesize this is a result of implicit vector space alignment. Evaluating such alignment, we see that larger models exhibit very high-quality linear alignments between corresponding concepts in different languages. Our experiments show that multilingual LLMs suffer from two familiar weaknesses: generalization works best for languages with similar typology, and for abstract concepts. For some models, e.g., the Llama-2 family of models, prompt-based embeddings align better than word embeddings, but the projections are less linear -- an observation that holds across almost all model families, indicating that some of the implicitly learned alignments are broken somewhat by prompt-based methods.
Since the onset of LLMs, translating natural language queries to structured SQL commands is assuming increasing. Unlike the previous reviews, this survey provides a comprehensive study of the evolution of LLM-based text-to-SQL systems, from early rule-based models to advanced LLM approaches, and how LLMs impacted this field. We discuss benchmarks, evaluation methods and evaluation metrics. Also, we uniquely study the role of integration of knowledge graphs for better contextual accuracy and schema linking in these systems. The current techniques fall into two categories: in-context learning of corpus and fine-tuning, which then leads to approaches such as zero-shot, few-shot learning from the end, and data augmentation. Finally, we highlight key challenges such as computational efficiency, model robustness, and data privacy with perspectives toward their development and improvements in potential areas for future of LLM-based text-to-SQL system.
Efficient scheduling is crucial for interactive Large Language Model (LLM) applications, where low request completion time directly impacts user engagement. Size-based scheduling algorithms like Shortest Remaining Process Time (SRPT) aim to reduce average request completion time by leveraging known or estimated request sizes and allowing preemption by incoming jobs with shorter service times. However, two main challenges arise when applying size-based scheduling to LLM systems. First, accurately predicting output lengths from prompts is challenging and often resource-intensive, making it impractical for many systems. As a result, the state-of-the-art LLM systems default to first-come, first-served scheduling, which can lead to head-of-line blocking and reduced system efficiency. Second, preemption introduces extra memory overhead to LLM systems as they must maintain intermediate states for unfinished (preempted) requests. In this paper, we propose TRAIL, a method to obtain output predictions from the target LLM itself. After generating each output token, we recycle the embedding of its internal structure as input for a lightweight classifier that predicts the remaining length for each running request. Using these predictions, we propose a prediction-based SRPT variant with limited preemption designed to account for memory overhead in LLM systems. This variant allows preemption early in request execution when memory consumption is low but restricts preemption as requests approach completion to optimize resource utilization. On the theoretical side, we derive a closed-form formula for this SRPT variant in an M/G/1 queue model, which demonstrates its potential value. In our system, we implement this preemption policy alongside our embedding-based prediction method.
Large Language Models (LLMs) are transforming programming practices, offering significant capabilities for code generation activities. While researchers have explored the potential of LLMs in various domains, this paper focuses on their use in programming tasks, drawing insights from user studies that assess the impact of LLMs on programming tasks. We first examined the user interaction behaviors with LLMs observed in these studies, from the types of requests made to task completion strategies. Additionally, our analysis reveals both benefits and weaknesses of LLMs showing mixed effects on the human and task. Lastly, we looked into what factors from the human, LLM or the interaction of both, affect the human's enhancement as well as the task performance. Our findings highlight the variability in human-LLM interactions due to the non-deterministic nature of both parties (humans and LLMs), underscoring the need for a deeper understanding of these interaction patterns. We conclude by providing some practical suggestions for researchers as well as programmers.
Recent advances in Large Language Models (LLMs) have demonstrated their potential as autonomous agents across various tasks. One emerging application is the use of LLMs in playing games. In this work, we explore a practical problem for the gaming industry: Can LLMs be used to measure game difficulty? We propose a general game-testing framework using LLM agents and test it on two widely played strategy games: Wordle and Slay the Spire. Our results reveal an interesting finding: although LLMs may not perform as well as the average human player, their performance, when guided by simple, generic prompting techniques, shows a statistically significant and strong correlation with difficulty indicated by human players. This suggests that LLMs could serve as effective agents for measuring game difficulty during the development process. Based on our experiments, we also outline general principles and guidelines for incorporating LLMs into the game testing process.
Simulated patient systems play a crucial role in modern medical education and research, providing safe, integrative learning environments and enabling clinical decision-making simulations. Large Language Models (LLM) could advance simulated patient systems by replicating medical conditions and patient-doctor interactions with high fidelity and low cost. However, ensuring the effectiveness and trustworthiness of these systems remains a challenge, as they require a large, diverse, and precise patient knowledgebase, along with a robust and stable knowledge diffusion to users. Here, we developed AIPatient, an advanced simulated patient system with AIPatient Knowledge Graph (AIPatient KG) as the input and the Reasoning Retrieval-Augmented Generation (Reasoning RAG) agentic workflow as the generation backbone. AIPatient KG samples data from Electronic Health Records (EHRs) in the Medical Information Mart for Intensive Care (MIMIC)-III database, producing a clinically diverse and relevant cohort of 1,495 patients with high knowledgebase validity (F1 0.89). Reasoning RAG leverages six LLM powered agents spanning tasks including retrieval, KG query generation, abstraction, checker, rewrite, and summarization. This agentic framework reaches an overall accuracy of 94.15% in EHR-based medical Question Answering (QA), outperforming benchmarks that use either no agent or only partial agent integration. Our system also presents high readability (median Flesch Reading Ease 77.23; median Flesch Kincaid Grade 5.6), robustness (ANOVA F-value 0.6126, p>0.1), and stability (ANOVA F-value 0.782, p>0.1). The promising performance of the AIPatient system highlights its potential to support a wide range of applications, including medical education, model evaluation, and system integration.
Evaluation of large language model (LLM) outputs requires users to make critical judgments about the best outputs across various configurations. This process is costly and takes time given the large amounts of data. LLMs are increasingly used as evaluators to filter training data, evaluate model performance or assist human evaluators with detailed assessments. To support this process, effective front-end tools are critical for evaluation. Two common approaches for using LLMs as evaluators are direct assessment and pairwise comparison. In our study with machine learning practitioners (n=15), each completing 6 tasks yielding 131 evaluations, we explore how task-related factors and assessment strategies influence criteria refinement and user perceptions. Findings show that users performed more evaluations with direct assessment by making criteria task-specific, modifying judgments, and changing the evaluator model. We conclude with recommendations for how systems can better support interactions in LLM-assisted evaluations.
This paper investigates the impact of verbose LLM translations on evaluation. We first demonstrate the prevalence of this behavior across several LLM outputs drawn from the WMT 2024 general shared task on machine translation. We then identify the primary triggers of verbosity, including safety, copyright concerns, and insufficient context in short input queries. Finally, we show that ignoring this behavior unfairly penalizes more verbose LLMs according to both automatic and human evaluations, highlighting the need to address this issue for more accurate future evaluations.
Large language models (LLMs) can be used to analyze cyber threat intelligence (CTI) data from cybercrime forums, which contain extensive information and key discussions about emerging cyber threats. However, to date, the level of accuracy and efficiency of LLMs for such critical tasks has yet to be thoroughly evaluated. Hence, this study assesses the performance of an LLM system built on the OpenAI GPT-3.5-turbo model [8] to extract CTI information. To do so, a random sample of more than 700 daily conversations from three cybercrime forums - XSS, Exploit_in, and RAMP - was extracted, and the LLM system was instructed to summarize the conversations and predict 10 key CTI variables, such as whether a large organization and/or a critical infrastructure is being targeted, with only simple human-language instructions. Then, two coders reviewed each conversation and evaluated whether the information extracted by the LLM was accurate. The LLM system performed well, with an average accuracy score of 96.23%, an average precision of 90% and an average recall of 88.2%. Various ways to enhance the model were uncovered, such as the need to help the LLM distinguish between stories and past events, as well as being careful with verb tenses in prompts. Nevertheless, the results of this study highlight the relevance of using LLMs for cyber threat intelligence.
Recognizing if LLM output can be grounded in evidence is central to many tasks in NLP: retrieval-augmented generation, summarization, document-grounded dialogue, and more. Current approaches to this kind of fact-checking are based on verifying each piece of a model generation against potential evidence using an LLM. However, this process can be very computationally expensive, requiring many calls to a model to check a single response. In this work, we show how to build small fact-checking models that have GPT-4-level performance but for 400x lower cost. We do this by constructing synthetic training data with GPT-4, which involves creating realistic yet challenging instances of factual errors via a structured generation procedure. Training on this data teaches models to check each fact in the claim and recognize synthesis of information across sentences. For evaluation, we unify datasets from recent work on fact-checking and grounding LLM generations into a new benchmark, LLM-AggreFact. Our best system MiniCheck-FT5 (770M parameters) outperforms all systems of comparable size and reaches GPT-4 accuracy. We release LLM-AggreFact, code for data synthesis, and models.
Text style transfer (TST) aims to modify the style of a text without altering its original meaning. Large language models (LLMs) demonstrate superior performance across multiple tasks, including TST. However, in zero-shot setups, they tend to directly copy a significant portion of the input text to the output without effectively changing its style. To enhance the stylistic variety and fluency of the text, we present sNeuron-TST, a novel approach for steering LLMs using style-specific neurons in TST. Specifically, we identify neurons associated with the source and target styles and deactivate source-style-only neurons to give target-style words a higher probability, aiming to enhance the stylistic diversity of the generated text. However, we find that this deactivation negatively impacts the fluency of the generated text, which we address by proposing an improved contrastive decoding method that accounts for rapid token probability shifts across layers caused by deactivated source-style neurons. Empirical experiments demonstrate the effectiveness of the proposed method on six benchmarks, encompassing formality, toxicity, politics, politeness, authorship, and sentiment.
Large model inference is shifting from cloud to edge due to concerns about the privacy of user interaction data. However, edge devices often struggle with limited computing power, memory, and bandwidth, requiring collaboration across multiple devices to run and speed up LLM inference. Pipeline parallelism, the mainstream solution, is inefficient for single-user scenarios, while tensor parallelism struggles with frequent communications. In this paper, we argue that tensor parallelism can be more effective than pipeline on low-resource devices, and present a compute- and memory-efficient tensor parallel inference system, named TPI-LLM, to serve 70B-scale models. TPI-LLM keeps sensitive raw data local in the users' devices and introduces a sliding window memory scheduler to dynamically manage layer weights during inference, with disk I/O latency overlapped with the computation and communication. This allows larger models to run smoothly on memory-limited devices. We analyze the communication bottleneck and find that link latency, not bandwidth, emerges as the main issue, so a star-based allreduce algorithm is implemented. Through extensive experiments on both emulated and real testbeds, TPI-LLM demonstrated over 80% less time-to-first-token and token latency compared to Accelerate, and over 90% compared to Transformers and Galaxy, while cutting the peak memory footprint of Llama 2-70B by 90%, requiring only 3.1 GB of memory for 70B-scale models.
Supervised fine-tuning (SFT) is crucial for adapting Large Language Models (LLMs) to specific tasks. In this work, we demonstrate that the order of training data can lead to significant training imbalances, potentially resulting in performance degradation. Consequently, we propose to mitigate this imbalance by merging SFT models fine-tuned with different data orders, thereby enhancing the overall effectiveness of SFT. Additionally, we introduce a novel technique, "parameter-selection merging," which outperforms traditional weighted-average methods on five datasets. Further, through analysis and ablation studies, we validate the effectiveness of our method and identify the sources of performance improvements.
Large Language Models (LLMs) have shown the impressive capability of answering questions in a wide range of scenarios. However, when LLMs face different types of questions, it is worth exploring whether LLMs are aware that some questions have limited answers and need to respond more deterministically but some do not. We refer to this as question awareness of LLMs. The lack of question awareness in LLMs leads to two phenomena that LLMs are: (1) too casual to answer non-open-ended questions or (2) too boring to answer open-ended questions. In this paper, we first evaluate the question awareness in LLMs. The experimental results show that LLMs have the issues of lacking awareness of questions in certain domains, e.g. factual knowledge, resulting in hallucinations during the generation. To mitigate these, we propose a method called Question Awareness Temperature Sampling (QuATS). This method enhances the question awareness of LLMs by adaptively adjusting the output distributions based on question features. The automatic adjustment in QuATS eliminates the need for manual temperature tuning in text generation and consistently improves model performance in various benchmarks.
Post-training, particularly reinforcement learning (RL) using self-play-generated data, has become a new learning paradigm for large language models (LLMs). However, scaling RL to develop a general reasoner remains a research challenge, as existing methods focus on task-specific reasoning without adequately addressing generalization across a broader range of tasks. Moreover, unlike traditional RL with limited action space, LLMs operate in an infinite space, making it crucial to search for valuable and diverse strategies to solve problems effectively. To address this, we propose searching within the action space on high-level abstract plans to enhance model generalization and introduce Critical Plan Step Learning (CPL), comprising: 1) searching on plan, using Monte Carlo Tree Search (MCTS) to explore diverse plan steps in multi-step reasoning tasks, and 2) learning critical plan steps through Step-level Advantage Preference Optimization (Step-APO), which integrates advantage estimates for step preference obtained via MCTS into Direct Preference Optimization (DPO). This combination helps the model effectively learn critical plan steps, enhancing both reasoning capabilities and generalization. Experimental results demonstrate that our method, trained exclusively on GSM8K and MATH, not only significantly improves performance on GSM8K (+10.5%) and MATH (+6.5%), but also enhances out-of-domain reasoning benchmarks, such as HumanEval (+12.2%), GPQA (+8.6%), ARC-C (+4.0%), MMLU-STEM (+2.2%), and BBH (+1.8%).
Code translation converts code from one programming language to another while maintaining its original functionality, which is crucial for software migration, system refactoring, and cross-platform development. Traditional rule-based methods rely on manually-written rules, which can be time-consuming and often result in less readable code. To overcome this, learning-based methods have been developed, leveraging parallel data to train models for automated code translation. More recently, the advance of Large Language Models (LLMs) further boosts learning-based code translation. Although promising, LLM-translated program still suffers from diverse quality issues (e.g., syntax errors and semantic errors). In particular, it can be challenging for LLMs to self-debug these errors when simply provided with the corresponding error messages. In this work, we propose a novel LLM-based multi-agent system TRANSAGENT, which enhances LLM-based code translation by fixing the syntax errors and semantic errors with the synergy between four LLM-based agents, including Initial Code Translator, Syntax Error Fixer, Code Aligner, and Semantic Error Fixer. The main insight of TRANSAGENT is to first localize the error code block in the target program based on the execution alignment between the target and source program, which can narrow down the fixing space and thus lower down the fixing difficulties. To evaluate TRANSAGENT, we first construct a new benchmark from recent programming tasks to mitigate the potential data leakage issue. On our benchmark, TRANSAGENT outperforms the latest LLM-based code translation technique UniTrans in both translation effectiveness and efficiency; additionally, our evaluation on different LLMs show the generalization of TRANSAGENT and our ablation study shows the contribution of each agent.
In human-written articles, we often leverage the subtleties of text style, such as bold and italics, to guide the attention of readers. These textual emphases are vital for the readers to grasp the conveyed information. When interacting with large language models (LLMs), we have a similar need -- steering the model to pay closer attention to user-specified information, e.g., an instruction. Existing methods, however, are constrained to process plain text and do not support such a mechanism. This motivates us to introduce PASTA -- Post-hoc Attention STeering Approach, a method that allows LLMs to read text with user-specified emphasis marks. To this end, PASTA identifies a small subset of attention heads and applies precise attention reweighting on them, directing the model attention to user-specified parts. Like prompting, PASTA is applied at inference time and does not require changing any model parameters. Experiments demonstrate that PASTA can substantially enhance an LLM's ability to follow user instructions or integrate new knowledge from user inputs, leading to a significant performance improvement on a variety of tasks, e.g., an average accuracy improvement of 22% for LLAMA-7B. Our code is publicly available at https://github.com/QingruZhang/PASTA .
The applications of large language models (LLMs) have been widely spread across all domains. However, the basic abilities such as the controllability of LLMs are still limited. To address this, we propose "Self-controller", a novel agentic framework bringing self-awareness into LLMs' reasoning logic. The core idea of this work is to maintain states based on the LLM's response, letting the LLM become self-aware of current status and think step by step in a multi-round chain-of-thought paradigm. Our experiment on the state of textual length has shown the controllability and effectiveness of the Self-controller. We further implement a binary search algorithm to accelerate the generation process based on the linearity and monotonicity of the textual length state. Another advantage of the Self-controller comes with DeepSeek's Context Caching technology, which significantly saves computational token consumption when a cluster of conversations shares the same prefix of context. Theoretically, we prove that in this scenario the extra time complexity is $O(c \log n)$. Results of the back-of-the-envelope estimation suggest that the token consumption of our method is no more than twice as much as that of the trivial single-round generation. Furthermore, our ablation study on word constraints demonstrates the Self-controller's consistent controllability across all foundation models.
This paper investigates the performance of several representative large models in the field of literature recommendation and explores potential biases. The results indicate that while some large models' recommendations can be somewhat satisfactory after simple manual screening, overall, the accuracy of these models in specific literature recommendation tasks is generally moderate. Additionally, the models tend to recommend literature that is timely, collaborative, and expands or deepens the field. In scholar recommendation tasks. There is no evidence to suggest that LLMs exacerbate inequalities related to gender, race, or the level of development of countries.
The impressive performance of Large Language Models (LLMs) has consistently surpassed numerous human-designed benchmarks, presenting new challenges in assessing the shortcomings of LLMs. Designing tasks and finding LLMs' limitations are becoming increasingly important. In this paper, we investigate the question of whether an LLM can discover its own limitations from the errors it makes. To this end, we propose a Self-Challenge evaluation framework with human-in-the-loop. Starting from seed instances that GPT-4 fails to answer, we prompt GPT-4 to summarize error patterns that can be used to generate new instances and incorporate human feedback on them to refine these patterns for generating more challenging data, iteratively. We end up with 8 diverse patterns, such as text manipulation and questions with assumptions. We then build a benchmark, SC-G4, consisting of 1,835 instances generated by GPT-4 using these patterns, with human-annotated gold responses. The SC-G4 serves as a challenging benchmark that allows for a detailed assessment of LLMs' abilities. Our results show that only 44.96\% of instances in SC-G4 can be answered correctly by GPT-4. Interestingly, our pilot study indicates that these error patterns also challenge other LLMs, such as Claude-3 and Llama-3, and cannot be fully resolved through fine-tuning. Our work takes the first step to demonstrate that LLMs can autonomously identify their inherent flaws and provide insights for future dynamic and automatic evaluation.
To combat the misuse of Large Language Models (LLMs), many recent studies have presented LLM-generated-text detectors with promising performance. When users instruct LLMs to generate texts, the instruction can include different constraints depending on the user's need. However, most recent studies do not cover such diverse instruction patterns when creating datasets for LLM detection. In this paper, we reveal that even task-oriented constraints -- constraints that would naturally be included in an instruction and are not related to detection-evasion -- cause existing powerful detectors to have a large variance in detection performance. We focus on student essay writing as a realistic domain and manually create task-oriented constraints based on several factors for essay quality. Our experiments show that the standard deviation (SD) of current detector performance on texts generated by an instruction with such a constraint is significantly larger (up to an SD of 14.4 F1-score) than that by generating texts multiple times or paraphrasing the instruction. We also observe an overall trend where the constraints can make LLM detection more challenging than without them. Finally, our analysis indicates that the high instruction-following ability of LLMs fosters the large impact of such constraints on detection performance.
Accurate diagnosis of ocular surface diseases is critical in optometry and ophthalmology, which hinge on integrating clinical data sources (e.g., meibography imaging and clinical metadata). Traditional human assessments lack precision in quantifying clinical observations, while current machine-based methods often treat diagnoses as multi-class classification problems, limiting the diagnoses to a predefined closed-set of curated answers without reasoning the clinical relevance of each variable to the diagnosis. To tackle these challenges, we introduce an innovative multi-modal diagnostic pipeline (MDPipe) by employing large language models (LLMs) for ocular surface disease diagnosis. We first employ a visual translator to interpret meibography images by converting them into quantifiable morphology data, facilitating their integration with clinical metadata and enabling the communication of nuanced medical insight to LLMs. To further advance this communication, we introduce a LLM-based summarizer to contextualize the insight from the combined morphology and clinical metadata, and generate clinical report summaries. Finally, we refine the LLMs' reasoning ability with domain-specific insight from real-life clinician diagnoses. Our evaluation across diverse ocular surface disease diagnosis benchmarks demonstrates that MDPipe outperforms existing standards, including GPT-4, and provides clinically sound rationales for diagnoses.
Empathy serves as a cornerstone in enabling prosocial behaviors, and can be evoked through sharing of personal experiences in stories. While empathy is influenced by narrative content, intuitively, people respond to the way a story is told as well, through narrative style. Yet the relationship between empathy and narrative style is not fully understood. In this work, we empirically examine and quantify this relationship between style and empathy using LLMs and large-scale crowdsourcing studies. We introduce a novel, theory-based taxonomy, HEART (Human Empathy and Narrative Taxonomy) that delineates elements of narrative style that can lead to empathy with the narrator of a story. We establish the performance of LLMs in extracting narrative elements from HEART, showing that prompting with our taxonomy leads to reasonable, human-level annotations beyond what prior lexicon-based methods can do. To show empirical use of our taxonomy, we collect a dataset of empathy judgments of stories via a large-scale crowdsourcing study with N=2,624 participants. We show that narrative elements extracted via LLMs, in particular, vividness of emotions and plot volume, can elucidate the pathways by which narrative style cultivates empathy towards personal stories. Our work suggests that such models can be used for narrative analyses that lead to human-centered social and behavioral insights.