Large language models (LLMs) achieve state-of-the-art performance on multiple language tasks, yet their safety guardrails can be circumvented, leading to harmful generations. In light of this, recent research on safety mechanisms has emerged, revealing that when safety representations or component are suppressed, the safety capability of LLMs are compromised. However, existing research tends to overlook the safety impact of multi-head attention mechanisms, despite their crucial role in various model functionalities. Hence, in this paper, we aim to explore the connection between standard attention mechanisms and safety capability to fill this gap in the safety-related mechanistic interpretability. We propose a novel metric which tailored for multi-head attention, the Safety Head ImPortant Score (Ships), to assess the individual heads' contributions to model safety. Based on this, we generalize Ships to the dataset level and further introduce the Safety Attention Head AttRibution Algorithm (Sahara) to attribute the critical safety attention heads inside the model. Our findings show that the special attention head has a significant impact on safety. Ablating a single safety head allows aligned model (e.g., Llama-2-7b-chat) to respond to 16 times more harmful queries, while only modifying 0.006% of the parameters, in contrast to the ~ 5% modification required in previous studies. More importantly, we demonstrate that attention heads primarily function as feature extractors for safety and models fine-tuned from the same base model exhibit overlapping safety heads through comprehensive experiments. Together, our attribution approach and findings provide a novel perspective for unpacking the black box of safety mechanisms within large models.
Molecular optimization (MO) is a crucial stage in drug discovery in which task-oriented generated molecules are optimized to meet practical industrial requirements. Existing mainstream MO approaches primarily utilize external property predictors to guide iterative property optimization. However, learning all molecular samples in the vast chemical space is unrealistic for predictors. As a result, errors and noise are inevitably introduced during property prediction due to the nature of approximation. This leads to discrepancy accumulation, generalization reduction and suboptimal molecular candidates. In this paper, we propose a text-guided multi-property molecular optimization method utilizing transformer-based diffusion language model (TransDLM). TransDLM leverages standardized chemical nomenclature as semantic representations of molecules and implicitly embeds property requirements into textual descriptions, thereby preventing error propagation during diffusion process. Guided by physically and chemically detailed textual descriptions, TransDLM samples and optimizes encoded source molecules, retaining core scaffolds of source molecules and ensuring structural similarities. Moreover, TransDLM enables simultaneous sampling of multiple molecules, making it ideal for scalable, efficient large-scale optimization through distributed computation on web platforms. Furthermore, our approach surpasses state-of-the-art methods in optimizing molecular structural similarity and enhancing chemical properties on the benchmark dataset. The code is available at: https://anonymous.4open.science/r/TransDLM-A901.
The Instruction-Driven Game Engine (IDGE) project aims to democratize game development by enabling a large language model (LLM) to follow free-form game descriptions and generate game-play processes. The IDGE allows users to create games simply by natural language instructions, which significantly lowers the barrier for game development. We approach the learning process for IDGEs as a Next State Prediction task, wherein the model autoregressively predicts the game states given player actions. The computation of game states must be precise; otherwise, slight errors could corrupt the game-play experience. This is challenging because of the gap between stability and diversity. To address this, we train the IDGE in a curriculum manner that progressively increases its exposure to complex scenarios. Our initial progress lies in developing an IDGE for Poker, which not only supports a wide range of poker variants but also allows for highly individualized new poker games through natural language inputs. This work lays the groundwork for future advancements in transforming how games are created and played.
Teaching large language models (LLMs) to generate text with citations to evidence sources can mitigate hallucinations and enhance verifiability in information-seeking systems. However, improving this capability requires high-quality attribution data, which is costly and labor-intensive. Inspired by recent advances in self-improvement that enhance LLMs without manual annotation, we present START, a Self-Taught AttRibuTion framework for iteratively improving the attribution capability of LLMs. First, to prevent models from stagnating due to initially insufficient supervision signals, START leverages the model to self-construct synthetic training data for warming up. To further self-improve the model's attribution ability, START iteratively utilizes fine-grained preference supervision signals constructed from its sampled responses to encourage robust, comprehensive, and attributable generation. Experiments on three open-domain question-answering datasets, covering long-form QA and multi-step reasoning, demonstrate significant performance gains of 25.13% on average without relying on human annotations and more advanced models. Further analysis reveals that START excels in aggregating information across multiple sources.
Many students struggle with math word problems (MWPs), often finding it difficult to identify key information and select the appropriate mathematical operations. Schema-based instruction (SBI) is an evidence-based strategy that helps students categorize problems based on their structure, improving problem-solving accuracy. Building on this, we propose a Schema-Based Instruction Retrieval-Augmented Generation (SBI-RAG) framework that incorporates a large language model (LLM). Our approach emphasizes step-by-step reasoning by leveraging schemas to guide solution generation. We evaluate its performance on the GSM8K dataset, comparing it with GPT-4 and GPT-3.5 Turbo, and introduce a "reasoning score" metric to assess solution quality. Our findings suggest that SBI-RAG enhances reasoning clarity and facilitates a more structured problem-solving process potentially providing educational benefits for students.
We present an UI agent for user interface (UI) interaction tasks, using Vision-Language Model Florence-2-Base. The agent's primary task is identifying the screen coordinates of the UI element corresponding to the user's command. It demonstrates very strong performance on Screenspot and OmniAct annotations, while maintaining a very small size of 0.27B parameters and minimal latency. Moreover, training needs small compute budget of 56 GPU-hours (worth about 40 USD). Relevant improvement comes from vision-specific multi-task training and MLLM-based data augmentation. We hope that decreased needs for expensive compute resources and manually annotated data will allow to facilitate more inclusive and sustainable research of UI agents.
Large Language Model (LLM)-driven interactive systems currently show potential promise in healthcare domains. Despite their remarkable capabilities, LLMs typically lack personalized recommendations and diagnosis analysis in sophisticated medical applications, causing hallucinations and performance bottlenecks. To address these challenges, this paper proposes MedAide, an LLM-based omni medical multi-agent collaboration framework for specialized healthcare services. Specifically, MedAide first performs query rewriting through retrieval-augmented generation to accomplish accurate medical intent understanding. Immediately, we devise a contextual encoder to obtain intent prototype embeddings, which are used to recognize fine-grained intents by similarity matching. According to the intent relevance, the activated agents collaborate effectively to provide integrated decision analysis. Extensive experiments are conducted on four medical benchmarks with composite intents. Experimental results from automated metrics and expert doctor evaluations show that MedAide outperforms current LLMs and improves their medical proficiency and strategic reasoning.
To tackle the AVeriTeC shared task hosted by the FEVER-24, we introduce a system that only employs publicly available large language models (LLMs) for each step of automated fact-checking, dubbed the Herd of Open LLMs for verifying real-world claims (HerO). For evidence retrieval, a language model is used to enhance a query by generating hypothetical fact-checking documents. We prompt pretrained and fine-tuned LLMs for question generation and veracity prediction by crafting prompts with retrieved in-context samples. HerO achieved 2nd place on the leaderboard with the AVeriTeC score of 0.57, suggesting the potential of open LLMs for verifying real-world claims. For future research, we make our code publicly available at https://github.com/ssu-humane/HerO.
Vector data is one of the two core data structures in geographic information science (GIS), essential for accurately storing and representing geospatial information. Shapefile, the most widely used vector data format, has become the industry standard supported by all major geographic information systems. However, processing this data typically requires specialized GIS knowledge and skills, creating a barrier for researchers from other fields and impeding interdisciplinary research in spatial data analysis. Moreover, while large language models (LLMs) have made significant advancements in natural language processing and task automation, they still face challenges in handling the complex spatial and topological relationships inherent in GIS vector data. To address these challenges, we propose ShapefileGPT, an innovative framework powered by LLMs, specifically designed to automate Shapefile tasks. ShapefileGPT utilizes a multi-agent architecture, in which the planner agent is responsible for task decomposition and supervision, while the worker agent executes the tasks. We developed a specialized function library for handling Shapefiles and provided comprehensive API documentation, enabling the worker agent to operate Shapefiles efficiently through function calling. For evaluation, we developed a benchmark dataset based on authoritative textbooks, encompassing tasks in categories such as geometric operations and spatial queries. ShapefileGPT achieved a task success rate of 95.24%, outperforming the GPT series models. In comparison to traditional LLMs, ShapefileGPT effectively handles complex vector data analysis tasks, overcoming the limitations of traditional LLMs in spatial analysis. This breakthrough opens new pathways for advancing automation and intelligence in the GIS field, with significant potential in interdisciplinary data analysis and application contexts.
PRefLexOR (Preference-based Recursive Language Modeling for Exploratory Optimization of Reasoning) combines preference optimization with concepts from Reinforcement Learning to enable models to self-teach through iterative reasoning improvements. We propose a recursive learning approach that engages the model in multi-step reasoning, revisiting, and refining intermediate steps before producing a final output in training and inference phases. Through multiple training stages, the model first learns to align its reasoning with accurate decision paths by optimizing the log odds between preferred and non-preferred responses. During this process, PRefLexOR builds a dynamic knowledge graph by generating questions from random text chunks and retrieval-augmentation to contextualize relevant details from the entire training corpus. In the second stage, preference optimization enhances model performance by using rejection sampling to fine-tune reasoning quality by continually producing in-situ training data while masking the reasoning steps. Recursive optimization within a thinking token framework introduces iterative feedback loops, where the model refines reasoning, achieving deeper coherence, consistency, and adaptability. Implemented in small language models with only 3 billion parameters, we should that even tiny models can iteratively teach themselves to reason with greater depth and reflectivity. Our implementation is straightforward and can be incorporated into any existing pretrained LLM. We focus our examples on applications in biological materials science and demonstrate the method in a variety of case studies that range from in-domain to cross-domain applications. Using reasoning strategies that include thinking and reflection modalities we build a multi-agent recursive self-improving inference approach to successively improve responses via repeated sampling in inference time.
In this paper, we address the challenge of optimizing training setups for Large Language Models (LLMs) of low-resource language with a limited amount of corpus. Existing works adopt multi-epoch, multi-lingual, and two-stage training to utilize the limited target language corpus efficiently. However, there is still a lack of understanding about the optimal hyperparameter setups for combining these three approaches to train LLMs. We exhaustively explore training setups for low-resource language LLM, combining these three approaches, and found the following insights for efficiently reducing the cost of hyperparameter search: (1) As the amount of target language corpus decreases, the optimal training approach shifts from monolingual single-stage training to multi-lingual two-stage training at a compute budget dependent threshold. (2) The optimal model scale remains stable regardless of the amount of target language corpus, allowing the use of the compute-optimal scale of monolingual training. (3) The optimal number of epochs can be extrapolated from smaller-scale experiments to larger scale using our proposed model. Also, we provide evidence that, in single-stage training, the target language validation loss follows a power law with respect to the target language ratio, with an exponent independent of the amount of data, model scale, and language pair.
Large language models have achieved remarkable capabilities, but aligning their outputs with human values and preferences remains a significant challenge. Existing alignment methods primarily focus on positive examples while overlooking the importance of negative responses in guiding models away from undesirable behaviors. For instance, the widely-used alignment datasets reveals a scarcity of explicit negative examples that contradict human values, hindering its ability to discourage harmful or biased outputs during training. To address this limitation, we propose NEAT, i.e., NEgative-prompt-driven AlignmenT, to introduce negative prompts to generate undesirable responses alongside positive examples during the optimization process. NEAT explicitly penalizes the model for producing harmful outputs, guiding it not only toward desirable behaviors but also steering it away from generating undesirable, biased responses. This dual feedback mechanism enables better alignment with human preferences, crucial in contexts where avoiding harm is paramount. Starting from a pre-trained language model, NEAT performs online alignment by incorporating a ranking loss derived from an expanded preference dataset containing both positive and negative examples. Extensive experiments validate NEAT's effectiveness in significantly enhancing language models' alignment with human values and preferences.
The development of Large Language Models (LLMs) has significantly advanced various AI applications in commercial and scientific research fields, such as scientific literature summarization, writing assistance, and knowledge graph construction. However, a significant challenge is the high risk of hallucination during LLM inference, which can lead to security concerns like factual inaccuracies, inconsistent information, and fabricated content. To tackle this issue, it is essential to develop effective methods for reducing hallucination while maintaining the original capabilities of the LLM. This paper introduces a novel approach called Iterative Model-level Contrastive Learning (Iter-AHMCL) to address hallucination. This method modifies the representation layers of pre-trained LLMs by using contrastive `positive' and `negative' models, trained on data with and without hallucinations. By leveraging the differences between these two models, we create a more straightforward pathway to eliminate hallucinations, and the iterative nature of contrastive learning further enhances performance. Experimental validation on four pre-trained foundation LLMs (LLaMA2, Alpaca, LLaMA3, and Qwen) finetuning with a specially designed dataset shows that our approach achieves an average improvement of 10.1 points on the TruthfulQA benchmark. Comprehensive experiments demonstrate the effectiveness of Iter-AHMCL in reducing hallucination while maintaining the general capabilities of LLMs.
Our information and communication environment has fallen short of the ideals that networked global communication might have served. Identifying all the causes of its pathologies is difficult, but existing recommender systems very likely play a contributing role. In this paper, which draws on the normative tools of philosophy of computing, informed by empirical and technical insights from natural language processing and recommender systems, we make the moral case for an alternative approach. We argue that existing recommenders incentivise mass surveillance, concentrate power, fall prey to narrow behaviourism, and compromise user agency. Rather than just trying to avoid algorithms entirely, or to make incremental improvements to the current paradigm, researchers and engineers should explore an alternative paradigm: the use of language model (LM) agents to source and curate content that matches users' preferences and values, expressed in natural language. The use of LM agents for recommendation poses its own challenges, including those related to candidate generation, computational efficiency, preference modelling, and prompt injection. Nonetheless, if implemented successfully LM agents could: guide us through the digital public sphere without relying on mass surveillance; shift power away from platforms towards users; optimise for what matters instead of just for behavioural proxies; and scaffold our agency instead of undermining it.
This study offers an initial evaluation of a human-in-the-loop system leveraging GPT-4 (a large language model or LLM), and Retrieval-Augmented Generation (RAG) to identify and define jargon terms in scientific abstracts, based on readers' self-reported knowledge. The system achieves fairly high recall in identifying jargon and preserves relative differences in readers' jargon identification, suggesting personalization as a feasible use-case for LLMs to support sense-making of complex information. Surprisingly, using only abstracts for context to generate definitions yields slightly more accurate and higher quality definitions than using RAG-based context from the fulltext of an article. The findings highlight the potential of generative AI for assisting science reporters, and can inform future work on developing tools to simplify dense documents.
Through end-to-end training to predict the next token, LLMs have become valuable tools for various tasks. Enhancing their core training in language modeling can improve numerous downstream applications. A successful approach to enhance language modeling uses a separate planning module to predict abstract labels of future sentences and conditions the LM on these predictions. However, this method is non-differentiable, preventing joint end-to-end tuning of the planner with the LM. We propose an effective method to improve this approach by enabling joint fine-tuning of the planner and the LM. We show that a naive way of approximating the gradient of selecting a label via the straight-through estimator is not effective. Instead, we propose to use the predicted label probabilities as mixing weights to condition the LM on a weighted average of label embeddings in a differentiable manner. This not only enables joint fine-tuning of the planner and the LM, but also allows the LM to draw on the full label distribution predicted by the planner, retaining more information. Our experimental results show consistent improvements in perplexity.
Machine learning (ML) exhibits promise in the clinical domain. However, it is constrained by data scarcity and ethical considerations, as the generation of clinical trials presents significant challenges due to stringent privacy regulations, high costs, and the extended duration required for conducting studies with human participants. Despite the advancements of large language models (LLMs) in general generation tasks, their potential in facilitating the generation of synthetic clinical trials is under-explored. To address this gap, we introduce a novel Retrieval-Reasoning few-shot framework that leverages LLMs to generate artificial yet realistic and diverse clinical trials with binary success/failure labels. Experiments conducted on real clinical trials from the \url{ClinicalTrials.gov} database demonstrate that our synthetic data can effectively augment real datasets. Furthermore, by fine-tuning a pre-trained model as a binary classifier on synthetic clinical trial datasets, we demonstrate that this augmentation enhances model training for downstream tasks such as trial outcome prediction. Our findings suggest that LLMs for synthetic clinical trial generation hold promise for accelerating clinical research and upholding ethical standards for patient privacy. The code is publicly available at https://anonymous.4open.science/r/Retrieval_Reasoning_Clinical_Trial_Generation-3EC4.
Messenger RNA (mRNA) plays a crucial role in protein synthesis, with its codon structure directly impacting biological properties. While Language Models (LMs) have shown promise in analyzing biological sequences, existing approaches fail to account for the hierarchical nature of mRNA's codon structure. We introduce Hierarchical Encoding for mRNA Language Modeling (HELM), a novel pre-training strategy that incorporates codon-level hierarchical structure into language model training. HELM modulates the loss function based on codon synonymity, aligning the model's learning process with the biological reality of mRNA sequences. We evaluate HELM on diverse mRNA datasets and tasks, demonstrating that HELM outperforms standard language model pre-training as well as existing foundation model baselines on seven diverse downstream property prediction tasks and an antibody region annotation tasks on average by around 8%. Additionally, HELM enhances the generative capabilities of language model, producing diverse mRNA sequences that better align with the underlying true data distribution compared to non-hierarchical baselines.
We study how features emerge, disappear, and persist across models fine-tuned on different domains of text. More specifically, we start from a base one-layer Transformer language model that is trained on a combination of the BabyLM corpus, and a collection of Python code from The Stack. This base model is adapted to two new domains of text: TinyStories, and the Lua programming language, respectively; and then these two models are merged using these two models using spherical linear interpolation. Our exploration aims to provide deeper insights into the stability and transformation of features across typical transfer-learning scenarios using small-scale models and sparse auto-encoders.
Large Language Models have demonstrated impressive performance in many pivotal web applications such as sensor data analysis. However, since LLMs are not designed for time series tasks, simpler models like linear regressions can often achieve comparable performance with far less complexity. In this study, we perform extensive experiments to assess the effectiveness of applying LLMs to key time series tasks, including forecasting, classification, imputation, and anomaly detection. We compare the performance of LLMs against simpler baseline models, such as single-layer linear models and randomly initialized LLMs. Our results reveal that LLMs offer minimal advantages for these core time series tasks and may even distort the temporal structure of the data. In contrast, simpler models consistently outperform LLMs while requiring far fewer parameters. Furthermore, we analyze existing reprogramming techniques and show, through data manifold analysis, that these methods fail to effectively align time series data with language and display pseudo-alignment behaviour in embedding space. Our findings suggest that the performance of LLM-based methods in time series tasks arises from the intrinsic characteristics and structure of time series data, rather than any meaningful alignment with the language model architecture.