Deciding which large language model (LLM) to use is a complex challenge. Pairwise ranking has emerged as a new method for evaluating human preferences for LLMs. This approach entails humans evaluating pairs of model outputs based on a predefined criterion. By collecting these comparisons, a ranking can be constructed using methods such as Elo. However, applying these algorithms as constructed in the context of LLM evaluation introduces several challenges. In this paper, we explore the effectiveness of ranking systems for head-to-head comparisons of LLMs. We formally define a set of fundamental principles for effective ranking and conduct a series of extensive evaluations on the robustness of several ranking algorithms in the context of LLMs. Our analysis uncovers key insights into the factors that affect ranking accuracy and efficiency, offering guidelines for selecting the most appropriate methods based on specific evaluation contexts and resource constraints.
Frontier AI safety policies highlight automation of AI research and development (R&D) by AI agents as an important capability to anticipate. However, there exist few evaluations for AI R&D capabilities, and none that are highly realistic and have a direct comparison to human performance. We introduce RE-Bench (Research Engineering Benchmark, v1), which consists of 7 challenging, open-ended ML research engineering environments and data from 71 8-hour attempts by 61 distinct human experts. We confirm that our experts make progress in the environments given 8 hours, with 82% of expert attempts achieving a non-zero score and 24% matching or exceeding our strong reference solutions. We compare humans to several public frontier models through best-of-k with varying time budgets and agent designs, and find that the best AI agents achieve a score 4x higher than human experts when both are given a total time budget of 2 hours per environment. However, humans currently display better returns to increasing time budgets, narrowly exceeding the top AI agent scores given an 8-hour budget, and achieving 2x the score of the top AI agent when both are given 32 total hours (across different attempts). Qualitatively, we find that modern AI agents possess significant expertise in many ML topics -- e.g. an agent wrote a faster custom Triton kernel than any of our human experts' -- and can generate and test solutions over ten times faster than humans, at much lower cost. We open-source the evaluation environments, human expert data, analysis code and agent trajectories to facilitate future research.
Large Language Model (LLM) agents are rapidly improving to handle increasingly complex web-based tasks. Most of these agents rely on general-purpose, proprietary models like GPT-4 and focus on designing better prompts to improve their planning abilities. However, general-purpose LLMs are not specifically trained to understand specialized web contexts such as HTML, and they often struggle with long-horizon planning. We explore an alternative approach that fine-tunes open-source LLMs using production-scale workflow data collected from over 250 domains corresponding to 6 billion tokens. This simple yet effective approach shows substantial gains over prompting-based agents on existing benchmarks -- ScribeAgent achieves state-of-the-art direct generation performance on Mind2Web and improves the task success rate by 7.3% over the previous best text-only web agents on WebArena. We further perform detailed ablation studies on various fine-tuning design choices and provide insights into LLM selection, training recipes, context window optimization, and effect of dataset sizes.
Language models perform differently across languages. It has been previously suggested that morphological typology may explain some of this variability (Cotterell et al., 2018). We replicate previous analyses and find additional new evidence for a performance gap between agglutinative and fusional languages, where fusional languages, such as English, tend to have better language modeling performance than morphologically more complex languages like Turkish. We then propose and test three possible causes for this performance gap: morphological alignment of tokenizers, tokenization quality, and disparities in dataset sizes and measurement. To test the morphological alignment hypothesis, we present MorphScore, a tokenizer evaluation metric, and supporting datasets for 22 languages. We find some evidence that tokenization quality explains the performance gap, but none for the role of morphological alignment. Instead we find that the performance gap is most reduced when training datasets are of equivalent size across language types, but only when scaled according to the so-called "byte-premium" -- the different encoding efficiencies of different languages and orthographies. These results suggest that no language is harder or easier for a language model to learn on the basis of its morphological typology. Differences in performance can be attributed to disparities in dataset size. These results bear on ongoing efforts to improve performance for low-performing and under-resourced languages.
Large language models (LLMs) deliver impressive results but face challenges from increasing model sizes and computational costs. Structured pruning reduces model size and speeds up inference but often causes uneven degradation across domains, leading to biased performance. To address this, we propose DRPruning, which incorporates distributionally robust optimization to restore balanced performance across domains, along with further improvements to enhance robustness. Experiments in monolingual and multilingual settings show that our method surpasses similarly sized models in pruning and continued pretraining over perplexity, downstream tasks, and instruction tuning. We further provide analysis demonstrating the robustness of our method towards various domains and distribution shifts. Furthermore, our method automatically determines optimal reference losses and data ratios, suggesting potential for broader applications. Our code is available at https://github.com/hexuandeng/DRPruning.
Large Language Models (LLMs) have demonstrated the potential to address some issues within the semiconductor industry. However, they are often general-purpose models that lack the specialized knowledge needed to tackle the unique challenges of this sector, such as the intricate physics and chemistry of semiconductor devices and processes. SemiKong, the first industry-specific LLM for the semiconductor domain, provides a foundation that can be used to develop tailored proprietary models. With SemiKong 1.0, we aim to develop a foundational model capable of understanding etching problems at an expert level. Our key contributions include (a) curating a comprehensive corpus of semiconductor-related texts, (b) creating a foundational model with in-depth semiconductor knowledge, and (c) introducing a framework for integrating expert knowledge, thereby advancing the evaluation process of domain-specific AI models. Through fine-tuning a pre-trained LLM using our curated dataset, we have shown that SemiKong outperforms larger, general-purpose LLMs in various semiconductor manufacturing and design tasks. Our extensive experiments underscore the importance of developing domain-specific LLMs as a foundation for company- or tool-specific proprietary models, paving the way for further research and applications in the semiconductor domain. Code and dataset will be available at https://github.com/aitomatic/semikong
In China, receptionist nurses face overwhelming workloads in outpatient settings, limiting their time and attention for each patient and ultimately reducing service quality. In this paper, we present the Personalized Intelligent Outpatient Reception System (PIORS). This system integrates an LLM-based reception nurse and a collaboration between LLM and hospital information system (HIS) into real outpatient reception setting, aiming to deliver personalized, high-quality, and efficient reception services. Additionally, to enhance the performance of LLMs in real-world healthcare scenarios, we propose a medical conversational data generation framework named Service Flow aware Medical Scenario Simulation (SFMSS), aiming to adapt the LLM to the real-world environments and PIORS settings. We evaluate the effectiveness of PIORS and SFMSS through automatic and human assessments involving 15 users and 15 clinical experts. The results demonstrate that PIORS-Nurse outperforms all baselines, including the current state-of-the-art model GPT-4o, and aligns with human preferences and clinical needs. Further details and demo can be found at https://github.com/FudanDISC/PIORS
Large language models (LLMs) under-perform on low-resource languages due to limited training data. We present a method to efficiently collect text data for low-resource languages from the entire Common Crawl corpus. Our approach, UnifiedCrawl, filters and extracts common crawl using minimal compute resources, yielding mono-lingual datasets much larger than previously available sources. We demonstrate that leveraging this data to fine-tuning multilingual LLMs via efficient adapter methods (QLoRA) significantly boosts performance on the low-resource language, while minimizing VRAM usage. Our experiments show large improvements in language modeling perplexity and an increase in few-shot prompting scores. Our work and released source code provide an affordable approach to improve LLMs for low-resource languages using consumer hardware. Our source code is available here at https://github.com/bethelmelesse/unifiedcrawl.
Graphical User Interface (GUI) grounding plays a crucial role in enhancing the capabilities of Vision-Language Model (VLM) agents. While general VLMs, such as GPT-4V, demonstrate strong performance across various tasks, their proficiency in GUI grounding remains suboptimal. Recent studies have focused on fine-tuning these models specifically for zero-shot GUI grounding, yielding significant improvements over baseline performance. We introduce a visual prompting framework that employs an iterative narrowing mechanism to further improve the performance of both general and fine-tuned models in GUI grounding. For evaluation, we tested our method on a comprehensive benchmark comprising various UI platforms and provided the code to reproduce our results.
Various watermarking methods (``watermarkers'') have been proposed to identify LLM-generated texts; yet, due to the lack of unified evaluation platforms, many critical questions remain under-explored: i) What are the strengths/limitations of various watermarkers, especially their attack robustness? ii) How do various design choices impact their robustness? iii) How to optimally operate watermarkers in adversarial environments? To fill this gap, we systematize existing LLM watermarkers and watermark removal attacks, mapping out their design spaces. We then develop WaterPark, a unified platform that integrates 10 state-of-the-art watermarkers and 12 representative attacks. More importantly, by leveraging WaterPark, we conduct a comprehensive assessment of existing watermarkers, unveiling the impact of various design choices on their attack robustness. We further explore the best practices to operate watermarkers in adversarial environments. We believe our study sheds light on current LLM watermarking techniques while WaterPark serves as a valuable testbed to facilitate future research.
The attention mechanism within the transformer architecture enables the model to weigh and combine tokens based on their relevance to the query. While self-attention has enjoyed major success, it notably treats all queries $q$ in the same way by applying the mapping $V^\top\text{softmax}(Kq)$, where $V,K$ are the value and key embeddings respectively. In this work, we argue that this uniform treatment hinders the ability to control contextual sparsity and relevance. As a solution, we introduce the $\textit{Selective Self-Attention}$ (SSA) layer that augments the softmax nonlinearity with a principled temperature scaling strategy. By controlling temperature, SSA adapts the contextual sparsity of the attention map to the query embedding and its position in the context window. Through theory and experiments, we demonstrate that this alleviates attention dilution, aids the optimization process, and enhances the model's ability to control softmax spikiness of individual queries. We also incorporate temperature scaling for value embeddings and show that it boosts the model's ability to suppress irrelevant/noisy tokens. Notably, SSA is a lightweight method which introduces less than 0.5% new parameters through a weight-sharing strategy and can be fine-tuned on existing LLMs. Extensive empirical evaluations demonstrate that SSA-equipped models achieve a noticeable and consistent accuracy improvement on language modeling benchmarks.
Large language model (LLM) agents show promise in an increasing number of domains. In many proposed applications, it is expected that the agent reasons over accumulated experience presented in an input prompt. We propose the OEDD (Operationalize Experience Despite Distraction) corpus, a human-annotator-validated body of scenarios with pre-scripted agent histories where the agent must make a decision based on disparate experiential information in the presence of a distractor. We evaluate three state-of-the-art LLMs (GPT-3.5 Turbo, GPT-4o, and Gemini 1.5 Pro) using a minimal chain-of-thought prompting strategy and observe that when (1) the input context contains over 1,615 tokens of historical interactions, (2) a crucially decision-informing premise is the rightful conclusion over two disparate environment premises, and (3) a trivial, but distracting red herring fact follows, all LLMs perform worse than random choice at selecting the better of two actions. Our code and test corpus are publicly available at: https://github.com/sonnygeorge/OEDD .
Large language model unlearning aims to remove harmful information that LLMs have learnt to prevent their use for malicious purposes. LLMU and RMU have been proposed as two methods for LLM unlearning, achieving impressive results on unlearning benchmarks. We study in detail the efficacy of these methods by evaluating their impact on general model capabilities on the WMDP benchmark as well as a biology benchmark we create. Our experiments show that RMU generally leads to better preservation of model capabilities, for similar or better unlearning. We further test the robustness of these methods and find that doing 5-shot prompting or rephrasing the question in simple ways can lead to an over ten-fold increase in accuracy on unlearning benchmarks. Finally, we show that training on unrelated data can almost completely recover pre-unlearning performance, demonstrating that these methods fail at truly unlearning. The code is available at: https://github.com/JaiDoshi/Knowledge-Erasure.
Linear Recurrent Neural Networks (LRNNs) such as Mamba, RWKV, GLA, mLSTM, and DeltaNet have emerged as efficient alternatives to Transformers for long sequences. However, both Transformers and LRNNs struggle to perform state-tracking, which may impair performance in tasks such as code evaluation. In one forward pass, current architectures are unable to solve even parity, the simplest state-tracking task, which non-linear RNNs can handle effectively. Recently, Sarrof et al. (2024) demonstrated that the failure of LRNNs like Mamba to solve parity stems from restricting the value range of their diagonal state-transition matrices to $[0, 1]$ and that incorporating negative values can resolve this issue. We extend this result to non-diagonal LRNNs such as DeltaNet. We prove that finite precision LRNNs with state-transition matrices having only positive eigenvalues cannot solve parity, while non-triangular matrices are needed to count modulo $3$. Notably, we also prove that LRNNs can learn any regular language when their state-transition matrices are products of identity minus vector outer product matrices, each with eigenvalues in the range $[-1, 1]$. Our experiments confirm that extending the eigenvalue range of Mamba and DeltaNet to include negative values not only enables them to solve parity but consistently improves their performance on state-tracking tasks. We also show that state-tracking enabled LRNNs can be pretrained stably and efficiently at scale (1.3B parameters), achieving competitive performance on language modeling and showing promise on code and math tasks.
Remote sensing (RS) visual grounding aims to use natural language expression to locate specific objects (in the form of the bounding box or segmentation mask) in RS images, enhancing human interaction with intelligent RS interpretation systems. Early research in this area was primarily based on horizontal bounding boxes (HBBs), but as more diverse RS datasets have become available, tasks involving oriented bounding boxes (OBBs) and segmentation masks have emerged. In practical applications, different targets require different grounding types: HBB can localize an object's position, OBB provides its orientation, and mask depicts its shape. However, existing specialized methods are typically tailored to a single type of RS visual grounding task and are hard to generalize across tasks. In contrast, large vision-language models (VLMs) exhibit powerful multi-task learning capabilities but struggle to handle dense prediction tasks like segmentation. This paper proposes GeoGround, a novel framework that unifies support for HBB, OBB, and mask RS visual grounding tasks, allowing flexible output selection. Rather than customizing the architecture of VLM, our work aims to elegantly support pixel-level visual grounding output through the Text-Mask technique. We define prompt-assisted and geometry-guided learning to enhance consistency across different signals. Experimental results show that GeoGround demonstrates strong performance across four RS visual grounding tasks, matching the performance of specialized methods on multiple benchmarks. Code available at https://github.com/zytx121/GeoGround
Compositional zero-shot learning (CZSL) aims to recognize novel compositions of attributes and objects learned from seen compositions. Previous works disentangle attribute and object by extracting shared and exclusive parts between image pairs sharing the same attribute (object), as well as aligning them with pretrained word embeddings to improve unseen attribute-object recognition. Despite the significant achievements of existing efforts, they are hampered by three limitations: (1) the efficacy of disentanglement is compromised due to the influence of the background and the intricate entanglement of attribute with object in the same parts. (2) existing word embeddings fail to capture complex multimodal semantic information. (3) overconfidence exhibited by existing models in seen compositions hinders their generalization to novel compositions. Being aware of these, we propose a novel framework named Multimodal Large Language Model (MLLM) embeddings and attribute smoothing guided disentanglement (TRIDENT) for CZSL. First, we leverage feature adaptive aggregation modules to mitigate the impact of background, and utilize learnable condition masks to capture multigranularity features for disentanglement. Then, the last hidden states of MLLM are employed as word embeddings for their superior representation capabilities. Moreover, we propose attribute smoothing with auxiliary attributes generated by Large Language Model (LLM) for seen compositions, addressing the issue of overconfidence by encouraging the model to learn more attributes in one given composition. Extensive experiments demonstrate that TRIDENT achieves state-of-the-art performance on three benchmarks.
The Structured Dialogue System, referred to as SuDoSys, is an innovative Large Language Model (LLM)-based chatbot designed to provide psychological counseling. SuDoSys leverages the World Health Organization (WHO)'s Problem Management Plus (PM+) guidelines to deliver stage-aware multi-turn dialogues. Existing methods for employing an LLM in multi-turn psychological counseling typically involve direct fine-tuning using generated dialogues, often neglecting the dynamic stage shifts of counseling sessions. Unlike previous approaches, SuDoSys considers the different stages of counseling and stores essential information throughout the counseling process, ensuring coherent and directed conversations. The system employs an LLM, a stage-aware instruction generator, a response unpacker, a topic database, and a stage controller to maintain dialogue flow. In addition, we propose a novel technique that simulates counseling clients to interact with the evaluated system and evaluate its performance automatically. When assessed using both objective and subjective evaluations, SuDoSys demonstrates its effectiveness in generating logically coherent responses. The system's code and program scripts for evaluation are open-sourced.
Various linear complexity models, such as Linear Transformer (LinFormer), State Space Model (SSM), and Linear RNN (LinRNN), have been proposed to replace the conventional softmax attention in Transformer structures. However, the optimal design of these linear models is still an open question. In this work, we attempt to answer this question by finding the best linear approximation to softmax attention from a theoretical perspective. We start by unifying existing linear complexity models as the linear attention form and then identify three conditions for the optimal linear attention design: 1) Dynamic memory ability; 2) Static approximation ability; 3) Least parameter approximation. We find that none of the current linear models meet all three conditions, resulting in suboptimal performance. Instead, we propose Meta Linear Attention (MetaLA) as a solution that satisfies these conditions. Our experiments on Multi-Query Associative Recall (MQAR) task, language modeling, image classification, and Long-Range Arena (LRA) benchmark demonstrate that MetaLA is more effective than the existing linear models.
Interactions with online Large Language Models raise privacy issues where providers can gather sensitive information about users and their companies from the prompts. While textual prompts can be sanitized using Differential Privacy, we show that it is difficult to anticipate the performance of an LLM on such sanitized prompt. Poor performance has clear monetary consequences for LLM services charging on a pay-per-use model as well as great amount of computing resources wasted. To this end, we propose a middleware architecture leveraging a Small Language Model to predict the utility of a given sanitized prompt before it is sent to the LLM. We experimented on a summarization task and a translation task to show that our architecture helps prevent such resource waste for up to 20% of the prompts. During our study, we also reproduced experiments from one of the most cited paper on text sanitization using DP and show that a potential performance-driven implementation choice dramatically changes the output while not being explicitly acknowledged in the paper.
The vision tokens in multimodal large language models usually exhibit significant spatial and temporal redundancy and take up most of the input tokens, which harms their inference efficiency. To solve this problem, some recent works were introduced to drop the unimportant tokens during inference where the importance of each token is decided only by the information in either the vision encoding stage or the prefilling stage. In this paper, we propose Multi-stage Token Dropping (MustDrop) to measure the importance of each token from the whole lifecycle, including the vision encoding stage, prefilling stage, and decoding stage. Concretely, in the visual encoding stage, MustDrop merges spatially adjacent tokens with high similarity, and establishes a key token set to retain the most vision-critical tokens, preventing them from being discarded in later stages. In the prefilling stage, MustDrop further compresses vision tokens by the guidance of text semantics, with a dual-attention filtering strategy. In the decoding stage, an output-aware cache policy is proposed to further reduce the size of the KV cache. By leveraging tailored strategies in the multi-stage process, MustDrop can more precisely recognize the important and redundant tokens, thus achieving an optimal balance between performance and efficiency. For instance, MustDrop reduces about 88.5\% FLOPs on LLaVA with a compression ratio of 92.2\% while maintaining comparable accuracy. Our codes are available at \url{https://github.com/liuting20/MustDrop}.