Large Language Models (LLMs) show promising potential in Software Engineering, especially for code-related tasks like code completion and code generation. LLMs' evaluation is generally centred around general metrics computed over benchmarks. While painting a macroscopic view of the benchmarks and of the LLMs' capacity, it is unclear how each programming task in these benchmarks assesses the capabilities of the LLMs. In particular, the difficulty level of the tasks in the benchmarks is not reflected in the score used to report the performance of the model. Yet, a model achieving a 90% score on a benchmark of predominantly easy tasks is likely less capable than a model achieving a 90% score on a benchmark containing predominantly difficult tasks. This paper devises a framework, HardEval, for assessing task difficulty for LLMs and crafting new tasks based on identified hard tasks. The framework uses a diverse array of prompts for a single task across multiple LLMs to obtain a difficulty score for each task of a benchmark. Using two code generation benchmarks, HumanEval+ and ClassEval, we show that HardEval can reliably identify the hard tasks within those benchmarks, highlighting that only 21% of HumanEval+ and 27% of ClassEval tasks are hard for LLMs. Through our analysis of task difficulty, we also characterize 6 practical hard task topics which we used to generate new hard tasks. Orthogonal to current benchmarking evaluation efforts, HardEval can assist researchers and practitioners in fostering better assessments of LLMs. The difficulty score can be used to identify hard tasks within existing benchmarks. This, in turn, can be leveraged to generate more hard tasks centred around specific topics either for evaluation or improvement of LLMs. HardEval generalistic approach can be applied to other domains such as code completion or Q/A.

Large language models (LLMs) (Hadi et al., 2023; Brown et al., 2020; OpenAI, 2023; Touvron et al., 2023a,b; Scao et al., 2022; Reid et al., 2024) have emerged as a dominant paradigm in natural language processing, achieving state-of-the-art performance across various tasks. A key principle, the Scaling Law (Kaplan et al., 2020), suggests that LLMs exhibit emergent abilities as model size increases, enhancing their capacity to understand context and handle long sequences (Xiong et al., 2023). This capacity growth allows LLMs to generate coherent and contextually accurate responses and enables various downstream applications, such as document summarization (Zhang et al., 2019, 2024a), code generation (Chen et al., 2021b), and conversational AI (Bordes et al., 2016; OpenAI, 2022),. Despite their success in various applications, the generation of LLMs incurs significant expenses, which escalate with increasing model size and sequence length. Notably, both the training (Strubell et al., 2020; Hoffmann et al., 2022; Dong et al., 2024a) and inference (Ainslie et al., 2023) stages involve frequent generation by LLMs, further contributing to these costs. Consequently, efficient LLMs have gained popularity in recent years (Hu et al., 2021; Wan et al., 2023). To address these challenges, quantization (Frantar et al., 2022; Lin et al., 2024; Dettmers et al., 2024; Xu et al., 2023) and pruning methods (Frankle and Carbin, 2018; Blalock et al., 2020) are employed to reduce model size. Additionally, managing long sequences presents another cost due to the transformer attention mechanism.

While large language models (LLMs) are extensively used, there are raising concerns regarding privacy, security, and copyright due to their opaque training data, which brings the problem of detecting pre-training data on the table. Current solutions to this problem leverage techniques explored in machine learning privacy such as Membership Inference Attacks (MIAs), which heavily depend on LLMs' capability of verbatim memorization. However, this reliance presents challenges, especially given the vast amount of training data and the restricted number of effective training epochs. In this paper, we propose an adaptive pre-training data detection method which alleviates this reliance and effectively amplify the identification. Our method adaptively locates \textit{surprising tokens} of the input. A token is surprising to a LLM if the prediction on the token is "certain but wrong", which refers to low Shannon entropy of the probability distribution and low probability of the ground truth token at the same time. By using the prediction probability of surprising tokens to measure \textit{surprising}, the detection method is achieved based on the simple hypothesis that seeing seen data is less surprising for the model compared with seeing unseen data. The method can be applied without any access to the the pre-training data corpus or additional training like reference models. Our approach exhibits a consistent enhancement compared to existing methods in diverse experiments conducted on various benchmarks and models, achieving a maximum improvement of 29.5\%. We also introduce a new benchmark Dolma-Book developed upon a novel framework, which employs book data collected both before and after model training to provide further evaluation.

Predicting influencers' views and public sentiment on social media is crucial for anticipating societal trends and guiding strategic responses. This study introduces a novel computational framework to predict opinion leaders' perspectives and the emotive reactions of the populace, addressing the inherent challenges posed by the unstructured, context-sensitive, and heterogeneous nature of online communication. Our research introduces an innovative module that starts with the automatic 5W1H (Where, Who, When, What, Why, and How) questions formulation engine, tailored to emerging news stories and trending topics. We then build a total of 60 anonymous opinion leader agents in six domains and realize the views generation based on an enhanced large language model (LLM) coupled with retrieval-augmented generation (RAG). Subsequently, we synthesize the potential views of opinion leaders and predicted the emotional responses to different events. The efficacy of our automated 5W1H module is corroborated by an average GPT-4 score of 8.83/10, indicative of high fidelity. The influencer agents exhibit a consistent performance, achieving an average GPT-4 rating of 6.85/10 across evaluative metrics. Utilizing the 'Russia-Ukraine War' as a case study, our methodology accurately foresees key influencers' perspectives and aligns emotional predictions with real-world sentiment trends in various domains.

While large language models (LLMs) have demonstrated impressive capabilities across various natural language processing tasks by acquiring rich factual knowledge from their broad training data, their ability to synthesize and logically reason with this knowledge in complex ways remains underexplored. In this work, we present a systematic evaluation of state-of-the-art LLMs' complex logical reasoning abilities through a novel benchmark of automatically generated complex reasoning questions over general domain and biomedical knowledge graphs. Our extensive experiments, employing diverse in-context learning techniques, reveal that LLMs excel at reasoning over general world knowledge but face significant challenges with specialized domain-specific knowledge. We find that prompting with explicit Chain-of-Thought demonstrations can substantially improve LLM performance on complex logical reasoning tasks with diverse logical operations. Interestingly, our controlled evaluations uncover an asymmetry where LLMs display proficiency at set union operations, but struggle considerably with set intersections - a key building block of logical reasoning. To foster further work, we will publicly release our evaluation benchmark and code.

We present two studies exploring how blind or low-vision (BLV) family members engage with their sighted children's artwork, strategies to support understanding and interpretation, and the potential role of technology, such as AI, therein. Our first study involved 14 BLV individuals, and the second included five groups of BLV individuals with their children. Through semi-structured interviews with AI descriptions of children's artwork and multi-sensory design probes, we found that BLV family members value artwork engagement as a bonding opportunity, preferring the child's storytelling and interpretation over other nonvisual representations. Additionally, despite some inaccuracies, BLV family members felt that AI-generated descriptions could facilitate dialogue with their children and aid self-guided art discovery. We close with specific design considerations for supporting artwork engagement in mixed visual-ability families, including enabling artwork access through various methods, supporting children's corrections of AI output, and distinctions in context vs. content and interpretation vs. description of children's artwork.

Automated soft moderation systems are unable to ascertain if a post supports or refutes a false claim, resulting in a large number of contextual false positives. This limits their effectiveness, for example undermining trust in health experts by adding warnings to their posts or resorting to vague warnings instead of granular fact-checks, which result in desensitizing users. In this paper, we propose to incorporate stance detection into existing automated soft-moderation pipelines, with the goal of ruling out contextual false positives and providing more precise recommendations for social media content that should receive warnings. We develop a textual deviation task called Contrastive Textual Deviation (CTD) and show that it outperforms existing stance detection approaches when applied to soft moderation.We then integrate CTD into the stateof-the-art system for automated soft moderation Lambretta, showing that our approach can reduce contextual false positives from 20% to 2.1%, providing another important building block towards deploying reliable automated soft moderation tools on social media.

We introduce WebApp1K, a practical code-generation benchmark to measure LLM ability to develop web apps. This benchmark aims to calibrate LLM output and aid the models to progressively improve code correctness and functionality. The benchmark is lightweight and easy to run. We present the initial version of WebApp1K, and share our findings of running the benchmark against the latest frontier LLMs. First, open source LLMs deliver impressive performance, closely trailing behind GPT-4o and Claude 3.5. Second, model size has strong correlation with code correctness. Third, no prompting techniques have been found to lift performance either universally to all models, or significantly to a single model.

Recently, autonomous agents built on large language models (LLMs) have experienced significant development and are being deployed in real-world applications. These agents can extend the base LLM's capabilities in multiple ways. For example, a well-built agent using GPT-3.5-Turbo as its core can outperform the more advanced GPT-4 model by leveraging external components. More importantly, the usage of tools enables these systems to perform actions in the real world, moving from merely generating text to actively interacting with their environment. Given the agents' practical applications and their ability to execute consequential actions, it is crucial to assess potential vulnerabilities. Such autonomous systems can cause more severe damage than a standalone language model if compromised. While some existing research has explored harmful actions by LLM agents, our study approaches the vulnerability from a different perspective. We introduce a new type of attack that causes malfunctions by misleading the agent into executing repetitive or irrelevant actions. We conduct comprehensive evaluations using various attack methods, surfaces, and properties to pinpoint areas of susceptibility. Our experiments reveal that these attacks can induce failure rates exceeding 80\% in multiple scenarios. Through attacks on implemented and deployable agents in multi-agent scenarios, we accentuate the realistic risks associated with these vulnerabilities. To mitigate such attacks, we propose self-examination detection methods. However, our findings indicate these attacks are difficult to detect effectively using LLMs alone, highlighting the substantial risks associated with this vulnerability.

Recently, vision-language instruct-tuning models have made significant progress due to their more comprehensive understanding of the world. In this work, we discovered that large-scale 3D parallel training on those models leads to an imbalanced computation load across different devices. The vision and language parts are inherently heterogeneous: their data distribution and model architecture differ significantly, which affects distributed training efficiency. We rebalanced the computational loads from data, model, and memory perspectives to address this issue, achieving more balanced computation across devices. These three components are not independent but are closely connected, forming an omniverse balanced training framework. Specifically, for the data, we grouped instances into new balanced mini-batches within and across devices. For the model, we employed a search-based method to achieve a more balanced partitioning. For memory optimization, we adaptively adjusted the re-computation strategy for each partition to utilize the available memory fully. We conducted extensive experiments to validate the effectiveness of our method. Compared with the open-source training code of InternVL-Chat, we significantly reduced GPU days, achieving about 1.8x speed-up. Our method's efficacy and generalizability were further demonstrated across various models and datasets. Codes will be released at https://github.com/ModelTC/OmniBal.