AgentOrca: A Dual-System Framework to Evaluate Language Agents on Operational Routine and Constraint Adherence

As language agents progressively automate critical tasks across domains, their ability to operate within operational constraints and safety protocols becomes essential. While extensive research has demonstrated these agents' effectiveness in downstream task completion, their reliability in following operational procedures and constraints remains largely unexplored. To this end, we present AgentOrca, a dual-system framework for evaluating language agents' compliance with operational constraints and routines. Our framework encodes action constraints and routines through both natural language prompts for agents and corresponding executable code serving as ground truth for automated verification. Through an automated pipeline of test case generation and evaluation across five real-world domains, we quantitatively assess current language agents' adherence to operational constraints. Our findings reveal notable performance gaps among state-of-the-art models, with large reasoning models like o1 demonstrating superior compliance while others show significantly lower performance, particularly when encountering complex constraints or user persuasion attempts.

Creative and Context-Aware Translation of East Asian Idioms with GPT-4

As a type of figurative language, an East Asian idiom condenses rich cultural background into only a few characters. Translating such idioms is challenging for human translators, who often resort to choosing a context-aware translation from an existing list of candidates. However, compiling a dictionary of candidate translations demands much time and creativity even for expert translators. To alleviate such burden, we evaluate if GPT-4 can help generate high-quality translations. Based on automatic evaluations of faithfulness and creativity, we first identify Pareto-optimal prompting strategies that can outperform translation engines from Google and DeepL. Then, at a low cost, our context-aware translations can achieve far more high-quality translations per idiom than the human baseline. We open-source all code and data to facilitate further research.

Can Editing LLMs Inject Harm?

Knowledge editing techniques have been increasingly adopted to efficiently correct the false or outdated knowledge in Large Language Models (LLMs), due to the high cost of retraining from scratch. Meanwhile, one critical but under-explored question is: can knowledge editing be used to inject harm into LLMs? In this paper, we propose to reformulate knowledge editing as a new type of safety threat for LLMs, namely Editing Attack, and conduct a systematic investigation with a newly constructed dataset EditAttack. Specifically, we focus on two typical safety risks of Editing Attack including Misinformation Injection and Bias Injection. For the risk of misinformation injection, we first categorize it into commonsense misinformation injection and long-tail misinformation injection. Then, we find that editing attacks can inject both types of misinformation into LLMs, and the effectiveness is particularly high for commonsense misinformation injection. For the risk of bias injection, we discover that not only can biased sentences be injected into LLMs with high effectiveness, but also one single biased sentence injection can cause a high bias increase in general outputs of LLMs, which are even highly irrelevant to the injected sentence, indicating a catastrophic impact on the overall fairness of LLMs. Then, we further illustrate the high stealthiness of editing attacks, measured by their impact on the general knowledge and reasoning capacities of LLMs, and show the hardness of defending editing attacks with empirical evidence. Our discoveries demonstrate the emerging misuse risks of knowledge editing techniques on compromising the safety alignment of LLMs.

MMSci: A Multimodal Multi-Discipline Dataset for PhD-Level Scientific Comprehension

The rapid advancement of Large Language Models (LLMs) and Large Multimodal Models (LMMs) has heightened the demand for AI-based scientific assistants capable of understanding scientific articles and figures. Despite progress, there remains a significant gap in evaluating models' comprehension of professional, graduate-level, and even PhD-level scientific content. Current datasets and benchmarks primarily focus on relatively simple scientific tasks and figures, lacking comprehensive assessments across diverse advanced scientific disciplines. To bridge this gap, we collected a multimodal, multidisciplinary dataset from open-access scientific articles published in Nature Communications journals. This dataset spans 72 scientific disciplines, ensuring both diversity and quality. We created benchmarks with various tasks and settings to comprehensively evaluate LMMs' capabilities in understanding scientific figures and content. Our evaluation revealed that these tasks are highly challenging: many open-source models struggled significantly, and even GPT-4V and GPT-4o faced difficulties. We also explored using our dataset as training resources by constructing visual instruction-following data, enabling the 7B LLaVA model to achieve performance comparable to GPT-4V/o on our benchmark. Additionally, we investigated the use of our interleaved article texts and figure images for pre-training LMMs, resulting in improvements on the material generation task. The source dataset, including articles, figures, constructed benchmarks, and visual instruction-following data, is open-sourced.

Finetuned Multimodal Language Models Are High-Quality Image-Text Data Filters

We propose a novel framework for filtering image-text data by leveraging fine-tuned Multimodal Language Models (MLMs). Our approach outperforms predominant filtering methods (e.g., CLIPScore) via integrating the recent advances in MLMs. We design four distinct yet complementary metrics to holistically measure the quality of image-text data. A new pipeline is established to construct high-quality instruction data for fine-tuning MLMs as data filters. Comparing with CLIPScore, our MLM filters produce more precise and comprehensive scores that directly improve the quality of filtered data and boost the performance of pre-trained models. We achieve significant improvements over CLIPScore on popular foundation models (i.e., CLIP and BLIP2) and various downstream tasks. Our MLM filter can generalize to different models and tasks, and be used as a drop-in replacement for CLIPScore. An additional ablation study is provided to verify our design choices for the MLM filter.

Large Language Models as Zero-shot Dialogue State Tracker through Function Calling

Large language models (LLMs) are increasingly prevalent in conversational systems due to their advanced understanding and generative capabilities in general contexts. However, their effectiveness in task-oriented dialogues (TOD), which requires not only response generation but also effective dialogue state tracking (DST) within specific tasks and domains, remains less satisfying. In this work, we propose a novel approach FnCTOD for solving DST with LLMs through function calling. This method improves zero-shot DST, allowing adaptation to diverse domains without extensive data collection or model tuning. Our experimental results demonstrate that our approach achieves exceptional performance with both modestly sized open-source and also proprietary LLMs: with in-context prompting it enables various 7B or 13B parameter models to surpass the previous state-of-the-art (SOTA) achieved by ChatGPT, and improves ChatGPT's performance beating the SOTA by 5.6% Avg. JGA. Individual model results for GPT-3.5 and GPT-4 are boosted by 4.8% and 14%, respectively. We also show that by fine-tuning on a small collection of diverse task-oriented dialogues, we can equip modestly sized models, specifically a 13B parameter LLaMA2-Chat model, with function-calling capabilities and DST performance comparable to ChatGPT while maintaining their chat capabilities. We plan to open-source experimental code and model.

GPT-4V as a Generalist Evaluator for Vision-Language Tasks

Automatically evaluating vision-language tasks is challenging, especially when it comes to reflecting human judgments due to limitations in accounting for fine-grained details. Although GPT-4V has shown promising results in various multi-modal tasks, leveraging GPT-4V as a generalist evaluator for these tasks has not yet been systematically explored. We comprehensively validate GPT-4V's capabilities for evaluation purposes, addressing tasks ranging from foundational image-to-text and text-to-image synthesis to high-level image-to-image translations and multi-images to text alignment. We employ two evaluation methods, single-answer grading and pairwise comparison, using GPT-4V. Notably, GPT-4V shows promising agreement with humans across various tasks and evaluation methods, demonstrating immense potential for multi-modal LLMs as evaluators. Despite limitations like restricted visual clarity grading and real-world complex reasoning, its ability to provide human-aligned scores enriched with detailed explanations is promising for universal automatic evaluator.

Do you really follow me? Adversarial Instructions for Evaluating the Robustness of Large Language Models

Large Language Models (LLMs) have shown remarkable proficiency in following instructions, making them valuable in customer-facing applications. However, their impressive capabilities also raise concerns about the amplification of risks posed by adversarial instructions, which can be injected into the model input by third-party attackers to manipulate LLMs' original instructions and prompt unintended actions and content. Therefore, it is crucial to understand LLMs' ability to accurately discern which instructions to follow to ensure their safe deployment in real-world scenarios. In this paper, we propose a pioneering benchmark for automatically evaluating the robustness of LLMs against adversarial instructions. The objective of this benchmark is to quantify the extent to which LLMs are influenced by injected adversarial instructions and assess their ability to differentiate between these adversarial instructions and original user instructions. Through experiments conducted with state-of-the-art instruction-following LLMs, we uncover significant limitations in their robustness against adversarial instruction attacks. Furthermore, our findings indicate that prevalent instruction-tuned models are prone to being overfitted to follow any instruction phrase in the prompt without truly understanding which instructions should be followed. This highlights the need to address the challenge of training models to comprehend prompts instead of merely following instruction phrases and completing the text.

Augmenting Language Models with Long-Term Memory

Existing large language models (LLMs) can only afford fix-sized inputs due to the input length limit, preventing them from utilizing rich long-context information from past inputs. To address this, we propose a framework, Language Models Augmented with Long-Term Memory (LongMem), which enables LLMs to memorize long history. We design a novel decoupled network architecture with the original backbone LLM frozen as a memory encoder and an adaptive residual side-network as a memory retriever and reader. Such a decoupled memory design can easily cache and update long-term past contexts for memory retrieval without suffering from memory staleness. Enhanced with memory-augmented adaptation training, LongMem can thus memorize long past context and use long-term memory for language modeling. The proposed memory retrieval module can handle unlimited-length context in its memory bank to benefit various downstream tasks. Typically, LongMem can enlarge the long-form memory to 65k tokens and thus cache many-shot extra demonstration examples as long-form memory for in-context learning. Experiments show that our method outperforms strong long-context models on ChapterBreak, a challenging long-context modeling benchmark, and achieves remarkable improvements on memory-augmented in-context learning over LLMs. The results demonstrate that the proposed method is effective in helping language models to memorize and utilize long-form contents. Our code is open-sourced at https://aka.ms/LongMem.

STEPS: A Benchmark for Order Reasoning in Sequential Tasks

Various human activities can be abstracted into a sequence of actions in natural text, i.e. cooking, repairing, manufacturing, etc. Such action sequences heavily depend on the executing order, while disorder in action sequences leads to failure of further task execution by robots or AI agents. Therefore, to verify the order reasoning capability of current neural models in sequential tasks, we propose a challenging benchmark , named STEPS. STEPS involves two subtask settings, focusing on determining the rationality of given next step in recipes and selecting the reasonable step from the multi-choice question, respectively. We describe the data construction and task formulations, and benchmark most of significant Large Language Models (LLMs). The experimental results demonstrate 1) The commonsense reasoning of action orders in sequential tasks are challenging to resolve via zero-shot prompting or few-shot in-context learning for LLMs; 2) Prompting method still significantly lags behind tuning-based method on STEPS.