Despite recent advancements in Large Language Models (LLMs), their performance on tasks involving long contexts remains sub-optimal. In-Context Learning (ICL) with few-shot examples may be an appealing solution to enhance LLM performance in this scenario; However, naively adding ICL examples with long context introduces challenges, including substantial token overhead added for each few-shot example and context mismatch between the demonstrations and the target query. In this work, we propose to automatically generate few-shot examples for long context QA tasks by recycling contexts. Specifically, given a long input context (1-3k tokens) and a query, we generate additional query-output pairs from the given context as few-shot examples, while introducing the context only once. This ensures that the demonstrations are leveraging the same context as the target query while only adding a small number of tokens to the prompt. We further enhance each demonstration by instructing the model to explicitly identify the relevant paragraphs before the answer, which improves performance while providing fine-grained attribution to the answer source. We apply our method on multiple LLMs and obtain substantial improvements (+23\% on average across models) on various QA datasets with long context, especially when the answer lies within the middle of the context. Surprisingly, despite introducing only single-hop ICL examples, LLMs also successfully generalize to multi-hop long-context QA using our approach.

The United States and China will play an important role in navigating safety and security challenges relating to advanced artificial intelligence. We sought to better understand how experts in each country describe safety and security threats from advanced artificial intelligence, extreme risks from AI, and the potential for international cooperation. Specifically, we compiled publicly-available statements from major technical and policy leaders in both the United States and China. We focused our analysis on advanced forms of artificial intelligence, such as artificial general intelligence (AGI), that may have the most significant impacts on national and global security. Experts in both countries expressed concern about risks from AGI, risks from intelligence explosions, and risks from AI systems that escape human control. Both countries have also launched early efforts designed to promote international cooperation around safety standards and risk management practices. Notably, our findings only reflect information from publicly available sources. Nonetheless, our findings can inform policymakers and researchers about the state of AI discourse in the US and China. We hope such work can contribute to policy discussions around advanced AI, its global security threats, and potential international dialogues or agreements to mitigate such threats.

What makes people 'click' on a first date and become mutually attracted to one another? While understanding and predicting the dynamics of romantic interactions used to be exclusive to human judgment, we show that Large Language Models (LLMs) can detect romantic attraction during brief getting-to-know-you interactions. Examining data from 964 speed dates, we show that ChatGPT (and Claude 3) can predict both objective and subjective indicators of speed dating success (r=0.12-0.23). ChatGPT's predictions of actual matching (i.e., the exchange of contact information) were not only on par with those of human judges who had access to the same information but incremental to speed daters' own predictions. While some of the variance in ChatGPT's predictions can be explained by common content dimensions (such as the valence of the conversations) the fact that there remains a substantial proportion of unexplained variance suggests that ChatGPT also picks up on conversational dynamics. In addition, ChatGPT's judgments showed substantial overlap with those made by the human observers (mean r=0.29), highlighting similarities in their representation of romantic attraction that is, partially, independent of accuracy.

Ensuring the general efficacy and goodness for human beings from medical large language models (LLM) before real-world deployment is crucial. However, a widely accepted and accessible evaluation process for medical LLM, especially in the Chinese context, remains to be established. In this work, we introduce "MedBench", a comprehensive, standardized, and reliable benchmarking system for Chinese medical LLM. First, MedBench assembles the currently largest evaluation dataset (300,901 questions) to cover 43 clinical specialties and performs multi-facet evaluation on medical LLM. Second, MedBench provides a standardized and fully automatic cloud-based evaluation infrastructure, with physical separations for question and ground truth. Third, MedBench implements dynamic evaluation mechanisms to prevent shortcut learning and answer remembering. Applying MedBench to popular general and medical LLMs, we observe unbiased, reproducible evaluation results largely aligning with medical professionals' perspectives. This study establishes a significant foundation for preparing the practical applications of Chinese medical LLMs. MedBench is publicly accessible at https://medbench.opencompass.org.cn.

Synthetic long-context LLM benchmarks (e.g., "needle-in-the-haystack") test only surface-level retrieval capabilities, but how well can long-context LLMs retrieve, synthesize, and reason over information across book-length inputs? We address this question by creating NoCha, a dataset of 1,001 minimally different pairs of true and false claims about 67 recently-published English fictional books, written by human readers of those books. In contrast to existing long-context benchmarks, our annotators confirm that the largest share of pairs in NoCha require global reasoning over the entire book to verify. Our experiments show that while human readers easily perform this task, it is enormously challenging for all ten long-context LLMs that we evaluate: no open-weight model performs above random chance (despite their strong performance on synthetic benchmarks), while GPT-4o achieves the highest accuracy at 55.8%. Further analysis reveals that (1) on average, models perform much better on pairs that require only sentence-level retrieval vs. global reasoning; (2) model-generated explanations for their decisions are often inaccurate even for correctly-labeled claims; and (3) models perform substantially worse on speculative fiction books that contain extensive world-building. The methodology proposed in NoCha allows for the evolution of the benchmark dataset and the easy analysis of future models.

Large language models (LLMs) for code are typically trained to align with natural language instructions to closely follow their intentions and requirements. However, in many practical scenarios, it becomes increasingly challenging for these models to navigate the intricate boundary between helpfulness and safety, especially against highly complex yet potentially malicious instructions. In this work, we introduce INDICT: a new framework that empowers LLMs with Internal Dialogues of Critiques for both safety and helpfulness guidance. The internal dialogue is a dual cooperative system between a safety-driven critic and a helpfulness-driven critic. Each critic provides analysis against the given task and corresponding generated response, equipped with external knowledge queried through relevant code snippets and tools like web search and code interpreter. We engage the dual critic system in both code generation stage as well as code execution stage, providing preemptive and post-hoc guidance respectively to LLMs. We evaluated INDICT on 8 diverse tasks across 8 programming languages from 5 benchmarks, using LLMs from 7B to 70B parameters. We observed that our approach can provide an advanced level of critiques of both safety and helpfulness analysis, significantly improving the quality of output codes ($+10\%$ absolute improvements in all models).

We present a novel extension to Retrieval Augmented Generation with the goal of mitigating factual inaccuracies in the output of large language models. Specifically, our method draws on the cognitive linguistic theory of frame semantics for the indexing and retrieval of factual information relevant to helping large language models answer queries. We conduct experiments to demonstrate the effectiveness of this method both in terms of retrieval effectiveness and in terms of the relevance of the frames and frame relations automatically generated. Our results show that this novel mechanism of Frame Semantic-based retrieval, designed to improve Retrieval Augmented Generation (FS-RAG), is effective and offers potential for providing data-driven insights into frame semantics theory. We provide open access to our program code and prompts.

The context window within a transformer provides a form of active memory for the current task, which can be useful for few-shot learning and conditional generation, both which depend heavily on previous context tokens. However, as the context length grows, the computational cost increases quadratically. Recent works have shown that saving a few initial tokens along with a fixed-sized sliding window leads to stable streaming generation with linear complexity in transformer-based Large Language Models (LLMs). However, they make suboptimal use of the fixed window by naively evicting all tokens unconditionally from the key-value (KV) cache once they reach the end of the window, resulting in tokens being forgotten and no longer able to affect subsequent predictions. To overcome this limitation, we propose a novel mechanism for storing longer sliding window contexts with the same total cache size by keeping separate cascading sub-cache buffers whereby each subsequent buffer conditionally accepts a fraction of the relatively more important tokens evicted from the previous buffer. Our method results in a dynamic KV cache that can store tokens from the more distant past than a fixed, static sliding window approach. Our experiments show improvements of 5.6% on long context generation (LongBench), 1.2% in streaming perplexity (PG19), and 0.6% in language understanding (MMLU STEM) using LLMs given the same fixed cache size. Additionally, we provide an efficient implementation that improves the KV cache latency from 1.33ms per caching operation to 0.54ms, a 59% speedup over previous work.

Personality is a fundamental construct in psychology, reflecting an individual's behavior, thinking, and emotional patterns. Previous researches have made some progress in personality detection, primarily by utilizing the whole text to predict personality. However, these studies generally tend to overlook psychological knowledge: they rarely apply the well-established correlations between emotion regulation and personality. Based on this, we propose a new personality detection method called EERPD. This method introduces the use of emotion regulation, a psychological concept highly correlated with personality, for personality prediction. By combining this feature with emotion features, it retrieves few-shot examples and provides process CoTs for inferring labels from text. This approach enhances the understanding of LLM for personality within text and improves the performance in personality detection. Experimental results demonstrate that EERPD significantly enhances the accuracy and robustness of personality detection, outperforming previous SOTA by 15.05/4.29 in average F1 on the two benchmark datasets.

AI Generated Text (AIGT) detectors are developed with texts from humans and LLMs of common tasks. Despite the diversity of plausible prompt choices, these datasets are generally constructed with a limited number of prompts. The lack of prompt variation can introduce prompt-specific shortcut features that exist in data collected with the chosen prompt, but do not generalize to others. In this paper, we analyze the impact of such shortcuts in AIGT detection. We propose Feedback-based Adversarial Instruction List Optimization (FAILOpt), an attack that searches for instructions deceptive to AIGT detectors exploiting prompt-specific shortcuts. FAILOpt effectively drops the detection performance of the target detector, comparable to other attacks based on adversarial in-context examples. We also utilize our method to enhance the robustness of the detector by mitigating the shortcuts. Based on the findings, we further train the classifier with the dataset augmented by FAILOpt prompt. The augmented classifier exhibits improvements across generation models, tasks, and attacks. Our code will be available at https://github.com/zxcvvxcz/FAILOpt.