While recent works (e.g. o1, DeepSeek R1) have demonstrated great promise of using long Chain-of-Thought (CoT) to improve reasoning capabilities of language models, scaling it up during test-time is challenging due to inefficient memory usage -- intermediate computations accumulate indefinitely in context even no longer needed for future thoughts. We propose PENCIL, which incorporates a reduction mechanism into the autoregressive generation process, allowing the model to recursively clean up intermediate thoughts based on patterns learned from training. With this reduction mechanism, PENCIL significantly reduces the maximal context length required during generation, and thus can generate longer thoughts with limited memory, solving larger-scale problems given more thinking time. For example, we demonstrate PENCIL achieves 97\% accuracy on the challenging Einstein's puzzle -- a task even large models like GPT-4 struggle with -- using only a small 25M-parameter transformer with 2048 context length. Theoretically, we prove PENCIL can perform universal space-efficient computation by simulating Turing machines with optimal time and space complexity, and thus can solve arbitrary computational tasks that would otherwise be intractable given context window constraints.

LLMs often adopt an assertive language style also when making false claims. Such ``overconfident hallucinations'' mislead users and erode trust. Achieving the ability to express in language the actual degree of uncertainty around a claim is therefore of great importance. We find that ``verbal uncertainty'' is governed by a single linear feature in the representation space of LLMs, and show that this has only moderate correlation with the actual ``semantic uncertainty'' of the model. We apply this insight and show that (1) the mismatch between semantic and verbal uncertainty is a better predictor of hallucinations than semantic uncertainty alone and (2) we can intervene on verbal uncertainty at inference time and reduce hallucinations on short-form answers, achieving an average relative reduction of 32%.

Hallucinations in generative AI, particularly in Large Language Models (LLMs), pose a significant challenge to the reliability of multilingual applications. Existing benchmarks for hallucination detection focus primarily on English and a few widely spoken languages, lacking the breadth to assess inconsistencies in model performance across diverse linguistic contexts. To address this gap, we introduce Poly-FEVER, a large-scale multilingual fact verification benchmark specifically designed for evaluating hallucination detection in LLMs. Poly-FEVER comprises 77,973 labeled factual claims spanning 11 languages, sourced from FEVER, Climate-FEVER, and SciFact. It provides the first large-scale dataset tailored for analyzing hallucination patterns across languages, enabling systematic evaluation of LLMs such as ChatGPT and the LLaMA series. Our analysis reveals how topic distribution and web resource availability influence hallucination frequency, uncovering language-specific biases that impact model accuracy. By offering a multilingual benchmark for fact verification, Poly-FEVER facilitates cross-linguistic comparisons of hallucination detection and contributes to the development of more reliable, language-inclusive AI systems. The dataset is publicly available to advance research in responsible AI, fact-checking methodologies, and multilingual NLP, promoting greater transparency and robustness in LLM performance. The proposed Poly-FEVER is available at: https://huggingface.co/datasets/HanzhiZhang/Poly-FEVER.

Large language models (LLMs) have been one of the most important discoveries in machine learning in recent years. LLM-based artificial intelligence (AI) assistants, such as ChatGPT, have consistently attracted the attention from researchers, investors, and the general public, driving the rapid growth of this industry. With the frequent introduction of new LLMs to the market, it becomes increasingly difficult to differentiate between them, creating a demand for new LLM comparison methods. In this research, the Consistency-focused Similarity Comparison Framework (ConSCompF) for generative large language models is proposed. It compares texts generated by two LLMs and produces a similarity score, indicating the overall degree of similarity between their responses. The main advantage of this framework is that it can operate on a small number of unlabeled data, such as chatbot instruction prompts, and does not require LLM developers to disclose any information about their product. To evaluate the efficacy of ConSCompF, two experiments aimed at identifying similarities between multiple LLMs are conducted. Additionally, these experiments examine the correlation between the similarity scores generated by ConSCompF and the differences in the outputs produced by other benchmarking techniques, such as ROUGE-L. Finally, a series of few-shot LLM comparison experiments is conducted to evaluate the performance of ConSCompF in a few-shot LLM comparison scenario. The proposed framework can be used for calculating similarity matrices of multiple LLMs, which can be effectively visualized using principal component analysis (PCA). The ConSCompF output may provide useful insights into data that might have been used during LLM training and help detect possible investment fraud attempts.

Co-speech gestures convey a wide variety of meanings and play an important role in face-to-face human interactions. These gestures significantly influence the addressee's engagement, recall, comprehension, and attitudes toward the speaker. Similarly, they impact interactions between humans and embodied virtual agents. The process of selecting and animating meaningful gestures has thus become a key focus in the design of these agents. However, automating this gesture selection process poses a significant challenge. Prior gesture generation techniques have varied from fully automated, data-driven methods, which often struggle to produce contextually meaningful gestures, to more manual approaches that require crafting specific gesture expertise and are time-consuming and lack generalizability. In this paper, we leverage the semantic capabilities of Large Language Models to develop a gesture selection approach that suggests meaningful, appropriate co-speech gestures. We first describe how information on gestures is encoded into GPT-4. Then, we conduct a study to evaluate alternative prompting approaches for their ability to select meaningful, contextually relevant gestures and to align them appropriately with the co-speech utterance. Finally, we detail and demonstrate how this approach has been implemented within a virtual agent system, automating the selection and subsequent animation of the selected gestures for enhanced human-agent interactions.

Temporal Context A wareness: A Defense Framework Against Multi-turn Manipulation Attacks on Large Language Models Prashant Kulkarni ORCID: 0009-0004-2344-4840 Mountain View, CA Assaf Namer ORCID: 0009-0008-5579-0544 Mountain View, CA Abstract --Many Large Language Models (LLMs) today are vulnerable to multi-turn manipulation attacks,where adversaries gradually build context through seemingly benign conversational turns to elicit harmful or unauthorized responses. These attacks exploit the temporal nature of dialogue to evade single-turn detection methods, posing a significant risk to the safe deployment of LLMs. This paper introduces the T emporal Context A wareness (TCA)framework, a novel defense mechanism designed to address this challenge by continuously analyzing semantic drift, cross-turn intention consistency, and evolving conversational patterns.The TCA framework integrates dynamic context embedding analysis, cross-turn consistency verification, and progressive risk scoring to detect and mitigate manipulation attempts effectively. Preliminary evaluations on simulated adversarial scenarios demonstrate the framework's potential to identify subtle manipulation patterns often missed by traditional detection techniques, offering a much-needed layer of security for conversational AI systems.In addition to outlining the design of TCA, we analyze diverse attack vectors and their progression across multi-turn conversations, providing valuable insights into adversarial tactics and their impact on LLM vulnerabilities. Our findings underscore the pressing need for robust, context-aware defenses in conversational AI systems and highlight the TCA framework as a promising direction for securing LLMs while preserving their utility in legitimate applications Index T erms --LLM Security, Multi-turn attacks, prompt security, obfuscation, prompt injection, security, trustworthy AI, jailbreak I. I NTRODUCTION Large Language Models (LLMs) have become integral to modern digital infrastructure, powering applications from customer service to healthcare assistance [Chen et al., 2023] [3].

Quizzes play a crucial role in education by reinforcing students' understanding of key concepts and encouraging self-directed exploration. However, compiling high-quality quizzes can be challenging and require deep expertise and insight into specific subject matter. Although LLMs have greatly enhanced the efficiency of quiz generation, concerns remain regarding the quality of these AI-generated quizzes and their educational impact on students. To address these issues, we introduce ConQuer, a concept-based quiz generation framework that leverages external knowledge sources. We employ comprehensive evaluation dimensions to assess the quality of the generated quizzes, using LLMs as judges. Our experiment results demonstrate a 4.8% improvement in evaluation scores and a 77.52% win rate in pairwise comparisons against baseline quiz sets. Ablation studies further underscore the effectiveness of each component in our framework. Code available at https://github.com/sofyc/ConQuer.

The malicious use or malfunction of advanced general-purpose AI (GPAI) poses risks that, according to leading experts, could lead to the 'marginalisation or extinction of humanity.' To address these risks, there are an increasing number of proposals for international agreements on AI safety. In this paper, we review recent (2023-) proposals, identifying areas of consensus and disagreement, and drawing on related literature to assess their feasibility. We focus our discussion on risk thresholds, regulations, types of international agreement and five related processes: building scientific consensus, standardisation, auditing, verification and incentivisation. Based on this review, we propose a treaty establishing a compute threshold above which development requires rigorous oversight. This treaty would mandate complementary audits of models, information security and governance practices, overseen by an international network of AI Safety Institutes (AISIs) with authority to pause development if risks are unacceptable. Our approach combines immediately implementable measures with a flexible structure that can adapt to ongoing research.

Social platforms have expanded opportunities for deliberation with the comments being used to inform one's opinion. However, using such information to form opinions is challenged by unsubstantiated or false content. To enhance the quality of opinion formation and potentially confer resistance to misinformation, we developed Iffy-Or-Not (ION), a browser extension that seeks to invoke critical thinking when reading texts. With three features guided by argumentation theory, ION highlights fallacious content, suggests diverse queries to probe them with, and offers deeper questions to consider and chat with others about. From a user study (N=18), we found that ION encourages users to be more attentive to the content, suggests queries that align with or are preferable to their own, and poses thought-provoking questions that expands their perspectives. However, some participants expressed aversion to ION due to misalignments with their information goals and thinking predispositions. Potential backfiring effects with ION are discussed.

How is the limited capacity of working memory efficiently used to support human linguistic behaviors? In this paper, we investigate strategic resource allocation as an efficiency principle for memory encoding in sentence processing. The idea is that working memory resources are dynamically and strategically allocated to prioritize novel and unexpected information, enhancing their representations to make them less susceptible to memory decay and interference. Theoretically, from a resource-rational perspective, we argue that this efficiency principle naturally arises from two functional assumptions about working memory, namely, its limited capacity and its noisy representation. Empirically, through naturalistic corpus data, we find converging evidence for strategic resource allocation in the context of dependency locality from both the production and the comprehension side, where non-local dependencies with less predictable antecedents are associated with reduced locality effect. However, our results also reveal considerable cross-linguistic variability, highlighting the need for a closer examination of how strategic resource allocation, as a universal efficiency principle, interacts with language-specific phrase structures.