Understanding uncertainty in large language models remains a fundamental challenge, particularly in creative tasks where multiple valid outputs exist. We present a geometric framework using credal sets - convex hulls of probability distributions - to quantify and decompose uncertainty in neural text generation, calibrated against human creative variation. Analyzing 500 creative writing prompts from the WritingPrompts dataset with 10 unique human continuations each, we evaluate four language models across five decoding strategies, generating 100,000 stories. Our credal set analysis reveals substantial gaps in capturing human creative variation, with the best model-human calibration reaching only 0.434 (Gemma-2B with temperature 0.7). We decompose total uncertainty into epistemic and aleatoric components, finding that the choice of decoding strategy contributes 39.4% to 72.0% of total epistemic uncertainty. Model scale shows weak correlation with calibration quality and no significant difference exists between base and instruction-tuned models in calibration quality. Our geometric framework provides actionable insights for improving generation systems for human-AI creative alignment. We release our complete experimental framework.

Retrieval-augmented generation agents development is hindered by the lack of process-level supervision to effectively guide agentic capabilities like task decomposition, retriever invocation, and stepwise decision-making. While reinforcement learning offers a potential solution, it suffers from sparse rewards and the limited reasoning capabilities of large language models (LLMs). Meanwhile, existing data synthesis methods only produce chain-of-thought rationales and fail to model environmental interactions. In this paper, we propose EviPath, an evidence-anchored reasoning path synthesis paradigm for RAG agent development. EviPath comprises: (i) Abductive Subtask Planning, which decomposes the problem into sub-questions and iteratively plans an optimal solution path based on the dependencies between them; (ii) Faithful Sub-question Answering, which uses supporting evidence to construct a proxy environment to generate reasoning thoughts and answers for each sub-question; and (iii) Conversational Fine-Tuning, which formats the complete agent-environment interaction trajectory into a dialogue format suitable for Supervised Fine-Tuning. EviPath allows LLMs to learn complex reasoning and tool-use capabilities directly from synthesized data. Extensive experiments on widely-used question-answering benchmarks show that an 8B parameter model trained with EviPath-synthesized data significantly and consistently outperforms state-of-the-art baselines with a double-digit absolute EM gain of 14.7% in open-domain question answering. Retrieval-augmented generation (RAG) agents, powered by large language models (LLMs) (Guo et al., 2025), can autonomously gather external knowledge and answer complex, multi-hop questions. Compared to vanilla RAG systems (Lewis et al., 2020), RAG agents minimize the need for human intervention, and adapt readily to downstream applications like math problem solving (Zhu et al., 2025), code generation (Zhang et al., 2023), and financial analysis (Wang et al., 2025c). Despite their promise, RAG agents are hard to develop since ground truth reasoning trajectories are unavailable. Mainstream multi-hop question answering datasets Y ang et al. (2018); Ho et al. (2020); Trivedi et al. (2022) provide final answers and supporting facts, while lacking step-wise supervision that is crucial to equip LLMs with agentic behaviors like question decomposition, search query reformulation, and plan refinement.

The rising cost of acquiring supervised data has driven significant interest in self-improvement for large language models (LLMs). Straightforward unsupervised signals like majority voting have proven effective in generating pseudo-labels for verifiable tasks, while their applicability to unverifiable tasks (e.g., translation) is limited by the open-ended character of responses. As a result, self-evaluation mechanisms (e.g., self-judging and entropy minimization) are predominantly used to derive pseudo-labels. However, self-evaluation relying on LLMs typically incurs high computational overhead and introduces overconfidence issues due to intrinsic biases. To address these challenges, we propose a novel self-evaluation-free approach for unverifiable tasks, designed for lightweight yet effective self-improvement. Inspired by majority voting commonly employed in verifiable tasks, we propose semantic voting as a novel mechanism that relaxes the principle of hard matching (i.e., exact matching) toward soft matching (i.e., semantic similarity). Soft matching is achieved by leveraging a lightweight sentence embedding model to quantify semantic similarity, thereby mitigating excessive computational burden and intrinsic bias-associated limitations of self-evaluation. Comprehensive experiments demonstrate that our method achieves substantial gains in computational efficiency and overall better performance than self-evaluation methods across diverse model architectures and tasks.

We introduce DafnyCOMP, a benchmark for evaluating large language models (LLMs) on compositional specification generation in Dafny. Unlike prior benchmarks that focus on single-function tasks, DafnyCOMP targets programs composed of multiple interacting functions with data dependencies, requiring reasoning across component boundaries. The benchmark consists of 300 automatically synthesized multi-function programs. We evaluate several state-of-the-art LLM families and find that, while they perform well on single-function verification, their performance drops sharply on compositional tasks. Analysis reveals systematic failures in cross-functional reasoning, including fragile specifications, misalignment between implementations and proofs, and unstable reasoning. DafnyCOMP thus provides a diagnostic tool for measuring progress toward reliable, verifiable, and compositional code generation with LLMs.

Large language models (LLMs) often display sycophancy, a tendency toward excessive agreeability. This behavior poses significant challenges for multi-agent debating systems (MADS) that rely on productive disagreement to refine arguments and foster innovative thinking. LLMs' inherent sycophancy can collapse debates into premature consensus, potentially undermining the benefits of multi-agent debate. While prior studies focus on user--LLM sycophancy, the impact of inter-agent sycophancy in debate remains poorly understood. To address this gap, we introduce the first operational framework that (1) proposes a formal definition of sycophancy specific to MADS settings, (2) develops new metrics to evaluate the agent sycophancy level and its impact on information exchange in MADS, and (3) systematically investigates how varying levels of sycophancy across agent roles (debaters and judges) affects outcomes in both decentralized and centralized debate frameworks. Our findings reveal that sycophancy is a core failure mode that amplifies disagreement collapse before reaching a correct conclusion in multi-agent debates, yields lower accuracy than single-agent baselines, and arises from distinct debater-driven and judge-driven failure modes. Building on these findings, we propose actionable design principles for MADS, effectively balancing productive disagreement with cooperation in agent interactions.

Abstract--Large Language Models (LLMs) are increasingly susceptible to jailbreak attacks, which are adversarial prompts that bypass alignment constraints and induce unauthorized or harmful behaviors. These vulnerabilities undermine the safety, reliability, and trustworthiness of LLM outputs, posing critical risks in domains such as healthcare, finance, and legal compliance. In this paper, we propose GuardNet, a hierarchical filtering framework that detects and filters jailbreak prompts prior to inference. GuardNet constructs structured graphs that combine sequential links, syntactic dependencies, and attention-derived token relations to capture both linguistic structure and contextual patterns indicative of jailbreak behavior. It then applies graph neural networks at two levels: (i) a prompt-level filter that detects global adversarial prompts, and (ii) a token-level filter that pinpoints fine-grained adversarial spans. Extensive experiments across three datasets and multiple attack settings show that GuardNet substantially outperforms prior defenses. Despite its structural complexity, GuardNet maintains acceptable latency and generalizes well in cross-domain evaluations, making it a practical and robust defense against jailbreak threats in real-world LLM deployments. I. Introduction Large Language Models (LLMs) have become central to a wide range of applications, powering systems in domains such as education [1], healthcare [2], finance [3], law [4], and customer support [5]. Their ability to understand and generate human-like text has enabled automation of complex tasks such as legal reasoning, clinical triage, financial analysis, and policy drafting. However, this general-purpose capability also makes LLMs vulnerable to misuse. In particular, LLMs are highly susceptible to prompt-based adversarial attacks, especially jailbreak prompts [6], [7], which are carefully engineered inputs designed to bypass alignment constraints and elicit unauthorized or harmful responses.

Standard training metrics like loss fail to explain the emergence of complex capabilities in large language models. We take a spectral approach to investigate the geometry of learned representations across pretraining and post-training, measuring effective rank (RankMe) and eigenspectrum decay ($α$-ReQ). With OLMo (1B-7B) and Pythia (160M-12B) models, we uncover a consistent non-monotonic sequence of three geometric phases during autoregressive pretraining. The initial "warmup" phase exhibits rapid representational collapse. This is followed by an "entropy-seeking" phase, where the manifold's dimensionality expands substantially, coinciding with peak n-gram memorization. Subsequently, a "compression-seeking" phase imposes anisotropic consolidation, selectively preserving variance along dominant eigendirections while contracting others, a transition marked with significant improvement in downstream task performance. We show these phases can emerge from a fundamental interplay of cross-entropy optimization under skewed token frequencies and representational bottlenecks ($d \ll |V|$). Post-training further transforms geometry: SFT and DPO drive "entropy-seeking" dynamics to integrate specific instructional or preferential data, improving in-distribution performance while degrading out-of-distribution robustness. Conversely, RLVR induces "compression-seeking", enhancing reward alignment but reducing generation diversity.

Mafia is a social deduction game where informed mafia compete against uninformed townsfolk. Its asymmetry of information and reliance on theory-of-mind reasoning mirror real-world multi-agent scenarios, making it a useful testbed for evaluating the social intelligence of large language models (LLMs). To support a systematic study, we introduce Mini-Mafia: a simplified four-player variant with one mafioso, one detective, and two villagers. We set the mafioso to kill a villager and the detective to investigate the mafioso during the night, reducing the game to a single day phase of discussion and voting. This setup isolates three interactive capabilities through role-specific win conditions: the mafioso must deceive, the villagers must detect deception, and the detective must effectively disclose information. To measure these skills, we have LLMs play against each other, creating the Mini-Mafia Benchmark: a two-stage framework that first estimates win rates within fixed opponent configurations, then aggregates performance across them using standardized scoring. Built entirely from model interactions without external data, the benchmark evolves as new models are introduced, with each one serving both as a new opponent and as a subject of evaluation. Our experiments reveal counterintuitive results, including cases where smaller models outperform larger ones. Beyond benchmarking, Mini-Mafia enables quantitative study of emergent multi-agent dynamics such as name bias and last-speaker advantage. It also contributes to AI safety by generating training data for deception detectors and by tracking models' deception capabilities against human baselines.

Agentic AI represents a paradigm shift in enhancing the capabilities of generative AI models. While these systems demonstrate immense potential and power, current evaluation techniques primarily focus on assessing their efficacy in identifying appropriate agents, tools, and parameters. However, a critical gap exists in evaluating the alignment between an Agentic AI system's tasks and its overarching goals. This paper introduces the Creative Adversarial Testing (CAT) framework, a novel approach designed to capture and analyze the complex relationship between Agentic AI tasks and the system's intended objectives. We validate the CAT framework through extensive simulation using synthetic interaction data modeled after Alexa+ audio services, a sophisticated Agentic AI system that shapes the user experience for millions of users globally. This synthetic data approach enables comprehensive testing of edge cases and failure modes while protecting user privacy. Our results demonstrate that the CAT framework provides unprecedented insights into goal-task alignment, enabling more effective optimization and development of Agentic AI systems.

We introduce ADAM (A Diverse Archive of Mankind), a framework for evaluating and improving multimodal large language models (MLLMs) in biographical reasoning. To the best of our knowledge, this is the first work to systematically examine LLM capabilities in biography, a critical yet underexplored dimension of factual knowledge. At its core, AdamDB is a multilingual and multimodal dataset covering over 4 million individuals across geography, time, and profession, while AdamBench provides cognitively structured evaluations based on Bloom's taxonomy, spanning six reasoning levels in both English and native languages. To address hallucinations, particularly for lesser-known individuals, we propose AdamRAG, a retrieval-augmented generation system tailored to biographical contexts. Experiments show that AdamRAG substantially improves open-source models and modestly benefits closed-source ones, with the largest gains on lower-order reasoning. Popularity strongly mediates accuracy, and multimodal input via face images offers smaller, less consistent improvements than retrieval. ADAM establishes the first benchmark and framework for cognitively, culturally, and multimodally grounded biographical evaluation, advancing the development of multilingual, accurate, and hallucination-resistant MLLMs.