Large language models (LLMs) have achieved impressive performance in code generation. However, due to the long-tail distribution of LLMs' training data, low-frequency terms are typically underrepresented in the training process. Consequently, LLMs often misunderstand or overlook problem-specific, low-frequency keywords during code generation, compromising the accuracy of the generated code. To address this, we propose a novel technique named SEK(\textbf{S}elf-\textbf{E}xplained \textbf{K}eywords), which empowers an LLM for better code generation by extracting and explaining the key terms in the problem description with the LLM itself and ranking them based on frequency. Comprehensive experiments across three benchmarks, i.e., HumanEval(+), MBPP(+), and APPS, with five representative LLMs, show that SEK can significantly improve LLMs in code generation, yielding substantial and consistent gains. For instance, SEK improves the Pass@1 of DeepSeek-Coder-V2-Instruct from 85.4\% to 93.3\% on the Humaneval benchmark. Further analysis confirms that SEK enables the LLMs to shift their attention from low-frequency keywords to their corresponding high-frequency counterparts.

Formal proofs are challenging to write even for experienced experts. Recent progress in Neural Theorem Proving (NTP) shows promise in expediting this process. However, the formal corpora available on the Internet are limited compared to the general text, posing a significant data scarcity challenge for NTP. To address this issue, this work proposes Alchemy, a general framework for data synthesis that constructs formal theorems through symbolic mutation. Specifically, for each candidate theorem in Mathlib, we identify all invocable theorems that can be used to rewrite or apply to it. Subsequently, we mutate the candidate theorem by replacing the corresponding term in the statement with its equivalent form or antecedent. As a result, our method increases the number of theorems in Mathlib by an order of magnitude, from 110k to 6M. Furthermore, we perform continual pretraining and supervised finetuning on this augmented corpus for large language models. Experimental results demonstrate the effectiveness of our approach, achieving a 5% absolute performance improvement on Leandojo benchmark. Additionally, our synthetic data achieve a 2.5% absolute performance gain on the out-of-distribution miniF2F benchmark. To provide further insights, we conduct a comprehensive analysis of synthetic data composition and the training paradigm, offering valuable guidance for developing a strong theorem prover.

Increasing use of large language models (LLMs) demand performant guardrails to ensure the safety of inputs and outputs of LLMs. When these safeguards are trained on imbalanced data, they can learn the societal biases. We present a light-weight, post-processing method for mitigating counterfactual fairness in closed-source text safety classifiers. Our approach involves building an ensemble that not only outperforms the input classifiers and policy-aligns them, but also acts as a debiasing regularizer. We introduce two threshold-agnostic metrics to assess the counterfactual fairness of a model, and demonstrate how combining these metrics with Fair Data Reweighting (FDW) helps mitigate biases. We create an expanded Open AI dataset, and a new templated LLM-generated dataset based on user-prompts, both of which are counterfactually balanced across identity groups and cover four key areas of safety; we will work towards publicly releasing these datasets. Our results show that our approach improves counterfactual fairness with minimal impact on model performance.

Evaluating retrieval-augmented generation (RAG) systems remains challenging, particularly for open-ended questions that lack definitive answers and require coverage of multiple sub-topics. In this paper, we introduce a novel evaluation framework based on sub-question coverage, which measures how well a RAG system addresses different facets of a question. We propose decomposing questions into sub-questions and classifying them into three types -- core, background, and follow-up -- to reflect their roles and importance. Using this categorization, we introduce a fine-grained evaluation protocol that provides insights into the retrieval and generation characteristics of RAG systems, including three commercial generative answer engines: You.com, Perplexity AI, and Bing Chat. Interestingly, we find that while all answer engines cover core sub-questions more often than background or follow-up ones, they still miss around 50% of core sub-questions, revealing clear opportunities for improvement. Further, sub-question coverage metrics prove effective for ranking responses, achieving 82% accuracy compared to human preference annotations. Lastly, we also demonstrate that leveraging core sub-questions enhances both retrieval and answer generation in a RAG system, resulting in a 74% win rate over the baseline that lacks sub-questions.

The rise of large language models (LLMs) for visually rich document understanding (VRDU) has kindled a need for prompt-response, document-based datasets. As annotating new datasets from scratch is labor-intensive, the existing literature has generated prompt-response datasets from available resources using simple templates. For the case of key information extraction (KIE), one of the most common VRDU tasks, past work has typically employed the template "What is the value for the {key}?". However, given the variety of questions encountered in the wild, simple and uniform templates are insufficient for creating robust models in research and industrial contexts. In this work, we present K2Q, a diverse collection of five datasets converted from KIE to a prompt-response format using a plethora of bespoke templates. The questions in K2Q can span multiple entities and be extractive or boolean. We empirically compare the performance of seven baseline generative models on K2Q with zero-shot prompting. We further compare three of these models when training on K2Q versus training on simpler templates to motivate the need of our work. We find that creating diverse and intricate KIE questions enhances the performance and robustness of VRDU models. We hope this work encourages future studies on data quality for generative model training.

Detecting fake news in large datasets is challenging due to its diversity and complexity, with traditional approaches often focusing on textual features while underutilizing semantic and emotional elements. Current methods also rely heavily on large annotated datasets, limiting their effectiveness in more nuanced analysis. To address these challenges, this paper introduces Emotion-\textbf{A}ware \textbf{M}ultimodal Fusion \textbf{P}rompt \textbf{L}\textbf{E}arning (\textbf{AMPLE}) framework to address the above issue by combining text sentiment analysis with multimodal data and hybrid prompt templates. This framework extracts emotional elements from texts by leveraging sentiment analysis tools. It then employs Multi-Head Cross-Attention (MCA) mechanisms and similarity-aware fusion methods to integrate multimodal data. The proposed AMPLE framework demonstrates strong performance on two public datasets in both few-shot and data-rich settings, with results indicating the potential of emotional aspects in fake news detection. Furthermore, the study explores the impact of integrating large language models with this method for text sentiment extraction, revealing substantial room for further improvement. The code can be found at :\url{https://github.com/xxm1215/MMM2025_few-shot/

Large Language Models (LLMs) are pivotal AI agents in complex tasks but still face challenges in open decision-making problems within complex scenarios. To address this, we use the language logic game ``Who is Undercover?'' (WIU) as an experimental platform to propose the Multi-Perspective Team Tactic (MPTT) framework. MPTT aims to cultivate LLMs' human-like language expression logic, multi-dimensional thinking, and self-perception in complex scenarios. By alternating speaking and voting sessions, integrating techniques like self-perspective, identity-determination, self-reflection, self-summary and multi-round find-teammates, LLM agents make rational decisions through strategic concealment and communication, fostering human-like trust. Preliminary results show that MPTT, combined with WIU, leverages LLMs' cognitive capabilities to create a decision-making framework that can simulate real society. This framework aids minority groups in communication and expression, promoting fairness and diversity in decision-making. Additionally, our Human-in-the-loop experiments demonstrate that LLMs can learn and align with human behaviors through interactive, indicating their potential for active participation in societal decision-making.

As large language models (LLMs) continue to evolve, their potential use in automating cyberattacks becomes increasingly likely. With capabilities such as reconnaissance, exploitation, and command execution, LLMs could soon become integral to autonomous cyber agents, capable of launching highly sophisticated attacks. In this paper, we introduce novel defense strategies that exploit the inherent vulnerabilities of attacking LLMs. By targeting weaknesses such as biases, trust in input, memory limitations, and their tunnel-vision approach to problem-solving, we develop techniques to mislead, delay, or neutralize these autonomous agents. We evaluate our defenses under black-box conditions, starting with single prompt-response scenarios and progressing to real-world tests using custom-built CTF machines. Our results show defense success rates of up to 90\%, demonstrating the effectiveness of turning LLM vulnerabilities into defensive strategies against LLM-driven cyber threats.

The integration of Large Language Models (LLMs) into autonomous driving systems demonstrates strong common sense and reasoning abilities, effectively addressing the pitfalls of purely data-driven methods. Current LLM-based agents require lengthy inference times and face challenges in interacting with real-time autonomous driving environments. A key open question is whether we can effectively leverage the knowledge from LLMs to train an efficient and robust Reinforcement Learning (RL) agent. This paper introduces RAPID, a novel \underline{\textbf{R}}obust \underline{\textbf{A}}daptive \underline{\textbf{P}}olicy \underline{\textbf{I}}nfusion and \underline{\textbf{D}}istillation framework, which trains specialized mix-of-policy RL agents using data synthesized by an LLM-based driving agent and online adaptation. RAPID features three key designs: 1) utilization of offline data collected from an LLM agent to distil expert knowledge into RL policies for faster real-time inference; 2) introduction of robust distillation in RL to inherit both performance and robustness from LLM-based teacher; and 3) employment of a mix-of-policy approach for joint decision decoding with a policy adapter. Through fine-tuning via online environment interaction, RAPID reduces the forgetting of LLM knowledge while maintaining adaptability to different tasks. Extensive experiments demonstrate RAPID's capability to effectively integrate LLM knowledge into scaled-down RL policies in an efficient, adaptable, and robust way. Code and checkpoints will be made publicly available upon acceptance.

In the realm of Large Language Model (LLM) functionalities, providing reliable information is paramount, yet reports suggest that LLM outputs lack consistency. This inconsistency, often at-tributed to randomness in token sampling, under-mines user trust as it leads to varying responses even for identical queries. In this paper, we present a new approach for evaluating semantic consistencies of LLM including comparison of alternative tech-niques. Our approach evaluates whether LLM re-sponses are semantically congruent for a given question, recognizing that as syntactically different sentences may convey the same meaning. Here-tofore, To enhance LLM consistency, two main approaches have been explored: Leverage external knowledge as context like the RAG pattern or use Zero-shot-CoT to improve performance of LLM itself. We apply our evaluation approach to these techniques, and demonstrate to compare the im-pact of these methods on LLM response con-sistency across different domains of question an-swering tasks. Using the TruthfulQA dataset to assess LLM responses, the study induces N re-sponses per question from the LLM and clusters semantically equivalent sentences to measure semantic consistency across 37 categories. Through this, it quantitatively analyzes the effectiveness of the aforementioned methods in improving LLM performance before and after their adoption.