As the capabilities of chatbots and their underlying LLMs continue to dramatically improve, evaluating their performance has increasingly become a major blocker to their further development. A major challenge is the available benchmarking datasets, which are largely static, outdated, and lacking in multilingual coverage, limiting their ability to capture subtle linguistic and cultural variations. This paper introduces MEDAL, an automated multi-agent framework for generating, evaluating, and curating more representative and diverse open-domain dialogue evaluation benchmarks. Our approach leverages several state-of-the-art LLMs to generate user-chatbot multilingual dialogues, conditioned on varied seed contexts. A strong LLM (GPT-4.1) is then used for a multidimensional analysis of the performance of the chatbots, uncovering noticeable cross-lingual performance differences. Guided by this large-scale evaluation, we curate a new meta-evaluation multilingual benchmark and human-annotate samples with nuanced quality judgments. This benchmark is then used to assess the ability of several reasoning and non-reasoning LLMs to act as evaluators of open-domain dialogues. We find that current LLMs struggle to detect nuanced issues, particularly those involving empathy and reasoning.
Self-disclosure, the sharing of one's thoughts and feelings, is affected by the perceived relationship between individuals. While chatbots are increasingly used for self-disclosure, the impact of a chatbot's framing on users' self-disclosure remains under-explored. We investigated how a chatbot's description of its relationship with users, particularly in terms of ephemerality, affects self-disclosure. Specifically, we compared a Familiar chatbot, presenting itself as a companion remembering past interactions, with a Stranger chatbot, presenting itself as a new, unacquainted entity in each conversation. In a mixed factorial design, participants engaged with either the Familiar or Stranger chatbot in two sessions across two days, with one conversation focusing on Emotional- and another Factual-disclosure. When Emotional-disclosure was sought in the first chatting session, Stranger-condition participants felt more comfortable self-disclosing. However, when Factual-disclosure was sought first, these differences were replaced by more enjoyment among Familiar-condition participants. Qualitative findings showed Stranger afforded anonymity and reduced judgement, whereas Familiar sometimes felt intrusive unless rapport was built via low-risk Factual-disclosure.
State-of-the-art large language models require specialized hardware and substantial energy to operate. As a consequence, cloud-based services that provide access to large language models have become very popular. In these services, the price users pay for an output provided by a model depends on the number of tokens the model uses to generate it -- they pay a fixed price per token. In this work, we show that this pricing mechanism creates a financial incentive for providers to strategize and misreport the (number of) tokens a model used to generate an output, and users cannot prove, or even know, whether a provider is overcharging them. However, we also show that, if an unfaithful provider is obliged to be transparent about the generative process used by the model, misreporting optimally without raising suspicion is hard. Nevertheless, as a proof-of-concept, we introduce an efficient heuristic algorithm that allows providers to significantly overcharge users without raising suspicion, highlighting the vulnerability of users under the current pay-per-token pricing mechanism. Further, to completely eliminate the financial incentive to strategize, we introduce a simple incentive-compatible token pricing mechanism. Under this mechanism, the price users pay for an output provided by a model depends on the number of characters of the output -- they pay a fixed price per character. Along the way, to illustrate and complement our theoretical results, we conduct experiments with several large language models from the $\texttt{Llama}$, $\texttt{Gemma}$ and $\texttt{Ministral}$ families, and input prompts from the LMSYS Chatbot Arena platform.
The conversational capabilities of Large Language Models (LLMs) suggest that they may be able to perform as automated talk therapists. It is crucial to know if these systems would be effective and adhere to known standards. We present a counsellor chatbot that focuses on motivating tobacco smokers to quit smoking. It uses a state-of-the-art LLM and a widely applied therapeutic approach called Motivational Interviewing (MI), and was evolved in collaboration with clinician-scientists with expertise in MI. We also describe and validate an automated assessment of both the chatbot's adherence to MI and client responses. The chatbot was tested on 106 participants, and their confidence that they could succeed in quitting smoking was measured before the conversation and one week later. Participants' confidence increased by an average of 1.7 on a 0-10 scale. The automated assessment of the chatbot showed adherence to MI standards in 98% of utterances, higher than human counsellors. The chatbot scored well on a participant-reported metric of perceived empathy but lower than typical human counsellors. Furthermore, participants' language indicated a good level of motivation to change, a key goal in MI. These results suggest that the automation of talk therapy with a modern LLM has promise.
The increasing sophistication and integration of Generative AI (GenAI) models into diverse applications introduce new security challenges that traditional methods struggle to address. This research explores the critical need for proactive security measures to mitigate the risks associated with malicious exploitation of GenAI systems. We present a framework encompassing key approaches, tools, and strategies designed to outmaneuver even advanced adversarial attacks, emphasizing the importance of securing GenAI innovation against potential liabilities. We also empirically prove the effectiveness of the said framework by testing it against the SPML Chatbot Prompt Injection Dataset. This work highlights the shift from reactive to proactive security practices essential for the safe and responsible deployment of GenAI technologies
LLM-based multi-agent systems (MAS) extend the capabilities of single LLMs by enabling cooperation among multiple specialized agents. However, most existing MAS frameworks rely on a single LLM to drive all agents, constraining the system's intelligence to the limit of that model. This paper explores the paradigm of heterogeneous LLM-driven MAS (X-MAS), where agents are powered by diverse LLMs, elevating the system's potential to the collective intelligence of diverse LLMs. We introduce X-MAS-Bench, a comprehensive testbed designed to evaluate the performance of various LLMs across different domains and MAS-related functions. As an extensive empirical study, we assess 27 LLMs across 5 domains (encompassing 21 test sets) and 5 functions, conducting over 1.7 million evaluations to identify optimal model selections for each domain-function combination. Building on these findings, we demonstrate that transitioning from homogeneous to heterogeneous LLM-driven MAS can significantly enhance system performance without requiring structural redesign. Specifically, in a chatbot-only MAS scenario, the heterogeneous configuration yields up to 8.4\% performance improvement on the MATH dataset. In a mixed chatbot-reasoner scenario, the heterogeneous MAS could achieve a remarkable 47\% performance boost on the AIME dataset. Our results underscore the transformative potential of heterogeneous LLMs in MAS, highlighting a promising avenue for advancing scalable, collaborative AI systems.
Large Language Models (LLMs) have shown strong capability in diverse software engineering tasks, e.g. code completion, bug fixing, and document generation. However, feature-driven development (FDD), a highly prevalent real-world task that involves developing new functionalities for large, existing codebases, remains underexplored. We therefore introduce SWE-Dev, the first large-scale dataset (with 14,000 training and 500 test samples) designed to evaluate and train autonomous coding systems on real-world feature development tasks. To ensure verifiable and diverse training, SWE-Dev uniquely provides all instances with a runnable environment and its developer-authored executable unit tests. This collection not only provides high-quality data for Supervised Fine-Tuning (SFT), but also enables Reinforcement Learning (RL) by delivering accurate reward signals from executable unit tests. Our extensive evaluations on SWE-Dev, covering 17 chatbot LLMs, 10 reasoning models, and 10 Multi-Agent Systems (MAS), reveal that FDD is a profoundly challenging frontier for current AI (e.g., Claude-3.7-Sonnet achieves only 22.45\% Pass@3 on the hard test split). Crucially, we demonstrate that SWE-Dev serves as an effective platform for model improvement: fine-tuning on training set enabled a 7B model comparable to GPT-4o on \textit{hard} split, underscoring the value of its high-quality training data. Code is available here \href{https://github.com/DorothyDUUU/SWE-Dev}{https://github.com/DorothyDUUU/SWE-Dev}.
As large language models (LLMs) are increasingly deployed in high-stakes settings, their ability to refuse ethically sensitive prompts-such as those involving hate speech or illegal activities-has become central to content moderation and responsible AI practices. While refusal responses can be viewed as evidence of ethical alignment and safety-conscious behavior, recent research suggests that users may perceive them negatively. At the same time, automated assessments of model outputs are playing a growing role in both evaluation and training. In particular, LLM-as-a-Judge frameworks-in which one model is used to evaluate the output of another-are now widely adopted to guide benchmarking and fine-tuning. This paper examines whether such model-based evaluators assess refusal responses differently than human users. Drawing on data from Chatbot Arena and judgments from two AI judges (GPT-4o and Llama 3 70B), we compare how different types of refusals are rated. We distinguish ethical refusals, which explicitly cite safety or normative concerns (e.g., "I can't help with that because it may be harmful"), and technical refusals, which reflect system limitations (e.g., "I can't answer because I lack real-time data"). We find that LLM-as-a-Judge systems evaluate ethical refusals significantly more favorably than human users, a divergence not observed for technical refusals. We refer to this divergence as a moderation bias-a systematic tendency for model-based evaluators to reward refusal behaviors more than human users do. This raises broader questions about transparency, value alignment, and the normative assumptions embedded in automated evaluation systems.
As Large Language Models (LLMs) rapidly advance, we introduce Hunyuan-TurboS, a novel large hybrid Transformer-Mamba Mixture of Experts (MoE) model. It synergistically combines Mamba's long-sequence processing efficiency with Transformer's superior contextual understanding. Hunyuan-TurboS features an adaptive long-short chain-of-thought (CoT) mechanism, dynamically switching between rapid responses for simple queries and deep "thinking" modes for complex problems, optimizing computational resources. Architecturally, this 56B activated (560B total) parameter model employs 128 layers (Mamba2, Attention, FFN) with an innovative AMF/MF block pattern. Faster Mamba2 ensures linear complexity, Grouped-Query Attention minimizes KV cache, and FFNs use an MoE structure. Pre-trained on 16T high-quality tokens, it supports a 256K context length and is the first industry-deployed large-scale Mamba model. Our comprehensive post-training strategy enhances capabilities via Supervised Fine-Tuning (3M instructions), a novel Adaptive Long-short CoT Fusion method, Multi-round Deliberation Learning for iterative improvement, and a two-stage Large-scale Reinforcement Learning process targeting STEM and general instruction-following. Evaluations show strong performance: overall top 7 rank on LMSYS Chatbot Arena with a score of 1356, outperforming leading models like Gemini-2.0-Flash-001 (1352) and o4-mini-2025-04-16 (1345). TurboS also achieves an average of 77.9% across 23 automated benchmarks. Hunyuan-TurboS balances high performance and efficiency, offering substantial capabilities at lower inference costs than many reasoning models, establishing a new paradigm for efficient large-scale pre-trained models.
As artificial intelligence (AI) regulations evolve and the regulatory landscape develops and becomes more complex, ensuring compliance with ethical guidelines and legal frameworks remains a challenge for AI developers. This paper introduces an AI-driven self-assessment chatbot designed to assist users in navigating the European Union AI Act and related standards. Leveraging a Retrieval-Augmented Generation (RAG) framework, the chatbot enables real-time, context-aware compliance verification by retrieving relevant regulatory texts and providing tailored guidance. By integrating both public and proprietary standards, it streamlines regulatory adherence, reduces complexity, and fosters responsible AI development. The paper explores the chatbot's architecture, comparing naive and graph-based RAG models, and discusses its potential impact on AI governance.
Mental-health stigma remains a pervasive social problem that hampers treatment-seeking and recovery. Existing resources for training neural models to finely classify such stigma are limited, relying primarily on social-media or synthetic data without theoretical underpinnings. To remedy this gap, we present an expert-annotated, theory-informed corpus of human-chatbot interviews, comprising 4,141 snippets from 684 participants with documented socio-cultural backgrounds. Our experiments benchmark state-of-the-art neural models and empirically unpack the challenges of stigma detection. This dataset can facilitate research on computationally detecting, neutralizing, and counteracting mental-health stigma. Our corpus is openly available at https://github.com/HanMeng2004/Mental-Health-Stigma-Interview-Corpus.
The recent explosion of large language models (LLMs), each with its own general or specialized strengths, makes scalable, reliable benchmarking more urgent than ever. Standard practices nowadays face fundamental trade-offs: closed-ended question-based benchmarks (eg MMLU) struggle with saturation as newer models emerge, while crowd-sourced leaderboards (eg Chatbot Arena) rely on costly and slow human judges. Recently, automated methods (eg LLM-as-a-judge) shed light on the scalability, but risk bias by relying on one or a few "authority" models. To tackle these issues, we propose Decentralized Arena (dearena), a fully automated framework leveraging collective intelligence from all LLMs to evaluate each other. It mitigates single-model judge bias by democratic, pairwise evaluation, and remains efficient at scale through two key components: (1) a coarse-to-fine ranking algorithm for fast incremental insertion of new models with sub-quadratic complexity, and (2) an automatic question selection strategy for the construction of new evaluation dimensions. Across extensive experiments across 66 LLMs, dearena attains up to 97% correlation with human judgements, while significantly reducing the cost. Our code and data will be publicly released on https://github.com/maitrix-org/de-arena.
Open-domain dialogue systems aim to generate natural and engaging conversations, providing significant practical value in real applications such as social robotics and personal assistants. The advent of large language models (LLMs) has greatly advanced this field by improving context understanding and conversational fluency. However, existing LLM-based dialogue systems often fall short in proactively understanding the user's chatting preferences and guiding conversations toward user-centered topics. This lack of user-oriented proactivity can lead users to feel unappreciated, reducing their satisfaction and willingness to continue the conversation in human-computer interactions. To address this issue, we propose a User-oriented Proactive Chatbot (UPC) to enhance the user-oriented proactivity. Specifically, we first construct a critic to evaluate this proactivity inspired by the LLM-as-a-judge strategy. Given the scarcity of high-quality training data, we then employ the critic to guide dialogues between the chatbot and user agents, generating a corpus with enhanced user-oriented proactivity. To ensure the diversity of the user backgrounds, we introduce the ISCO-800, a diverse user background dataset for constructing user agents. Moreover, considering the communication difficulty varies among users, we propose an iterative curriculum learning method that trains the chatbot from easy-to-communicate users to more challenging ones, thereby gradually enhancing its performance. Experiments demonstrate that our proposed training method is applicable to different LLMs, improving user-oriented proactivity and attractiveness in open-domain dialogues.
Recent advances in generative AI, such as ChatGPT, have transformed access to information in education, knowledge-seeking, and everyday decision-making. However, in many developing regions, access remains a challenge due to the persistent digital divide. To help bridge this gap, we developed WaLLM - a custom AI chatbot over WhatsApp, a widely used communication platform in developing regions. Beyond answering queries, WaLLM offers several features to enhance user engagement: a daily top question, suggested follow-up questions, trending and recent queries, and a leaderboard-based reward system. Our service has been operational for over 6 months, amassing over 14.7K queries from approximately 100 users. In this paper, we present WaLLM's design and a systematic analysis of logs to understand user interactions. Our results show that 55% of user queries seek factual information. "Health and well-being" was the most popular topic (28%), including queries about nutrition and disease, suggesting users view WaLLM as a reliable source. Two-thirds of users' activity occurred within 24 hours of the daily top question. Users who accessed the "Leaderboard" interacted with WaLLM 3x as those who did not. We conclude by discussing implications for culture-based customization, user interface design, and appropriate calibration of users' trust in AI systems for developing regions.
This paper presents a chatbot-based system to engage young Spanish people in the awareness of certain mental disorders through a self-disclosure technique. The study was carried out in a population of teenagers aged between 12 and 18 years. The dialogue engine mixes closed and open conversations, so certain controlled messages are sent to focus the chat on a specific disorder, which will change over time. Once a set of trial questions is answered, the system can initiate the conversation on the disorder under the focus according to the user's sensibility to that disorder, in an attempt to establish a more empathetic communication. Then, an open conversation based on the GPT-3 language model is initiated, allowing the user to express themselves with more freedom. The results show that these systems are of interest to young people and could help them become aware of certain mental disorders.
The emergence of Generative AI (Gen AI) and Large Language Models (LLMs) has enabled more advanced chatbots capable of human-like interactions. However, these conversational agents introduce a broader set of operational risks that extend beyond traditional cybersecurity considerations. In this work, we propose a novel, instrumented risk-assessment metric that simultaneously evaluates potential threats to three key stakeholders: the service-providing organization, end users, and third parties. Our approach incorporates the technical complexity required to induce erroneous behaviors in the chatbot--ranging from non-induced failures to advanced prompt-injection attacks--as well as contextual factors such as the target industry, user age range, and vulnerability severity. To validate our metric, we leverage Garak, an open-source framework for LLM vulnerability testing. We further enhance Garak to capture a variety of threat vectors (e.g., misinformation, code hallucinations, social engineering, and malicious code generation). Our methodology is demonstrated in a scenario involving chatbots that employ retrieval-augmented generation (RAG), showing how the aggregated risk scores guide both short-term mitigation and longer-term improvements in model design and deployment. The results underscore the importance of multi-dimensional risk assessments in operationalizing secure, reliable AI-driven conversational systems.
Large Language Models (LLMs) are a transformational technology, fundamentally changing how people obtain information and interact with the world. As people become increasingly reliant on them for an enormous variety of tasks, a body of academic research has developed to examine these models for inherent biases, especially political biases, often finding them small. We challenge this prevailing wisdom. First, by comparing 31 LLMs to legislators, judges, and a nationally representative sample of U.S. voters, we show that LLMs' apparently small overall partisan preference is the net result of offsetting extreme views on specific topics, much like moderate voters. Second, in a randomized experiment, we show that LLMs can promulgate their preferences into political persuasiveness even in information-seeking contexts: voters randomized to discuss political issues with an LLM chatbot are as much as 5 percentage points more likely to express the same preferences as that chatbot. Contrary to expectations, these persuasive effects are not moderated by familiarity with LLMs, news consumption, or interest in politics. LLMs, especially those controlled by private companies or governments, may become a powerful and targeted vector for political influence.
As large language models (LLMs) improve in their capacity to serve as personal AI assistants, their ability to output uniquely tailored, personalized responses that align with the soft preferences of their users is essential for enhancing user satisfaction and retention. However, untrained lay users have poor prompt specification abilities and often struggle with conveying their latent preferences to AI assistants. To address this, we leverage activation steering to guide LLMs to align with interpretable preference dimensions during inference. In contrast to memory-based personalization methods that require longer user history, steering is extremely lightweight and can be easily controlled by the user via an linear strength factor. We embed steering into three different interactive chatbot interfaces and conduct a within-subjects user study (n=14) to investigate how end users prefer to personalize their conversations. The results demonstrate the effectiveness of preference-based steering for aligning real-world conversations with hidden user preferences, and highlight further insights on how diverse values around control, usability, and transparency lead users to prefer different interfaces.
Battles, or side-by-side comparisons in so called arenas that elicit human preferences, have emerged as a popular approach to assessing the output quality of LLMs. Recently, this idea has been extended to retrieval-augmented generation (RAG) systems. While undoubtedly representing an advance in evaluation, battles have at least two drawbacks, particularly in the context of complex information-seeking queries: they are neither explanatory nor diagnostic. Recently, the nugget evaluation methodology has emerged as a promising approach to evaluate the quality of RAG answers. Nuggets decompose long-form LLM-generated answers into atomic facts, highlighting important pieces of information necessary in a "good" response. In this work, we apply our AutoNuggetizer framework to analyze data from roughly 7K Search Arena battles provided by LMArena in a fully automatic manner. Our results show a significant correlation between nugget scores and human preferences, showcasing promise in our approach to explainable and diagnostic system evaluations.
Measuring progress is fundamental to the advancement of any scientific field. As benchmarks play an increasingly central role, they also grow more susceptible to distortion. Chatbot Arena has emerged as the go-to leaderboard for ranking the most capable AI systems. Yet, in this work we identify systematic issues that have resulted in a distorted playing field. We find that undisclosed private testing practices benefit a handful of providers who are able to test multiple variants before public release and retract scores if desired. We establish that the ability of these providers to choose the best score leads to biased Arena scores due to selective disclosure of performance results. At an extreme, we identify 27 private LLM variants tested by Meta in the lead-up to the Llama-4 release. We also establish that proprietary closed models are sampled at higher rates (number of battles) and have fewer models removed from the arena than open-weight and open-source alternatives. Both these policies lead to large data access asymmetries over time. Providers like Google and OpenAI have received an estimated 19.2% and 20.4% of all data on the arena, respectively. In contrast, a combined 83 open-weight models have only received an estimated 29.7% of the total data. We show that access to Chatbot Arena data yields substantial benefits; even limited additional data can result in relative performance gains of up to 112% on the arena distribution, based on our conservative estimates. Together, these dynamics result in overfitting to Arena-specific dynamics rather than general model quality. The Arena builds on the substantial efforts of both the organizers and an open community that maintains this valuable evaluation platform. We offer actionable recommendations to reform the Chatbot Arena's evaluation framework and promote fairer, more transparent benchmarking for the field