The widespread deployment of Large Language Models (LLMs) has intensified concerns about subtle social biases embedded in their outputs. Existing guardrails often fail when faced with indirect or contextually complex bias-inducing prompts. To address these limitations, we propose a unified framework for both systematic bias evaluation and targeted mitigation. Our approach introduces six novel Metamorphic Relations (MRs) that, based on metamorphic testing principles, transform direct bias-inducing inputs into semantically equivalent yet adversarially challenging variants. These transformations enable an automated method for exposing hidden model biases: when an LLM responds inconsistently or unfairly across MR-generated variants, the underlying bias becomes detectable. We further show that the same MRs can be used to generate diverse bias-inducing samples for fine-tuning, directly linking the testing process to mitigation. Using six state-of-the-art LLMs - spanning open-source and proprietary models - and a representative subset of 385 questions from the 8,978-item BiasAsker benchmark covering seven protected groups, our MRs reveal up to 14% more hidden biases compared to existing tools. Moreover, fine-tuning with both original and MR-mutated samples significantly enhances bias resiliency, increasing safe response rates from 54.7% to over 88.9% across models. These results highlight metamorphic relations as a practical mechanism for improving fairness in conversational AI.

Recent reasoning large language models (LLMs) excel in complex tasks but encounter significant computational and memory challenges due to long sequence lengths. KV cache compression has emerged as an effective approach to greatly enhance the efficiency of reasoning. However, existing methods often focus on prompt compression or token eviction with local attention score, overlooking the long-term importance of tokens. We propose G-KV, a KV cache eviction method that employs a global scoring mechanism, combining local and historical attention scores to more accurately assess token importance. Additionally, we introduce post-training techniques, including reinforcement learning and distillation, to optimize models for compressed KV cache settings. The code of this paper is available on: https://github.com/microsoft/G-KV. Large language models (LLMs) have garnered widespread attention and applications. Recently released reasoning models have demonstrated remarkable performance (Guo et al., 2025; Team et al., 2025; Y ang et al., 2025), even in addressing complex tasks such as mathematics and coding. These reasoning models achieve significant improvements across various problems through long chain-of-thought (CoT) (Wei et al., 2022), enabling iterative reflection and verification. However, the long CoT of reasoning models typically consists of thousands or even tens of thousands of tokens. This imposes a substantial increase in computational costs and KV cache memory consumption. Notably, the computation of attention becomes a critical bottleneck, as its complexity scales quadratically with the sequence length. To overcome the bottlenecks of memory and computational complexity, numerous optimization methods for KV cache or attention mechanisms have been proposed (Li et al., 2024a). Among these, some methods prune the KV cache of tokens, significantly reducing computational overhead and memory consumption.

Data collection often reflects human decisions. In healthcare, for instance, a referral for a diagnostic test is influenced by the patient's health, their preferences, available resources, and the practitioner's recommendations. Despite the extensive literature on the informativeness of missingness, its implications on the performance of Large Language Models (LLMs) have not been studied. Through a series of experiments on data from Columbia University Medical Center, a large urban academic medical center, and MIMIC-IV, we demonstrate that patterns of missingness significantly impact zero-shot predictive performance. Notably, the explicit inclusion of missingness indicators at prompting benefits some while hurting other LLMs' zero-shot predictive performance and calibration, suggesting an inconsistent impact. The proposed aggregated analysis and theoretical insights suggest that larger models benefit from these interventions, while smaller models can be negatively impacted. The LLM paradigm risks obscuring the impact of missingness, often neglected even in conventional ML, even further. We conclude that there is a need for more transparent accounting and systematic evaluation of the impact of representing (informative) missingness on downstream performance.

With the continuous advancement of image generation technology, advanced models such as GPT-Image-1 and Qwen-Image have achieved remarkable text-to-image consistency and world knowledge However, these models still fall short in photorealistic image generation. Even on simple T2I tasks, they tend to produce " fake" images with distinct AI artifacts, often characterized by "overly smooth skin" and "oily facial sheens". To recapture the original goal of "indistinguishable-from-reality" generation, we propose RealGen, a photorealistic text-to-image framework. RealGen integrates an LLM component for prompt optimization and a diffusion model for realistic image generation. Inspired by adversarial generation, RealGen introduces a "Detector Reward" mechanism, which quantifies artifacts and assesses realism using both semantic-level and feature-level synthetic image detectors. We leverage this reward signal with the GRPO algorithm to optimize the entire generation pipeline, significantly enhancing image realism and detail. Furthermore, we propose RealBench, an automated evaluation benchmark employing Detector-Scoring and Arena-Scoring. It enables human-free photorealism assessment, yielding results that are more accurate and aligned with real user experience. Experiments demonstrate that RealGen significantly outperforms general models like GPT-Image-1 and Qwen-Image, as well as specialized photorealistic models like FLUX-Krea, in terms of realism, detail, and aesthetics. The code is available at https://github.com/yejy53/RealGen.

Test-time compute scaling has emerged as a powerful paradigm for enhancing mathematical reasoning in large language models (LLMs) by allocating additional computational resources during inference. However, current methods employ uniform resource distribution across all reasoning sub-problems, creating fundamental bottlenecks where challenging sub-problems receive insufficient attention while routine operations consume disproportionate resources. This uniform allocation creates performance bottlenecks where additional computational resources yield diminishing returns. Inspired by dual-process theory, we propose \textbf{SCALE} (Selective Resource Allocation), a framework that selectively allocates computational resources based on sub-problem difficulty. SCALE operates through four stages: (1) problem decomposition into sequential reasoning sub-problems, (2) difficulty assessment of each sub-problem to distinguish between routine operations and computationally challenging sub-problems, (3) selective processing mode assignment between System 1 for simple sub-problems and System 2 for complex ones, and (4) sequential execution with context propagation. By concentrating resources on challenging sub-problems while processing routine operations efficiently, SCALE achieves substantial performance improvements with superior resource utilization. Extensive experiments demonstrate that SCALE significantly outperforms uniform scaling baselines, achieving accuracy improvements of up to 13.75 percentage points (57.50% to 71.25% on AIME25) while reducing computational costs by 33%-53%, representing a major advance in test-time scaling that addresses fundamental limitations of current approaches.

Synthetic personae experiments have become a prominent method in Large Language Model alignment research, yet the representativeness and ecological validity of these personae vary considerably between studies. Through a review of 63 peer-reviewed studies published between 2023 and 2025 in leading NLP and AI venues, we reveal a critical gap: task and population of interest are often underspecified in persona-based experiments, despite personalization being fundamentally dependent on these criteria. Our analysis shows substantial differences in user representation, with most studies focusing on limited sociodemographic attributes and only 35% discussing the representativeness of their LLM personae. Based on our findings, we introduce a persona transparency checklist that emphasizes representative sampling, explicit grounding in empirical data, and enhanced ecological validity. Our work provides both a comprehensive assessment of current practices and practical guidelines to improve the rigor and ecological validity of persona-based evaluations in language model alignment research.

Recent work on autonomous scientific discovery has leveraged LLM-based agents to integrate problem specification, experiment planning, and execution into end-to-end systems. However, these frameworks are often confined to narrow application domains, offer limited real-time interaction with researchers, and lack principled mechanisms for determining when to halt exploration, resulting in inefficiencies, reproducibility challenges, and under-utilized human expertise. To address these gaps, we propose \textit{SelfAI}, a general multi-agent platform that combines a User Agent for translating high-level research objectives into standardized experimental configurations, a Cognitive Agent powered by LLMs with optimal stopping criteria to iteratively refine hyperparameter searches, and an Experiment Manager responsible for orchestrating parallel, fault-tolerant training workflows across heterogeneous hardware while maintaining a structured knowledge base for continuous feedback. We further introduce two novel evaluation metrics, Score and $\text{AUP}_D$, to quantify discovery efficiency and search diversity. Across regression, NLP, computer vision, scientific computing, medical imaging, and drug discovery benchmarks, SelfAI consistently achieves strong performance and reduces redundant trials compared to classical Bayesian optimization and LLM-based baselines, while enabling seamless interaction with human researchers.

Background Recent developments in artificial intelligence (AI), particularly Large Language Models (LLMs), have shown promise in automating previously unavailable aspects of student writing assessment and providing detailed, individuated feedback. Our previous research demonstrated that AI systems can reliably assess student writing using standardized rubrics, achieving consistency 2 rates of over 99% over five iterations (Heywood & Carrier, 2024). However, while these systems excel at providing holistic assessment using broad categories, their potential to provide detailed, granular feedback about specific writing errors has not yet been fully explored . This study builds upon our earlier work by developing and testing a sophisticated error classification system that can identify, categorize, and describe writing errors at both the word and sentence levels. The system employs a detailed taxonomy of errors based on established linguistic theory in the area of error classification (Corder, 1967, 1975, 1981; Richards, 1971, 1974; James, 1998). The AI analysis is implemented through carefully designed API calls to Claude 3.5 Sonnet in Python. With this enhanced error classification system, the present study analyzes an error ridden dialogue from an infamous text, English as she is spoke (Fonseca et al., 2004). We also provide the results of a review of the AI analysis by a human panel of experts.

Recent research has explored LLMs as scalable tools for relevance labeling, but studies indicate they are susceptible to priming effects, where prior relevance judgments influence later ones. Although psychological theories link personality traits to such biases, it is unclear whether simulated personalities in LLMs exhibit similar effects. We investigate how Big Five personality profiles in LLMs influence priming in relevance labeling, using multiple LLMs on TREC 2021 and 2022 Deep Learning Track datasets. Our results show that certain profiles, such as High Openness and Low Neuroticism, consistently reduce priming susceptibility. Additionally, the most effective personality in mitigating priming may vary across models and task types. Based on these findings, we propose personality prompting as a method to mitigate threshold priming, connecting psychological evidence with LLM-based evaluation practices.

Large Language Models' (LLMs) programming capabilities enable their participation in open-source games: a game-theoretic setting in which players submit computer programs in lieu of actions. These programs offer numerous advantages, including interpretability, inter-agent transparency, and formal verifiability; additionally, they enable program equilibria, solutions that leverage the transparency of code and are inaccessible within normal-form settings. We evaluate the capabilities of leading open- and closed-weight LLMs to predict and classify program strategies and evaluate features of the approximate program equilibria reached by LLM agents in dyadic and evolutionary settings. We identify the emergence of payoff-maximizing, cooperative, and deceptive strategies, characterize the adaptation of mechanisms within these programs over repeated open-source games, and analyze their comparative evolutionary fitness. We find that open-source games serve as a viable environment to study and steer the emergence of cooperative strategy in multi-agent dilemmas.