A new arXiv preprint introduces a framework that converts static, single-turn language model tasks into dynamic, multi-turn conversations where user intent changes over time. The study finds that leading language models, which perform well in static settings, experience significant performance drops when required to track and respond to evolving user intent, revealing a consistent limitation across model families.
Why it matters: This work highlights a key shortcoming in current LLMs that could impact their effectiveness as collaborative agents in real-world, interactive scenarios.
A new arXiv preprint introduces ConfidenceBench, a benchmark designed to evaluate how well large language models (LLMs) verbalize their confidence in answers using Brier scores. Testing 15 prominent LLMs, the study finds that top-performing models in accuracy are not always the best-calibrated in confidence, with some models showing severe miscalibration. The benchmark works via prompting, making it applicable to both open- and closed-source models.
Why it matters: This work highlights that confidence calibration is a distinct and critical aspect of LLM reliability, with implications for deploying these models in high-stakes or safety-sensitive applications.
A new arXiv preprint introduces Int-Bench, a benchmark designed to evaluate how large language models (LLMs) act as tutors during problem-solving. The study finds that LLMs, when compared to human tutors, tend to intervene both more frequently and earlier, often providing full solutions instead of incremental hints. This behavior suggests that current AI assistants may prioritize immediate task completion over fostering deeper reasoning or learning.
Why it matters: As AI tutors become more widely used, understanding their intervention patterns is important to ensure they support, rather than undermine, genuine learning.
A new arXiv preprint introduces TopoGuard, a set of graph theory-based methods designed to defend Retrieval Augmented Generation (RAG) systems against split-knowledge attacks. These attacks involve injecting individually benign documents that, when combined, create misleading associations undetectable by existing per-document filters. TopoGuard constructs a semantic similarity graph from retrieved documents to identify suspicious contexts, and in experiments on HotpotQA, it detected 21 times more attacks than LlamaGuard-2-8B at a 1% false positive rate.
Why it matters: The work highlights a previously under-addressed vulnerability in RAG systems and proposes a practical, efficient defense that could improve the security of AI applications relying on document retrieval.
A new arXiv preprint introduces the PhantomFill benchmark, showing that large language models (LLMs) frequently fabricate information when required to fill structured fields, such as in JSON forms, even when they would otherwise admit to having no data in free text. Across 13 models, the study found that required fields led to coerced fabrication nearly every time, while free-text responses were mostly honest. The benchmark quantifies this phenomenon and highlights that current evaluation practices may overlook this critical failure mode.
Why it matters: This finding exposes a significant reliability risk for LLMs in real-world applications that depend on structured outputs, suggesting that required fields can silently induce systematic hallucination.
A new arXiv preprint reports that quantizing large language models—a common step to make them more efficient—can introduce measurable bias in open-ended text generation, even when standard safety checks show no change. The study finds that quantized models are more likely to produce stereotyped responses across eight languages, a phenomenon not detected by typical refusal rate or multiple-choice fairness metrics. The authors introduce QuantiBias, a benchmark designed to reveal this hidden bias.
Why it matters: This work highlights a previously overlooked risk in LLM deployment, showing that quantization can silently increase bias without triggering standard safety evaluations.
A new arXiv preprint demonstrates that large language models (LLMs), when paired with prediction market data, can quantify how information sources bias strategic predictions. Analyzing 111 Ukraine-related prediction markets, the study finds that English news context systematically biases territorial forecasts, with predictions favoring territorial capture being incorrect 64–72% of the time. The analysis suggests that this bias originates mainly from the text sources rather than the LLMs themselves, and persists across multiple model architectures.
Why it matters: The findings highlight that AI systems relying on real-world text sources can inherit and propagate significant biases, which could impact strategic decision-making in high-stakes domains.
A new arXiv preprint examines political bias in large language models (LLMs) using both abstract policy questionnaires and real-world Swiss referenda. The study finds that while LLMs appear left-leaning on surveys, their responses to actual referenda are more centrist and sometimes show a general aversion to change rather than a clear left-right bias. The language in which questions are posed can also significantly affect model responses.
Why it matters: This suggests that previous claims of consistent leftward bias in LLMs may not generalize to real-world decision contexts, raising questions about how AI political alignment should be measured and interpreted.
A new arXiv preprint introduces the Frontier Financial Judgement benchmark, developed with professional equity analysts, to test AI agents' ability to identify news that could impact stock valuations. The best-performing AI agent matched expert human labels just 52.4% of the time, and false-positive rates varied widely across models. The benchmark consists of 656 items, including both synthetic and real news articles, designed to reflect real-world financial information challenges.
Why it matters: The results highlight a significant gap in current AI capabilities for reliably filtering valuation-relevant financial news, which is critical for safe and effective deployment in equity analysis.
A new arXiv preprint presents a controlled evaluation of how different evidence interfaces affect the validity of large language model (LLM)-generated policy audit reports. Using 60 AGORA policy cases, the study finds that reports can appear substantively plausible even when citing irrelevant internal evidence, especially in a shuffled control condition. The hybrid evidence interface was rated most useful by human reviewers, and the findings suggest that simply providing internal model access does not guarantee meaningful transparency.
Why it matters: The work highlights a significant governance risk: AI-generated audit reports may sound convincing while relying on irrelevant or misleading evidence, emphasizing the need for careful evidence design in AI oversight.
A systematic human evaluation of ten leading large language models (LLMs) across 62 multidomain questions found that all models exhibit response drift—deviations from expert-validated references. Most models showed high rates of drift (78-81%), while two had notably lower rates (47-49%). The extent and pattern of drift varied by domain and question, and automated metrics accounted for less than 2% of the variance in human judgments.
Why it matters: The findings highlight a widespread and domain-dependent limitation in current LLMs that cannot be reliably detected by automated metrics alone, underscoring the need for human evaluation in assessing model reliability.
A new arXiv preprint finds that large language models (LLMs) often change their answers' correctness when mathematically equivalent problems are presented in different surface forms, such as story, symbolic, or equation formats. Introducing code-based reasoning, where models generate and execute Python code, improves performance for some LLMs but does not fully resolve this sensitivity; errors instead shift to issues like protocol violations or code execution failures. The study highlights that even subtle changes in problem wording can significantly affect LLM reliability.
Why it matters: This work suggests that the way problems are presented remains a critical factor in LLM reliability for mathematical tasks, with implications for evaluation and deployment in real-world applications.
A new arXiv preprint introduces Rushes, a dataset of over 44,000 decision events from interactive narrative games, capturing how thousands of users make sequential choices. The study finds that leading large language models, including GPT-5, do not outperform simple baselines in predicting individual user choices, highlighting a persistent 'Engagement Gap.' This suggests that current alignment methods, which optimize for population-level preferences, may be inadequate for capturing diverse, context-dependent user behaviors.
Why it matters: The findings raise questions about the ability of current AI alignment techniques to personalize responses and adapt to individual users, a key challenge for future AI systems.
A new framework for generating synthetic lung CT slices combines Optimal Transport Conditional Flow Matching with geometric latent space filtering to enhance privacy. While the method reduces direct image memorization and some privacy risks, official challenge results show that deeper anatomical identity can still be inferred from generated images. This suggests that current filtering techniques are insufficient to fully protect patient identity in medical image synthesis.
Why it matters: The findings underscore a significant limitation in privacy-preserving medical image generation, with implications for data sharing and regulatory compliance.
A new arXiv preprint audits five commonly used diversity metrics for selecting large language models (LLMs) in ensemble majority voting. The study finds that these metrics are largely entangled with model capability, rather than measuring true diversity. After controlling for capability, only a modest residual link remains between shared errors and ensemble voting gains, challenging the assumption that diversity metrics reliably predict ensemble improvement.
Why it matters: This result questions the reliability of widely used diversity metrics in LLM ensemble construction, potentially impacting how practitioners combine models for better performance.
A new arXiv preprint introduces Merge-Adversarial Training, a method designed to make watermarks in open-source large language models (LLMs) more robust to removal by model merging—a common post-training modification. The approach reportedly increases watermark detection rates by up to 51 percentage points at a 1% false positive rate, without degrading model performance. The study also evaluates watermark durability across several realistic model merging scenarios.
Why it matters: This work addresses a key challenge in tracing the origins of text generated by open-source LLMs, even after models are merged or modified post-release.
A new arXiv preprint reports that while human-written natural-language policies can significantly improve the performance of frozen large language model (LLM) agents in the TextWorldExpress environment, the TextGrad method—which attempts to learn such policies from agent experience using natural-language feedback—does not yield better results than fixed prompting. The study finds that the main obstacle is not in executing policy updates, but in reliably generating and selecting effective policies from agent trajectories, even when using richer feedback or advanced search techniques.
Why it matters: This highlights a key limitation in current approaches to learning agent policies from experience using natural-language feedback, suggesting that simply leveraging agent trajectories is insufficient for policy optimization in LLM agents.
A new arXiv preprint finds that large language models (LLMs) are unreliable at distinguishing their own generated content from human-written responses in short-answer educational tasks. While LLMs perform well at detecting AI-generated code and longer reflective writing, they often misclassify their own short answers as more human-like than authentic student work. The study also shows that prompt variations can significantly affect detection accuracy in reflective writing, but have less impact on programming tasks.
Why it matters: This highlights a critical limitation for using LLMs as automated detectors of AI-generated student work, especially for short-answer formats that are common in education.
A new arXiv preprint introduces Instruct-FD, a benchmark designed to test whether full-duplex spoken dialogue systems can reliably follow explicit turn-taking instructions. When evaluated on six leading systems, the highest adherence rate was only 64.4%, with proactive conversational behaviors such as backchanneling and interruption posing particular difficulties. The results suggest that current systems have significant limitations in controllable turn management.
Why it matters: This highlights a key challenge for deploying adaptable full-duplex speech systems in real-world applications where flexible turn-taking is required.
A new arXiv preprint reports that while mean cross-positional attention degradation in transformer language models follows a consistent exponential-then-plateau pattern, it does not causally limit contextual retrieval. The study, spanning several popular LLM architectures, finds that interventions designed to boost attention to function tokens do not improve—and can sometimes harm—model performance. The results suggest that function tokens matter for their hidden state computations rather than the attention they receive.
Why it matters: This challenges common assumptions about the causal role of attention patterns in LLM interpretability and optimization, with potential implications for model analysis and efficiency strategies.