Relatively many past AI safety discussions have centered around the dangers of unbounded utility maximisation by RL agents, illustrated by scenarios like the "paperclip maximiser" or by specification gaming in general. Unbounded maximisation is problematic for many reasons. We wanted to verify whether these RL runaway optimisation problems are still relevant with LLMs as well. Turns out, strangely, this is indeed clearly the case. The problem is not that the LLMs just lose context or become incoherent. The problem is that in various scenarios, LLMs lose context in very specific ways, which systematically resemble runaway optimisers in the following distinct ways: 1) Ignoring homeostatic targets and "defaulting" to unbounded maximisation instead. 2) It is equally concerning that the "default" meant also reverting back to single-objective optimisation. Our findings also suggest that long-running scenarios are important. Systematic failures emerge after periods of initially successful behaviour. In some trials the LLMs were successful until the end. This means, while current LLMs do conceptually grasp biological and economic alignment, they exhibit randomly triggered problematic behavioural tendencies under sustained long-running conditions, particularly involving multiple or competing objectives. Once they flip, they usually do not recover. Even though LLMs look multi-objective and bounded on the surface, the underlying mechanisms seem to be actually still biased towards being single-objective and unbounded.

In the United States, schizophrenia (SCZ) carries a race and sex disparity that may be explained by clinician bias - a belief held by a clinician about a patient that prevents impartial clinical decision making. The emergency department (ED) is marked by higher rates of stress that lead to clinicians relying more on implicit biases during decision making. In this work, we considered a large cohort of psychiatric patients in the ED from the Mount Sinai Health System (MSHS) in New York City to investigate the effects of clinician bias on SCZ diagnosis while controlling for known risk factors and patient sociodemographic information. Clinician bias was quantified as the ratio of negative to total sentences within a patient's first ED note. We utilized a logistic regression to predict SCZ diagnosis given patient race, sex, age, history of trauma or substance use disorder, and the ratio of negative sentences. Our findings showed that an increased ratio of negative sentences is associated with higher odds of obtaining a SCZ diagnosis [OR (95% CI)=1.408 (1.361-1.456)]. Identifying as male [OR (95% CI)=1.112 (1.055-1.173)] or Black [OR (95% CI)=1.081(1.031-1.133)] increased one's odds of being diagnosed with SCZ. However, from an intersectional lens, Black female patients with high SES have the highest odds of obtaining a SCZ diagnosis [OR (95% CI)=1.629 (1.535-1.729)]. Results such as these suggest that SES does not act as a protective buffer against SCZ diagnosis in all patients, demanding more attention to the quantification of health disparities. Lastly, we demonstrated that clinician bias is operational with real world data and related to increased odds of obtaining a stigmatizing diagnosis such as SCZ.

We present a comparative docking experiment that aligns human-subject interview data with large language model (LLM)-driven synthetic personas to evaluate fidelity, divergence, and blind spots in AI-enabled simulation. Fifteen early-stage startup founders were interviewed about their hopes and concerns regarding AI-powered validation, and the same protocol was replicated with AI-generated founder and investor personas. A structured thematic synthesis revealed four categories of outcomes: (1) Convergent themes - commitment-based demand signals, black-box trust barriers, and efficiency gains were consistently emphasized across both datasets; (2) Partial overlaps - founders worried about outliers being averaged away and the stress of real customer validation, while synthetic personas highlighted irrational blind spots and framed AI as a psychological buffer; (3) Human-only themes - relational and advocacy value from early customer engagement and skepticism toward moonshot markets; and (4) Synthetic-only themes - amplified false positives and trauma blind spots, where AI may overstate adoption potential by missing negative historical experiences. We interpret this comparative framework as evidence that LLM-driven personas constitute a form of hybrid social simulation: more linguistically expressive and adaptable than traditional rule-based agents, yet bounded by the absence of lived history and relational consequence. Rather than replacing empirical studies, we argue they function as a complementary simulation category - capable of extending hypothesis space, accelerating exploratory validation, and clarifying the boundaries of cognitive realism in computational social science.

Evaluating large language models (LLMs) on their ability to generate high-quality, accurate, situationally aware answers to clinical questions requires going beyond conventional benchmarks to assess how these systems behave in complex, high-stake clincal scenarios. Traditional evaluations are often limited to multiple-choice questions that fail to capture essential competencies such as contextual reasoning, awareness and uncertainty handling etc. To address these limitations, we evaluate our agentic, RAG-based clinical support assistant, DR.INFO, using HealthBench, a rubric-driven benchmark composed of open-ended, expert-annotated health conversations. On the Hard subset of 1,000 challenging examples, DR.INFO achieves a HealthBench score of 0.51, substantially outperforming leading frontier LLMs (GPT-5, o3, Grok 3, GPT-4, Gemini 2.5, etc.) across all behavioral axes (accuracy, completeness, instruction following, etc.). In a separate 100-sample evaluation against similar agentic RAG assistants (OpenEvidence, Pathway.md), it maintains a performance lead with a health-bench score of 0.54. These results highlight DR.INFOs strengths in communication, instruction following, and accuracy, while also revealing areas for improvement in context awareness and completeness of a response. Overall, the findings underscore the utility of behavior-level, rubric-based evaluation for building a reliable and trustworthy AI-enabled clinical support assistant.

Scholarly work and communication encompass the entire system in which research and creative works are created, evaluated for quality, disseminated to the academic community and beyond, used, and preserved for future use. It includes formal publications, such as journal articles and books, as well as informal sharing through preprints, conference presentations, data sharing, and broader engagement with scholarly works and research outputs. Scholarly knowledge serves as the primary engine of progress, shaping our world and guiding our collective future. It forms the backbone of technological advancement, public health systems, and sustainable environmental practices. Obtained through rigorous methods of observation, experimentation, and validation, it is a reliable resource that helps societies solve complex problems and improve the quality of life by achieving sustainable development goals (SDGs) [6]. To be truly useful, scholarly knowledge must first be systematically extracted and organized. However, the scholarly community of today faces the problem of an overload of scientific papers in their respective domains. There is an increasing number of papers published every year (currently, 3 million), in addition to more than 200 million papers that have already been published . This gives rise to the research question: "How can we provide a reliable and living scholarly knowledge base that empowers researchers to query, synthesize, and analyze the vast body of scholarly knowledge?"

Large Language Models (LLMs) often generate repetitive and monotonous outputs, especially in tasks like story generation, due to limited creative diversity when given the same input prompt. To address this challenge, we propose a novel decoding strategy, Avoidance Decoding, that modifies token logits by penalizing similarity to previously generated outputs, thereby encouraging more diverse multi-branch stories. This penalty adaptively balances two similarity measures: (1) Concept-level Similarity Penalty, which is prioritized in early stages to diversify initial story concepts, and (2) Narrative-level Similarity Penalty, which is increasingly emphasized later to ensure natural yet diverse plot development. Notably, our method achieves up to 2.6 times higher output diversity and reduces repetition by an average of 30% compared to strong baselines, while effectively mitigating text degeneration. Furthermore, we reveal that our method activates a broader range of neurons, demonstrating that it leverages the model's intrinsic creativity.

Existing agents for solving tasks such as ML engineering rely on prompting powerful language models. As a result, these agents do not improve with more experience. In this paper, we show that agents backed by weaker models that improve via reinforcement learning (RL) can outperform agents backed by much larger, but static models. We identify two major challenges with RL in this setting. First, actions can take a variable amount of time (e.g., executing code for different solutions), which leads to asynchronous policy gradient updates that favor faster but suboptimal solutions. To tackle variable-duration actions, we propose duration-aware gradient updates in a distributed asynchronous RL framework to amplify high-cost but high-reward actions. Second, using only test split performance as a reward provides limited feedback. A program that is nearly correct is treated the same as one that fails entirely. To address this, we propose environment instrumentation to offer partial credit, distinguishing almost-correct programs from those that fail early (e.g., during data loading). Environment instrumentation uses a separate static language model to insert print statement to an existing program to log the agent's experimental progress, from which partial credit can be extracted as reward signals for learning. Our experimental results on MLEBench suggest that performing gradient updates on a much smaller model (Qwen2.5-3B) trained with RL outperforms prompting a much larger model (Claude-3.5-Sonnet) with agent scaffolds, by an average of 22% across 12 Kaggle tasks.

We introduce CoreThink, a state-of-the-art Reasoning Layer built upon a novel reasoning method called General Symbolics. This approach diverges from reasoning paradigms such as test-time scaling, Supervised Fine-Tuning (SFT), and Reinforcement Learning with Verifiable Rewards (RLVR). CoreThink General Symbolic Reasoner (GSR) is specifically structured around three key use cases: tool-calling, code generation, and planning, demonstrating exemplary performance across a total of seven benchmarks in their respective areas. Notably, we are achieving SOTA scores of 66.66% on Livecodebench v6, 89% on Instruction-Following Evals, and 24.4% on ARC-AGI-2. We also present an agentic coding IDE, developed using the principles of General Symbolics, which achieves a state-of-the-art accuracy of 62.3% on SWE-Bench Lite. We are able to achieve these improvements without any fine-tuning or training costs. Our Reasoning Layer is designed to provide a pure performance uplift, ensuring that a model's accuracy on reasoning tasks is never negatively impacted. We argue that incumbent methods will eventually lead to diminishing returns in LLM performance, necessitating the development of new reasoning techniques. This technical report details our approach at a high level and the availability of the CoreThink models for reasoning-intensive use cases.

Real-world adoption of closed-loop insulin delivery systems (CLIDS) in type 1 diabetes remains low, driven not by technical failure, but by diverse behavioral, psychosocial, and social barriers. We introduce ChatCLIDS, the first benchmark to rigorously evaluate LLM-driven persuasive dialogue for health behavior change. Our framework features a library of expert-validated virtual patients, each with clinically grounded, heterogeneous profiles and realistic adoption barriers, and simulates multi-turn interactions with nurse agents equipped with a diverse set of evidence-based persuasive strategies. ChatCLIDS uniquely supports longitudinal counseling and adversarial social influence scenarios, enabling robust, multi-dimensional evaluation. Our findings reveal that while larger and more reflective LLMs adapt strategies over time, all models struggle to overcome resistance, especially under realistic social pressure. These results highlight critical limitations of current LLMs for behavior change, and offer a high-fidelity, scalable testbed for advancing trustworthy persuasive AI in healthcare and beyond.

We present L-MARS (Legal Multi-Agent Workflow with Orchestrated Reasoning and Agentic Search), a system that reduces hallucination and uncertainty in legal question answering through coordinated multi-agent reasoning and retrieval. Unlike single-pass retrieval-augmented generation (RAG), L-MARS decomposes queries into subproblems, issues targeted searches across heterogeneous sources (Serper web, local RAG, CourtListener case law), and employs a Judge Agent to verify sufficiency, jurisdiction, and temporal validity before answer synthesis. This iterative reasoning-search-verification loop maintains coherence, filters noisy evidence, and grounds answers in authoritative law. We evaluated L-MARS on LegalSearchQA, a new benchmark of 200 up-to-date multiple choice legal questions in 2025. Results show that L-MARS substantially improves factual accuracy, reduces uncertainty, and achieves higher preference scores from both human experts and LLM-based judges. Our work demonstrates that multi-agent reasoning with agentic search offers a scalable and reproducible blueprint for deploying LLMs in high-stakes domains requiring precise legal retrieval and deliberation.