Training Large Language Models (LLMs) for multi-turn Tool-Integrated Reasoning (TIR) - where models iteratively reason, generate code, and verify through execution - remains challenging for existing reinforcement learning (RL) approaches. Current RL methods, exemplified by Group Relative Policy Optimization (GRPO), suffer from coarse-grained, trajectory-level rewards that provide insufficient learning signals for complex multi-turn interactions, leading to training stagnation. To address this issue, we propose Group Turn Policy Optimization (GTPO), a novel RL algorithm specifically designed for training LLMs on multi-turn TIR tasks. GTPO introduces three key innovations: (1) turn-level reward assignment that provides fine-grained feedback for individual turns, (2) return-based advantage estimation where normalized discounted returns are calculated as advantages, and (3) self-supervised reward shaping that exploits self-supervision signals from generated code to densify sparse binary outcome-based rewards. Our comprehensive evaluation demonstrates that GTPO outperforms GRPO by 3.0% on average across diverse reasoning benchmarks, establishing its effectiveness for advancing complex mathematical reasoning in the real world.

Contemporary machine learning models, including large language models, exhibit remarkable capabilities in static tasks yet falter in non-stationary environments due to rigid architectures that hinder continual adaptation and lifelong learning. Building upon the nested learning paradigm, which decomposes models into multi-level optimization problems with fixed update frequencies, this work proposes dynamic nested hierarchies as the next evolutionary step in advancing artificial intelligence and machine learning. Dynamic nested hierarchies empower models to autonomously adjust the number of optimization levels, their nesting structures, and update frequencies during training or inference, inspired by neuroplasticity to enable self-evolution without predefined constraints. This innovation addresses the anterograde amnesia in existing models, facilitating true lifelong learning by dynamically compressing context flows and adapting to distribution shifts. Through rigorous mathematical formulations, theoretical proofs of convergence, expressivity bounds, and sublinear regret in varying regimes, alongside empirical demonstrations of superior performance in language modeling, continual learning, and long-context reasoning, dynamic nested hierarchies establish a foundational advancement toward adaptive, general-purpose intelligence.

Modeling genomic sequences faces two unsolved challenges: the information density varies widely across different regions, while there is no clearly defined minimum vocabulary unit. Relying on either four primitive bases or independently designed DNA tokenizers, existing approaches with naive masked language modeling pre-training often fail to adapt to the varying complexities of genomic sequences. Leveraging Token Merging techniques, this paper introduces a hierarchical architecture that jointly optimizes a dynamic ge-nomic tokenizer and latent Transformers with context-aware pre-training tasks. As for network structures, the tokeniza-tion module automatically chunks adjacent bases into words by stacking multiple layers of the differentiable token merging blocks with local-window constraints, then a Latent Encoder captures the global context of these merged words by full-attention blocks. Symmetrically employing a Latent Decoder and a Local Decoder, MergeDNA learns with two pre-training tasks: Merged Token Reconstruction simultaneously trains the dynamic tokenization module and adaptively filters important tokens, while Adaptive Masked Token Modeling learns to predict these filtered tokens to capture informative contents. Extensive experiments show that MergeDNA achieves superior performance on three popular DNA benchmarks and several multi-omics tasks with fine-tuning or zero-shot evaluation, outperforming typical tokenization methods and large-scale DNA foundation models.

IT environments typically have logging mechanisms to monitor system health and detect issues. However, the huge volume of generated logs makes manual inspection impractical, highlighting the importance of automated log analysis in IT Software Support. In this paper, we propose a log analytics tool that leverages Large Language Models (LLMs) for log data processing and issue diagnosis, enabling the generation of automated insights and summaries. We further present a novel approach for efficiently running LLMs on CPUs to process massive log volumes in minimal time without compromising output quality. We share the insights and lessons learned from deployment of the tool - in production since March 2024 - scaled across 70 software products, processing over 2000 tickets for issue diagnosis, achieving a time savings of 300+ man hours and an estimated $15, 444 per month in manpower costs compared to the traditional log analysis practices.

Manual grading of programming assignments in introductory computer science courses can be time-consuming and prone to inconsistencies. While unit testing is commonly used for automatic evaluation, it typically follows a binary pass/fail model and does not give partial marks. Recent advances in large language models (LLMs) offer the potential for automated, scalable, and more objective grading. This paper compares two AI-based grading techniques: \textit{Direct}, where the AI model applies a rubric directly to student code, and \textit{Reverse} (a newly proposed approach), where the AI first fixes errors, then deduces a grade based on the nature and number of fixes. Each method was evaluated on both the instructor's original grading scale and a tenfold expanded scale to assess the impact of range on AI grading accuracy. To assess their effectiveness, AI-assigned scores were evaluated against human tutor evaluations on a range of coding problems and error types. Initial findings suggest that while the Direct approach is faster and straightforward, the Reverse technique often provides a more fine-grained assessment by focusing on correction effort. Both methods require careful prompt engineering, particularly for allocating partial credit and handling logic errors. To further test consistency, we also used synthetic student code generated using Gemini Flash 2.0, which allowed us to evaluate AI graders on a wider range of controlled error types and difficulty levels. We discuss the strengths and limitations of each approach, practical considerations for prompt design, and future directions for hybrid human-AI grading systems that aim to improve consistency, efficiency, and fairness in CS courses.

Automatic algorithm configuration tools such as irace efficiently tune parameter values but leave algorithmic code unchanged. This paper introduces a first version of irace-evo, an extension of irace that integrates code evolution through large language models (LLMs) to jointly explore parameter and code spaces. The proposed framework enables multi-language support (e.g., C++, Python), reduces token consumption via progressive context management, and employs the Always-From-Original principle to ensure robust and controlled code evolution. We evaluate irace-evo on the Construct, Merge, Solve & Adapt (CMSA) metaheuristic for the Variable-Sized Bin Packing Problem (VSBPP). Experimental results show that irace-evo can discover new algorithm variants that outperform the state-of-the-art CMSA implementation while maintaining low computational and monetary costs. Notably, irace-evo generates competitive algorithmic improvements using lightweight models (e.g., Claude Haiku 3.5) with a total usage cost under 2 euros. These results demonstrate that coupling automatic configuration with LLM-driven code evolution provides a powerful, cost-efficient avenue for advancing heuristic design and metaheuristic optimization.

Subnational location data of disaster events are critical for risk assessment and disaster risk reduction. Disaster databases such as EM-DAT often report locations in unstructured textual form, with inconsistent granularity or spelling, that make it difficult to integrate with spatial datasets. We present a fully automated LLM-assisted workflow that processes and cleans textual location information using GPT-4o, and assigns geometries by cross-checking three independent geoinformation repositories: GADM, OpenStreetMap and Wikidata. Based on the agreement and availability of these sources, we assign a reliability score to each location while generating subnational geometries. Applied to the EM-DAT dataset from 2000 to 2024, the workflow geocodes 14,215 events across 17,948 unique locations. Unlike previous methods, our approach requires no manual intervention, covers all disaster types, enables cross-verification across multiple sources, and allows flexible remapping to preferred frameworks. Beyond the dataset, we demonstrate the potential of LLMs to extract and structure geographic information from unstructured text, offering a scalable and reliable method for related analyses.

We present Ask WhAI, a systems-level framework for inspecting and perturbing belief states in multi-agent interactions. The framework records and replays agent interactions, supports out-of-band queries into each agent's beliefs and rationale, and enables counterfactual evidence injection to test how belief structures respond to new information. We apply the framework to a medical case simulator notable for its multi-agent shared memory (a time-stamped electronic medical record, or EMR) and an oracle agent (the LabAgent) that holds ground truth lab results revealed only when explicitly queried. We stress-test the system on a multi-specialty diagnostic journey for a child with an abrupt-onset neuropsychiatric presentation. Large language model agents, each primed with strong role-specific priors ("act like a neurologist", "act like an infectious disease specialist"), write to a shared medical record and interact with a moderator across sequential or parallel encounters. Breakpoints at key diagnostic moments enable pre- and post-event belief queries, allowing us to distinguish entrenched priors from reasoning or evidence-integration effects. The simulation reveals that agent beliefs often mirror real-world disciplinary stances, including overreliance on canonical studies and resistance to counterevidence, and that these beliefs can be traced and interrogated in ways not possible with human experts. By making such dynamics visible and testable, Ask WhAI offers a reproducible way to study belief formation and epistemic silos in multi-agent scientific reasoning.

Large language models (LLMs) have achieved remarkable results on tasks framed as reasoning problems, yet their true ability to perform procedural reasoning, executing multi-step, rule-based computations remains unclear. Unlike algorithmic systems, which can deterministically execute long-horizon symbolic procedures, LLMs often degrade under extended reasoning chains, but there is no controlled, interpretable benchmark to isolate and measure this collapse. We introduce Finite-State Machine (FSM) Execution as a minimal, fully interpretable framework for evaluating the procedural reasoning capacity of LLMs. In our setup, the model is given an explicit FSM definition and must execute it step-by-step given input actions, maintaining state consistency over multiple turns. This task requires no world knowledge, only faithful application of deterministic transition rules, making it a direct probe of the model's internal procedural fidelity. We measure both Turn Accuracy and Task Accuracy to disentangle immediate computation from cumulative state maintenance. Empirical results reveal systematic degradation as task horizon or branching complexity increases. Models perform significantly worse when rule retrieval involves high branching factors than when memory span is long. Larger models show improved local accuracy but remain brittle under multi-step reasoning unless explicitly prompted to externalize intermediate steps. FSM-based evaluation offers a transparent, complexity-controlled probe for diagnosing this failure mode and guiding the design of inductive biases that enable genuine long-horizon procedural competence. By grounding reasoning in measurable execution fidelity rather than surface correctness, this work helps establish a rigorous experimental foundation for understanding and improving the algorithmic reliability of LLMs.

With this paper, we survey techniques for improving the predictive accuracy of pretrained large language models by allocating additional compute at inference time. In categorizing test-time scaling methods, we place special emphasis on how a problem is decomposed into subproblems and on the topological organization of these subproblems whether sequential, parallel, or tree-structured. This perspective allows us to unify diverse approaches such as Chain-of-Thought, Branch-Solve-Merge, and Tree-of-Thought under a common lens. We further synthesize existing analyses of these techniques, highlighting their respective strengths and weaknesses, and conclude by outlining promising directions for future research