Large Language Models (LLMs) are widely used for downstream tasks such as tabular classification, where ensuring fairness in their outputs is critical for inclusivity, equal representation, and responsible AI deployment. This study introduces a novel approach to enhancing LLM performance and fairness through the concept of a dynamic validation set, which evolves alongside the test set, replacing the traditional static validation approach. We also propose an iterative algorithm, SMITE, to select optimal in-context examples, with each example set validated against its corresponding dynamic validation set. The in-context set with the lowest total error is used as the final demonstration set. Our experiments across four different LLMs show that our proposed techniques significantly improve both predictive accuracy and fairness compared to baseline methods. To our knowledge, this is the first study to apply dynamic validation in the context of in-context learning for LLMs.

Prompt engineering significantly influences the reliability and clinical utility of Large Language Models (LLMs) in medical applications. Current optimization approaches inadequately address domain-specific medical knowledge and safety requirements. This paper introduces EMPOWER, a novel evolutionary framework that enhances medical prompt quality through specialized representation learning, multi-dimensional evaluation, and structure-preserving algorithms. Our methodology incorporates: (1) a medical terminology attention mechanism, (2) a comprehensive assessment architecture evaluating clarity, specificity, clinical relevance, and factual accuracy, (3) a component-level evolutionary algorithm preserving clinical reasoning integrity, and (4) a semantic verification module ensuring adherence to medical knowledge. Evaluation across diagnostic, therapeutic, and educational tasks demonstrates significant improvements: 24.7% reduction in factually incorrect content, 19.6% enhancement in domain specificity, and 15.3% higher clinician preference in blinded evaluations. The framework addresses critical challenges in developing clinically appropriate prompts, facilitating more responsible integration of LLMs into healthcare settings.

Recent advances in the intrinsic reasoning capabilities of large language models (LLMs) have given rise to LLM-based agent systems that exhibit near-human performance on a variety of automated tasks. However, although these systems share similarities in terms of their use of LLMs, different reasoning frameworks of the agent system steer and organize the reasoning process in different ways. In this survey, we propose a systematic taxonomy that decomposes agentic reasoning frameworks and analyze how these frameworks dominate framework-level reasoning by comparing their applications across different scenarios. Specifically, we propose an unified formal language to further classify agentic reasoning systems into single-agent methods, tool-based methods, and multi-agent methods. After that, we provide a comprehensive review of their key application scenarios in scientific discovery, healthcare, software engineering, social simulation, and economics. We also analyze the characteristic features of each framework and summarize different evaluation strategies. Our survey aims to provide the research community with a panoramic view to facilitate understanding of the strengths, suitable scenarios, and evaluation practices of different agentic reasoning frameworks.

While large language models (LLMs) are rapidly advancing scientific research, they continue to struggle with core biological reasoning tasks essential for translational and biomedical discovery. To address this limitation, we created and curated eight comprehensive benchmark datasets comprising over 300,000 verifiable question-and-answer pairs, each targeting critical challenges in drug discovery including target druggability, modality suitability, and drug perturbation effects. Using this resource, we developed the OwkinZero models by post-training open-source LLMs through a Reinforcement Learning from Verifiable Rewards strategy. Our results demonstrate that specialized 8-32B OwkinZero models substantially outperform larger, state-of-the-art commercial LLMs on these biological benchmarks. Remarkably, we uncover evidence of a key aspect of generalization: specialist models trained on a single task consistently outperform their base models on previously unseen tasks. This generalization effect is further amplified in our comprehensive OwkinZero models, which were trained on a mixture of datasets and achieve even broader cross-task improvements. This study represents a significant step toward addressing the biological reasoning blind spot in current LLMs, demonstrating that targeted reinforcement learning on carefully curated data can unlock generalizable performance in specialized models, thereby accelerating AI-driven biological discovery.

In this paper, we introduce a novel learning paradigm for Adaptive Large Language Model (LLM) agents that eliminates the need for fine-tuning the underlying LLMs. Existing approaches are often either rigid, relying on static, handcrafted reflection workflows, or computationally intensive, requiring gradient updates of LLM model parameters. In contrast, our method enables low-cost continual adaptation via memory-based online reinforcement learning. We formalise this as a Memory-augmented Markov Decision Process (M-MDP), equipped with a neural case-selection policy to guide action decisions. Past experiences are stored in an episodic memory, either differentiable or non-parametric. The policy is continually updated based on environmental feedback through a memory rewriting mechanism, whereas policy improvement is achieved through efficient memory reading (retrieval). We instantiate our agent model in the deep research setting, namely \emph{Memento}, which attains top-1 on GAIA validation ($87.88\%$ Pass@$3$) and $79.40\%$ on the test set. It reaches $66.6\%$ F1 and $80.4\%$ PM on the DeepResearcher dataset, outperforming the state-of-the-art training-based method, while case-based memory adds $4.7\%$ to $9.6\%$ absolute points on out-of-distribution tasks. Our approach offers a scalable and efficient pathway for developing generalist LLM agents capable of continuous, real-time learning without gradient updates, advancing machine learning towards open-ended skill acquisition and deep research scenarios. The code is available at https://github.com/Agent-on-the-Fly/Memento.

High-quality mathematical and logical datasets with verifiable answers are essential for strengthening the reasoning capabilities of large language models (LLMs). While recent data augmentation techniques have facilitated the creation of large-scale benchmarks, existing LLM-generated datasets often suffer from limited reliability, diversity, and scalability. To address these challenges, we introduce PuzzleClone, a formal framework for synthesizing verifiable data at scale using Satisfiability Modulo Theories (SMT). Our approach features three key innovations: (1) encoding seed puzzles into structured logical specifications, (2) generating scalable variants through systematic variable and constraint randomization, and (3) ensuring validity via a reproduction mechanism. Applying PuzzleClone, we construct a curated benchmark comprising over 83K diverse and programmatically validated puzzles. The generated puzzles span a wide spectrum of difficulty and formats, posing significant challenges to current state-of-the-art models. We conduct post training (SFT and RL) on PuzzleClone datasets. Experimental results show that training on PuzzleClone yields substantial improvements not only on PuzzleClone testset but also on logic and mathematical benchmarks. Post training raises PuzzleClone average from 14.4 to 56.2 and delivers consistent improvements across 7 logic and mathematical benchmarks up to 12.5 absolute percentage points (AMC2023 from 52.5 to 65.0). Our code and data are available at https://github.com/HiThink-Research/PuzzleClone.

DPO (Direct Preference Optimization) has become a widely used offline preference optimization algorithm due to its simplicity and training stability. However, DPO is prone to overfitting and collapse. To address these challenges, we propose Linear Preference Optimization (LPO), a novel alignment framework featuring three key innovations. First, we introduce gradient decoupling by replacing the log-sigmoid function with an absolute difference loss, thereby isolating the optimization dynamics. Second, we improve stability through an offset constraint combined with a positive regularization term to preserve the chosen response quality. Third, we implement controllable rejection suppression using gradient separation with straightforward estimation and a tunable coefficient that linearly regulates the descent of the rejection probability. Through extensive experiments, we demonstrate that LPO consistently improves performance on various tasks, including general text tasks, math tasks, and text-to-speech (TTS) tasks. These results establish LPO as a robust and tunable paradigm for preference alignment, and we release the source code, models, and training data publicly.

Tokenization remains a fundamental yet underexplored bottleneck in natural language processing, with strategies largely static despite remarkable progress in model architectures. We present SupraTok, a novel tokenization architecture that reimagines subword segmentation through three innovations: cross-boundary pattern learning that discovers multi-word semantic units, entropy-driven data curation that optimizes training corpus quality, and multi-phase curriculum learning for stable convergence. Our approach extends Byte-Pair Encoding by learning "superword" tokens, coherent multi-word expressions that preserve semantic unity while maximizing compression efficiency. SupraTok achieves 31% improvement in English tokenization efficiency (5.91 versus 4.51 characters per token) compared to OpenAI's o200k tokenizer and 30% improvement over Google's Gemma 3 tokenizer (256k vocabulary), while maintaining competitive performance across 38 languages. When integrated with a GPT-2 scale model (124M parameters) trained on 10 billion tokens from the FineWeb-Edu dataset, SupraTok yields 8.4% improvement on HellaSWAG and 9.5% on MMLU benchmarks without architectural modifications. While these results are promising at this scale, further validation at larger model scales is needed. These findings suggest that efficient tokenization can complement architectural innovations as a path to improved language model performance.

Algorithmic discovery has traditionally relied on human ingenuity and extensive experimentation. Here we investigate whether a prominent scientific computing algorithm, the Kalman Filter, can be discovered through an automated, data-driven, evolutionary process that relies on Cartesian Genetic Programming (CGP) and Large Language Models (LLM). We evaluate the contributions of both modalities (CGP and LLM) in discovering the Kalman filter under varying conditions. Our results demonstrate that our framework of CGP and LLM-assisted evolution converges to near-optimal solutions when Kalman optimality assumptions hold. When these assumptions are violated, our framework evolves interpretable alternatives that outperform the Kalman filter. These results demonstrate that combining evolutionary algorithms and generative models for interpretable, data-driven synthesis of simple computational modules is a potent approach for algorithmic discovery in scientific computing.

Evaluating large language models (LLMs) in specific domain like tourism remains challenging due to the prohibitive cost of annotated benchmarks and persistent issues like hallucinations. We propose $\textbf{L}$able-Free $\textbf{E}$valuation of LLM on $\textbf{T}$ourism using Expert $\textbf{T}$ree-$\textbf{o}$f-$\textbf{T}$hought (LETToT), a framework that leverages expert-derived reasoning structures-instead of labeled data-to access LLMs in tourism. First, we iteratively refine and validate hierarchical ToT components through alignment with generic quality dimensions and expert feedback. Results demonstrate the effectiveness of our systematically optimized expert ToT with 4.99-14.15\% relative quality gains over baselines. Second, we apply LETToT's optimized expert ToT to evaluate models of varying scales (32B-671B parameters), revealing: (1) Scaling laws persist in specialized domains (DeepSeek-V3 leads), yet reasoning-enhanced smaller models (e.g., DeepSeek-R1-Distill-Llama-70B) close this gap; (2) For sub-72B models, explicit reasoning architectures outperform counterparts in accuracy and conciseness ($p<0.05$). Our work established a scalable, label-free paradigm for domain-specific LLM evaluation, offering a robust alternative to conventional annotated benchmarks.