The paper introduces StyleShield, a novel flow matching framework for conditional text style transfer that exposes the fragility of AI-generated content (AIGC) detectors. It operates in continuous token embedding space to blur the statistical boundary between human and AI writing, challenging the reliability of current detection services.
This paper introduces H-probes, linear probes designed to extract hierarchical structure, specifically depth and pairwise distance, from the latent representations of large language models. The research shows these probes robustly find low-dimensional subspaces crucial for performance in synthetic tree traversal tasks, generalizing well both within and out-of-domain.
O trabalho propõe $S^3$ (Stratified Scaling Search), um método de busca guiado por verificador para melhorar a qualidade de geração em modelos de linguagem de difusão durante o tempo de inferência. Ele realoca a computação no processo de denoising, avaliando e reamostrando seletivamente candidatos promissores para favorecer saídas de maior qualidade.
This research explores Exponential Moving Average (EMA) traces as a minimal recurrent context to delineate the capabilities and limitations of fixed-coefficient accumulation in sequence models. It demonstrates that EMA traces excel at encoding temporal structure, matching advanced models on structural tasks, yet fundamentally fail to capture token identity, resulting in significantly reduced performance for language modeling.
This research investigates the training-time mechanisms of refusal in safety-aligned language models, specifically comparing supervised fine-tuning with R2D2-style dynamic adversarial fine-tuning. Findings show R2D2 initially achieves strong refusal on HarmBench but then partially reopens, while SFT remains consistently less robust.
CL-bench Life is a new human-curated benchmark designed to assess whether frontier language models can effectively learn from complex, messy real-life contexts. It comprises 405 context-task pairs to test models' ability to reason over personal and social experiences.
KMMMU is a new native Korean benchmark for evaluating multimodal understanding in Korean cultural and institutional settings, featuring 3,466 questions from native exams. The study shows that current AI models achieve only 42.05% accuracy on the full set, with significant failures in culturally and discipline-specific problems.
This paper proposes Math Takes Two, a new benchmark designed to assess the emergence of mathematical reasoning in language models through communication. It tests whether two agents, without prior mathematical knowledge, can develop a shared symbolic protocol to solve a visually grounded task where a numerical system facilitates extrapolation.
Este artigo de pesquisa propõe uma nova abordagem para otimização de documentos, transformando-os para melhor alinhamento com sistemas de recuperação via Reinforcement Learning (GRPO), utilizando melhorias de ranking como recompensa. O método, aplicável a retrievers de caixa preta, demonstrou ganhos em tarefas de recuperação de código e documentos visuais.
This paper investigates multi-step rewriting attacks on diffusion language model watermarks, which are used to verify AI text authorship. The findings show that watermarked texts can have their detection compromised after multiple rewrites by other language models, even those unaware of the watermark key.
FlowLM introduces a novel flow matching language model, adapted from pre-trained diffusion models through efficient fine-tuning. This method enables high-quality, few-step text generation that significantly outperforms traditional diffusion sampling with fewer training epochs.
This paper introduces TraFL, a novel post-training approach for diffusion language models that addresses "trajectory locking" observed in reward-maximizing methods. TraFL, a trajectory-balance objective, outperforms other methods across mathematical reasoning and code generation benchmarks.
Agent Island is a new multiagent simulation environment for language models, serving as a dynamic benchmark designed to mitigate saturation and contamination. Models like openai/gpt-5.5 are ranked based on their performance in games involving cooperation, conflict, and persuasion.
Grokers is an innovative architecture for building persistent, structured comprehension of typed knowledge graphs through bottom-up inductive traversal. Unlike RAG, it shifts intelligence to write time, where autonomous Groker agents analyze and enrich attributes via language models for all future queries at zero cost.
This paper explores training language models to forecast the empirical success of research ideas by evaluating pairs of ideas against objective outcomes. SFT significantly boosts performance beyond GPT-5, and RLVR can train models to discover interpretable reasoning paths for this forecasting task.
CoCoDA proposes a framework for tool-augmented language models, utilizing a co-evolving compositional code DAG to manage and retrieve tools efficiently. This approach addresses challenges in scaling tool libraries by encoding typed, compositional structures and pruning candidates through symbolic signature unification.
This research introduces OP-Mix, a novel algorithm for efficient data mixing throughout the entire lifecycle of language model training. It addresses the challenge of combining diverse data sources for pretraining, continual learning, and adaptation, proposing a unified online decision-making solution.
DisaBench introduces a participatory evaluation framework to assess disability-related harms in large language models, addressing the inadequacy of general-purpose safety benchmarks. It features a co-created taxonomy of twelve harm categories, a methodology pairing benign and adversarial prompts, and a dataset with human-annotated labels, revealing subtle harms often missed by standard evaluations.
Hebatron is a Hebrew-specialized open-weight large language model built on NVIDIA's Nemotron-3 Mixture-of-Experts (MoE) architecture. It achieves a 73.8% Hebrew reasoning average, outperforming competitors and offering significantly higher inference throughput by activating fewer parameters per pass.
The Consilium Protocol, derived from Byzantine Fault Tolerance, is introduced for structured multi-model AI deliberation, treating inter-model disagreement as an epistemic signal. The study demonstrates that cognitive personas determine epistemic behavior and that RLHF alignment training creates measurable epistemic blind spots.