NorBERTo is a new ModernBERT model trained on a 331 billion token Brazilian Portuguese corpus (Aurora-PT), designed for long-context support and efficient attention mechanisms. It achieves state-of-the-art results among evaluated encoder models on semantic similarity, textual entailment, and classification tasks using datasets like ASSIN 2 and PLUE.
This paper introduces SCALAR (Structured Critic--Actor Loop for AI Reasoning), an Actor--Critic--Judge pipeline applied to theoretical physics problems. It investigates how the interaction between researchers and AI agents affects results in physics reasoning tasks, demonstrating that multi-turn dialogue significantly improves over single-shot attempts.
This paper evaluates speculative decoding with EAGLE3 as an inference-time optimization for PayPal's Commerce Agent, powered by fine-tuned Nemotron models. The study demonstrates significant performance improvements, including 22-49% throughput increase and 18-33% latency reduction at zero additional hardware cost.
CoAuthorAI is a human-in-the-loop system designed for scientific book writing, tackling LLM challenges like inconsistency and unreliable citations. It combines retrieval-augmented generation, expert outlines, and automatic reference linking, demonstrated by a high satisfaction rate and a published book.
This research introduces the Tool-Augmented Markov Decision Process (TA-MDP) to formally model multimodal agentic decision-making, addressing theoretical gaps in reinforcement fine-tuning for Large Vision-Language Models (LVLMs). It specifically investigates how composite verifiable rewards affect GRPO convergence and why training on small datasets generalizes to out-of-distribution domains for agentic LVLMs.
Super Apriel, a 15B-parameter supernet, has been released, offering four trained mixer choices per decoder layer to enable multiple speed/quality presets from a single checkpoint. This allows for 2.9x to 10.7x decode throughput gains with 96% to 77% quality retention, and also facilitates speculative decoding without a separate draft model.
This comprehensive replication study evaluates the efficacy of DExperts, an inference-time mitigation technique, to reduce toxicity in Large Language Models. The research establishes baseline toxicity measurements, implements DExperts to mitigate explicit toxicity, and stress-tests the method against implicit hate speech.
Large language models struggle with complex long-context reasoning tasks despite supporting extensive inputs. ProxyCoT is a novel training framework designed to transfer reasoning capabilities from short proxy contexts to full long contexts, outperforming strong baselines.
FLUID is a new framework designed to efficiently adapt Autoregressive (AR) backbones to the diffusion paradigm for parallel text generation. It enables initialization from GPT-style models and introduces an entropy-driven mechanism called Elastic Horizons, achieving state-of-the-art performance with significantly reduced training costs.
This article highlights Alibaba's Qwen model family as the largest and most complete open-source AI offering in 2026, detailing the Qwen3 series and the advanced Qwen3.5. It emphasizes the flagship Qwen3-235B-A22B's competitive performance against Gemini 2.5 Pro and discusses Alibaba's broader AI strategy.
GPT-Rosalind, a new OpenAI tool based on GPT-4 and fine-tuned on scientific data, has been launched to accelerate life sciences research. It addresses the data bottleneck by optimizing hypothesis generation, literature analysis, and experimental design, with the potential to reduce drug development costs and timelines.
Dental-TriageBench introduces the first expert-annotated benchmark for multimodal reasoning in hierarchical dental triage, comprising 246 authentic, de-identified cases. The research highlights a substantial performance gap between 19 MLLMs and junior dentists, particularly in treatment-level triage tasks requiring multiple referral domains.
The text introduces the Qiushi Discovery Engine, an LLM-based agentic system for autonomous scientific discovery on a real optical platform. It demonstrates end-to-end discovery by combining nonlinear research phases, Meta-Trace memory, and a dual-layer architecture, successfully reproducing a published experiment.
This research explores how humans communicate with limited vocabularies, comparing their strategies to computational sampling algorithms powered by large language models. The study reveals that human language production under constraint often mirrors greedy sampling, although more skilled individuals exhibit non-greedy revision behaviors.
This research investigates the balance between fluency and faithfulness in literary translation, comparing human, Google Translate, and TranslateGemma performance across 106 novels in 16 source languages. It reveals a consistent negative correlation between fluency and faithfulness, particularly for human and Google Translate, and indicates that segment length significantly impacts automatic evaluation.
A large-scale study re-evaluates Retrieval-Augmented Generation (RAG) in medical question answering, finding only small and inconsistent improvements over no-retrieval baselines. It suggests that the choice of the backbone model is more critical than retrieval methods, and the main bottleneck is the model's ability to effectively use retrieved evidence.
This paper derives scaling rules for Gated Delta Networks to address the computational demands of training and scaling Large Language Models. Experiments validate that these configurations enable stable learning-rate transfer across various model widths, unlike standard parametrization.
This commentary introduces PEEL, a working scaffolding combining deterministic distant reading with LLM interpretation, grounded in Peircean semiotics and abductive reasoning. Applied to AI-generated condensations, PEEL reveals systematic distortions invisible without non-AI measurement, implying deterministic instruments must accompany AI tools to ensure fidelity and epistemic authority.
This guide shows developers how to slash AI API costs by up to 95%, advocating for cheaper alternatives like DeepSeek V4 Flash over GPT-4o. It emphasizes a 40x price difference for similar output quality, helping developers manage project budgets effectively.
Google's Gemini 3.5 Flash revolutionizes AI speed, offering instant, top-tier intelligence for coding and complex reasoning tasks. This new model sets a new standard for performance, outperforming previous versions and challenging rivals.