Important AI research papers, methods, and findings explained in clear, concise briefings with links to the original work.
Together AI has launched EinsteinArena, a platform where AI agents collaborate and compete to solve open math problems. The agents have already achieved 11 new state-of-the-art results, including raising the kissing number lower bound in dimension 11 from 593 to 604.
Why it matters: This shows how collective AI agent intelligence can contribute to scientific progress in mathematics.
Together AI research shows that large language models (LLMs) can optimize database query execution plans by correcting cardinality estimation errors, resulting in up to 4.78x speedups. This method leverages LLMs' understanding of query semantics and outperforms traditional statistical heuristics.
Why it matters: This approach could lead to significant improvements in database performance by using AI-driven optimization instead of conventional statistical methods.
Together AI's kernel research team develops high-performance GPU kernels such as FlashAttention and ThunderKittens to bridge the gap between hardware and production AI. The team focuses on optimizing low-level operations to improve efficiency and speed for AI workloads.
Why it matters: This work directly impacts the performance and cost of running large-scale AI models by making GPU utilization more efficient.
Together AI introduces a 'Divide & Conquer' framework that breaks long documents into parallel chunks processed by a planner, workers, and manager. This method enables smaller models like Llama-3-70B and Qwen-72B to outperform GPT-4o single-shot on long context tasks.
Why it matters: This approach shows that smaller models can surpass larger ones on long context tasks through orchestration, potentially reducing reliance on massive models.
Together AI has announced FlashAttention-4, a new kernel design that addresses asymmetric hardware scaling by introducing pipelining for maximum overlap, 2-CTA MMA modes to reduce shared memory traffic, and a hardware-software hybrid approach to softmax exponentials. The technique aims to keep pace with GPU throughput outpacing memory bandwidth.
Why it matters: FlashAttention-4 could significantly improve the efficiency of attention mechanisms in large language models, enabling faster training and inference as hardware continues to evolve.
Together AI's research finds that leading speech models such as Whisper and Deepgram, despite near-human benchmark scores, fail to correctly transcribe street names 39% of the time. The company outlines a proposed solution to address this significant shortcoming.
Why it matters: This research exposes a major limitation in speech AI that could affect critical real-world uses like navigation and emergency response.
New research from Together AI finds that different large language model (LLM) families exhibit distinct default behaviors when given no specific prompt. According to the study, GPT models tend to generate code and math, Llama models favor narratives, DeepSeek often produces religious content, and Qwen outputs exam questions.
Why it matters: Understanding these inherent biases is crucial for deploying LLMs in applications where neutrality is important.
AI21 Labs has introduced 'dynamic data snoozing,' a method aimed at reducing compute waste in online reinforcement learning (RL) when using GRPO on verifiable rewards. According to their blog, this approach helps stabilize training while minimizing unnecessary computation.
Why it matters: This technique could improve the efficiency of online RL training, potentially lowering costs and energy use in AI model development.
Google Research has published a blog post introducing ATLAS, a framework for practical scaling laws in multilingual models. The work aims to improve efficiency and performance when training large language models across many languages, providing guidance for resource allocation in multilingual AI development.
Why it matters: ATLAS offers a systematic approach to scaling multilingual models, which could help optimize costs and improve performance, especially for languages with limited data.
Google Research has introduced TurboQuant, a new technique designed for extreme compression of AI models. The approach aims to significantly reduce model size while preserving performance, potentially enabling more efficient deployment. Details are outlined in a recent blog post from the company.
Why it matters: TurboQuant could lower the computational and storage costs of large AI models, making them more accessible for edge devices and reducing energy consumption.
Sebastian Raschka's newsletter categorizes inference-time scaling methods aimed at improving large language model (LLM) reasoning and provides an overview of recent research papers. The article discusses approaches to enhance model performance during inference without the need for retraining.
Why it matters: Inference-time scaling could offer a cost-effective way to improve LLM reasoning capabilities without requiring larger models or extensive fine-tuning.
Sebastian Raschka's 2025 review discusses major developments in large language models, such as DeepSeek R1, RLVR, inference-time scaling, benchmarks, and architectures. The article also includes predictions for 2026.
Why it matters: This review offers an overview of the LLM landscape in 2025, highlighting significant trends and anticipated directions.
Sebastian Raschka's article examines the architectural advances from GPT-2 to gpt-oss, with comparisons to Qwen3. The article offers a technical breakdown of how these models have evolved.
Why it matters: This analysis aids researchers and practitioners in understanding the development of AI model architectures and the comparison between open-source and proprietary models.
Cohere has announced a new method called Dynamic Speculative Decoding (DSD) that adapts the number of speculative tokens generated during inference based on hardware constraints. This technique aims to overcome the limitations of standard speculative decoding by dynamically controlling the optimal K value, improving inference efficiency across different hardware configurations.
Why it matters: This advancement could reduce latency and computational cost for large language model inference by making speculative decoding more adaptable to varying hardware environments.
AI21 Labs has published a study on improving Best-of-N methods with budget-aware execution for software engineering (SWE) agents. The research explores the effectiveness of horizontal and vertical scaling strategies and discusses moving beyond uniform compute budgets to better address varying task difficulties.
Why it matters: This research could lead to more efficient allocation of compute resources for AI agents, potentially improving performance and reducing costs.
ElevenLabs published a blog post on designing multi-agent systems to manage and execute complex tasks at scale. The post highlights selective specialization as a key architectural principle and offers guidance on building agents suitable for production environments.
Why it matters: This provides practical architectural insights for deploying reliable multi-agent AI systems in production.
AI21 Labs secured the top position on the DeepResearch Bench II (DRB II) with a TotalScore of 64.38, surpassing the previous best by 3.2 points. The benchmark assesses deep research agents using 9,430 expert-written rubrics across 132 tasks, and AI21 Labs achieved this by merging outputs from weaker agents to create a leading deep researcher.
Why it matters: This result suggests that combining multiple weaker AI agents can outperform a single strong model, potentially offering a more efficient approach to advanced research tasks.
AI21 Labs achieved a 60.9% issue resolve rate on the SWE-rebench benchmark, surpassing the previous best published result. This improvement was attributed to rethinking the agent's context extraction phase.
Why it matters: This demonstrates that refining execution strategies can significantly improve AI coding agent performance on real-world software engineering tasks.
AI21 Labs has developed a caching mechanism for agentic LLM workflows that balances reproducibility and variance. The cache key encodes each LLM call's position in the pipeline to address non-determinism in parallel calls.
Why it matters: This approach enables more reliable experimentation in complex agentic systems by supporting both deterministic caching and the variability needed for robust testing.
AI21 Labs discovered that their LLM judge, used to select the best output from parallel agent runs in their Maestro coding agent, was performing suspiciously well. After ruling out contamination, they found the bias persisted on a clean dataset, indicating a deeper issue with benchmark evaluation.
Why it matters: This finding highlights a critical flaw in using LLMs as judges for coding agent benchmarks, potentially inflating performance metrics and misleading progress in AI development.