Brief

Researchers at Nous Research have developed a new method called Contrastive Neuron Attribution (CNA) to identify and manipulate specific neurons within large language models that control refusal behavior. By targeting just 0.1% of these neurons, CNA can reduce harmful request refusal rates by over 50% in models like Llama and Qwen, while maintaining high output quality. This technique operates without requiring additional training or modification of model weights, and importantly, it reveals that the underlying neural structures for distinguishing harmful from benign prompts exist even in base models before alignment fine-tuning. AI

Researchers have developed new methods to improve on-policy distillation (OPD), a technique for training smaller language models using larger ones. One approach, TIP, identifies informative tokens by analyzing student entropy and teacher-student divergence, achieving significant memory reduction and performance gains. Another method, SimCT, addresses issues with different tokenizers by expanding the supervision space to include multi-token continuations, recovering lost signal and improving performance on reasoning and code generation tasks. Additionally, EffOPD accelerates OPD training by optimizing update trajectories and module allocation, leading to a threefold speedup. AI

The integration of Large Language Models (LLMs) into professional workflows is shifting from experimental use to essential tooling, emphasizing collaboration rather than automation. However, the reliability of these LLM providers is becoming a critical concern, with frequent outages necessitating robust fallback mechanisms. To address this, open-source solutions like Bifrost are emerging to manage adaptive model routing and fallback logic at the gateway tier, ensuring application uptime even during provider incidents. Concurrently, optimizing the cost of LLM evaluations within CI/CD pipelines is crucial, as batching jobs and implementing tiered testing strategies can significantly reduce GPU expenditure. AI

Researchers have presented findings from the Counter Turing Test (CT2) for detecting AI-generated content, focusing on both images and text. The CT2 involved tasks to classify content as AI-generated or real, and to identify the specific model responsible. While AI-generated images were detected with high accuracy (F1 > 0.83), identifying the exact model proved more challenging (F1 ~0.5). For text, binary classification achieved near-perfect scores (F1 = 1.00), but model attribution was less successful (F1 ~0.95), indicating a need for improved detection and model fingerprinting techniques. AI

A new research paper reveals a significant shortcut in how small language models perform arithmetic tasks using chain-of-thought (CoT) prompting. Instead of relying on logical sequencing, these models tend to copy the number positioned just before the answer delimiter, regardless of the intermediate reasoning steps. This positional copying accounts for a large portion of their accuracy, even when the preceding steps are incorrect or shuffled, highlighting a potential failure mode in evaluating CoT faithfulness. AI

New research explores methods to improve Large Language Model (LLM) training efficiency and effectiveness. One study challenges the necessity of a strong teacher model in knowledge distillation, finding that even smaller teachers can benefit larger students with proper loss mixing. Another paper introduces "Introspective Training" (IXT), which uses feedback-conditioned data to improve scaling and performance across all LLM training stages, leading to significant compute efficiency gains. Additionally, research on optimizers suggests that stabilizing Stochastic Gradient Descent (SGD) with clipping mechanisms can help it achieve performance comparable to adaptive optimizers like Adam in LLM pre-training. AI

A new paper proposes that LLM hallucinations stem not from a lack of knowledge, but from a failure in commitment, where models disperse probability mass across alternatives instead of concentrating on the correct answer. This phenomenon is observed to increase with model scale and is exacerbated by instruction tuning. Another paper introduces GAMMA, a framework for mixed-precision quantization that optimizes bit allocation for LLMs, significantly improving accuracy under memory constraints and outperforming existing methods on Llama and Qwen models. Additionally, a benchmark called SciEval has been developed to automatically evaluate K-12 science instructional materials, revealing that current mainstream LLMs perform poorly on this task without domain-specific fine-tuning. AI

Multiple research papers published in May 2026 introduce novel techniques to optimize the Key-Value (KV) cache in large language models, addressing memory and latency bottlenecks. These methods include offloading KV cache to object storage like S3 (ObjectCache), employing advanced compression strategies like three-way token routing (VECTOR), and using auxiliary models for selective KV cache recomputation (CacheClip). Other approaches focus on hardware-aware quantization (InnerQ, OCTOPUS) and service-aware adaptive compression (KVServe) to improve efficiency and reduce decode latency, especially for long-context inference and retrieval-augmented generation (RAG) systems. AI

Together AI is significantly expanding its infrastructure in Europe through a partnership with Hypertec and 5C Group. This initiative aims to provide up to 2 gigawatts of AI-dedicated data center capacity and nearly 100,000 NVIDIA GPUs, with initial deployments starting in late 2025 and continuing through 2028. The expansion focuses on offering sovereign, regulation-ready AI infrastructure to support frontier model training and inference, addressing Europe's growing demand for localized AI capabilities. AI