Anthropic has updated its specialized LLM lineup with version 5.2. Claude Fable 5.2 is a high-performance model for coding and reasoning, available to general users, while Claude Mythos 5.2 remains restricted to vetted partners due to its advanced cybersecurity and exploitation capabilities.
NVIDIA has released an open-source tool, NVIDIA Snapshot, for Kubernetes-based GPU workloads. By checkpointing fully initialized models and their states, the system allows inference replicas to resume in seconds rather than minutes, effectively eliminating repetitive initialization steps.
Researchers have proposed FFN-Perturbed Self-Distillation (FPD) to address the challenge of removing sensitive information from Multimodal Large Language Models (MLLMs). The approach minimizes utility degradation by using the original teacher distribution as a reference while perturbing specific FFN activations associated with private data.
Google has updated its open-source JAX training framework, MaxText, to support training LLMs at extreme sequence lengths on Cloud TPU hardware. By introducing four distinct context-parallelism strategies, the update enables efficient training of models with up to 14 million tokens, overcoming memory and compute bottlenecks.
Amazon has expanded the capabilities of the AgenticRetrieveStream API, allowing developers to specify the iteration level for knowledge base requests. You can now choose between zero, one, or multiple retrieval steps to balance speed, cost, and complexity per request.
Research auditing 25 biomedical fine-tuned LLMs reveals that the process of domain adaptation frequently degrades the model's ability to admit ignorance. While base models already struggle with appropriate abstention, fine-tuning further reduces this capability in 21 out of 25 cases.
Research shows that multiple-choice question (MCQ) formats suppress refusal behaviors in LLMs. Experiments on Llama-3.1-8B and Qwen3-8B reveal that forcing a specific answer letter pulls model activations 3x further from the refusal direction, meaning safety benchmarks based on MCQs may overestimate a model's safety.
Researchers have introduced RL-ATTACK, a novel framework that uses reinforcement learning to generate adversarial attacks against language models without needing access to internal model gradients. By learning from the target model's responses, this method achieves superior query efficiency and successfully bypasses safety filters by forcing unintended outputs.
NVIDIA has introduced new optimization tools in TensorRT LLM VisualGen for the FLUX 3 Image Klein model. By utilizing NVFP4 quantization, kernel fusion, sparse attention, and step caching, developers can achieve up to 3.46x faster text-to-image generation on NVIDIA Blackwell hardware.
NVIDIA has introduced five co-design pillars for its Vera Rubin NVL72 platform, aimed at maximizing throughput in agentic AI workloads. By combining new precision formats like NVFP4, wide expert parallelism, and the Dynamo distributed serving layer, the system treats memory as a unified, tiered resource across GPUs, host CPUs, and specialized storage.
Researchers have identified that LLMs suffer from a bias toward providing immediate, often incorrect answers in ambiguous situations instead of asking clarifying questions. They introduced BACE (Belief-Aware Critic with Epistemic tracking), a lightweight method that runs on top of a frozen language model to strategically decide when to request clarification.
Dharma-AI has introduced a 2-billion-parameter model specialized in extracting structured JSON fields from various documents. It supports both open extraction and schema-directed modes, consistently outperforming frontier LLMs on field-level F1 benchmarks while significantly reducing operational costs.
NVIDIA introduced a workflow for turning production agent failures into structured test cases using NeMo Compass and NeMo Eval Author. This process analyzes trace logs to identify recurring issues, generates targeted evaluations, and validates harness improvements.
A new method called FlagRot optimizes key-value (KV) cache quantization for long-context models. By selectively rotating key segments based on attention error rather than applying global rotation, it reduces accuracy loss in 2-bit or 4-bit caches.
Researchers introduced SLATE, a benchmark for evaluating long-horizon, large-scale API selection in enterprise software environments. Unlike current benchmarks, it forces models to navigate roughly 1,000 tools to complete complex tasks.
New research shows that reasoning history can be rewritten into compact state records before new tool results are introduced. This method effectively reduces future reasoning tokens and total agent spending by eliminating redundant thoughts without requiring model retraining.
Researchers have developed a method to calculate the theoretical upper bound for Multi-Draft Speculative Decoding (MDSD) efficiency by solving the dual of the optimal transport problem. The study reveals that current verification algorithms often fall short of this theoretical limit, and that sampling without replacement significantly improves acceptance rates compared to sampling with replacement.
Collinear AI has launched CWE-bench v2, a benchmarking suite featuring 100 tasks covering 91 common vulnerability types. The tool allows developers to programmatically evaluate how effectively their AI agents find and fix security weaknesses in open-source repositories.
Google has deployed a new suite of Cheminformatics Skills for its Antigravity platform, enabling AI agents to manage end-to-end drug discovery pipelines. By integrating the RDKit toolkit with LLM-based orchestration, the system can perform literature research, data analysis, and chemical synthesis planning in approximately one hour.
Research shows that identical persona instructions produce different behavioral outcomes in LLMs when switching from English to Korean. While buyers followed instructions similarly, sellers became significantly more accommodating and polite in Korean, leading to higher agreement rates and lower prices.
Researchers have introduced MEMENTO, a specialized agentic framework built on Qwen3-8B designed to manage long context by compacting history into concise notes. By utilizing fine-tuning on synthetic trajectories followed by reinforcement learning, it achieves high accuracy on book-level tasks while strictly limiting context usage to 20K tokens, even for inputs as large as 2.5 million tokens.
Amazon introduced a multi-agent system built on Bedrock AgentCore that automates engineering document evaluation and substitution. The architecture combines multimodal retrieval, deterministic Python-based tolerance checking, and hierarchical LLM orchestration.
Apple researchers have unveiled an online training infrastructure for mobile UI agents that utilizes a two-part architecture to split navigational logic from visual interpretation. By employing step-level feedback provided by an LLM-as-a-Judge and a novel value estimation technique known as GiGPPO, these agents demonstrate significantly better adaptability than static models trained on pre-recorded sequences.
Recent research shows that many attention heads in large language models are redundant and specifically contribute to verbatim memorisation of training data. By selectively pruning these heads at inference time, researchers achieved a 91.5% reduction in memorisation events while simultaneously improving performance on standard benchmarks.