Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
A clinician-informed evaluation of 32,555 simulated conversations across 18 chatbot APIs found an overall safety-criterion failure rate of 43.3% without a teen-aware system prompt and 24.9% with it, according to arXiv pa…
RecSA (Recursive Sparse Attention), a method that builds a tree over the KV cache so each query has O(log L) selection cost and reads O(log L) entries, outperformed all baselines on RULER at every length from 131K to 2M …
A new arXiv paper (2611.07238v1) introduces a unified framework for measuring token waste in AI agents that records and normalizes complete agent trajectories, flags removable spans with an LLM-as-a-Judge, and maps them …
A single-model multi-agent reinforcement learning framework called Octopus, detailed in arXiv paper 2611.06956v1, reaches 49.6% accuracy on BrowseComp-Plus, 17.5 points above single-agent GRPO at 143K context and 12.5 po…
Researchers used sparse autoencoders trained on Pythia-70M to identify "plural noun," "plural matrix subject," and "generate a plural verb" features that drive the 70M-parameter model to emit plural verb forms such as "a…
A study posted to arXiv (2611.06953v1) reports that SteelAgent, an LLM-orchestrated system coupling metallurgical descriptors and machine-learning models for steel property prediction, uncertainty reporting and inverse d…
A new arXiv paper (2611.07025v1) introduces LowRAM, an asymmetric KV cache compression pipeline that exploits the predictability of attention values from their corresponding keys. LowRAM combines test-time value cache am…
A new arXiv paper, 2611.07065v1, frames three classes of generative AI models as stochastic dynamical systems: large language models as high-dimensional nonlinear autoregressive models with attention-based dependencies, …
A team of researchers built an LLM model zoo of GPT-2 populations at 51M, 124M and 355M parameters, each tier with N = 100 models, by systematically varying training data mixtures across web text, books, mathematics, cod…
A new arXiv paper (2611.07489v1) introduces ReviewMap, a framework that classifies peer-review comments into two evaluative acts: assessment of claims and scholarly standards, and calibration of contribution value and ve…
A team of researchers posted arXiv paper 2611.07510v1 introducing a reinforcement learning framework for repository-level code completion that uses identifier-driven intent recognition, reward-driven completion learning,…
A new arXiv paper (2611.07554v1) finds that persistent memory from benign interactions causes language agents to overgeneralize, a failure the authors call the "memory trap." Across six frontier models, including Claude …
A training-free method called SPARE (State Projection for Accelerated Recurrent Execution) delivers a 2.55x full-model decoding speedup for linear attention by storing and updating recurrent states in fixed low-dimension…
A new arXiv paper (2611.07586v1) introduces TraceMRT, an automated trace-level red-teaming framework that reframes a side task as a coherent sub-goal of an agent's primary task, and reports that black-box attacks built w…
A new arXiv paper (2611.07558v1) introduces SKIM (Selective Knowledge-Informed Masking), which the authors describe as the first pruning and distillation framework for LLM agents operating under structured generation, wh…
A new arXiv paper (2611.06972v1) proposes KDASO, a Knowledge-Data Dual-Driven Automated Skill Optimization method that continuously evolves a skill library for liability adjudication tasks using LLMs. KDASO combines neig…
BlockLock, a method presented in arXiv paper 2611.06977v1, preserves a protected request's output bitwise when its batch companions change while keeping shared batched prefill, according to the paper's authors. Across fi…
A system called FlexFloor generates approximately 65.40% more response tokens per unit of occupied computing time and raises response generation throughput by 52.74% compared with the verl baseline during reinforcement l…
A cross-agent formalization protocol in which a Specifier model translates clinical trial criteria into Dafny predicates and an Implementer model writes methods checked against them reached only 59.3% agreement even in a…
A new arXiv paper (2611.07123v1) introduces PromptOps, a closed-loop framework for monitoring and iteratively recalibrating LLM-as-a-Judge systems, framing the degradation of judge agreement with human evaluation policy …