Researchers propose CARE (Contrastive Anchor-based Rubric Evolution), a method that uses anchor responses from frontier models to dynamically evolve rubrics during LLM post-training, addressing reward…
Hugging Face released gr.Workflow, a feature built into Gradio that turns AI pipelines into interactive drag-and-drop canvases where each node is runnable and every intermediate result is visible, whi…
A new study from arXiv (2608.20348v1) characterizes the clinical lost-in-the-middle (CLitM) effect in electronic health record (EHR) processing, where large language models (LLMs) retrieve information…
A new study from researchers including Shayan Shahrabi finds that bitsandbytes 4-bit (INT4/NF4) post-training quantization significantly amplifies proactive interference in large language models, redu…
Researchers released ArguLens, an open-source automated essay scoring system that decomposes scoring into three components: a discourse-move classifier, a LightGBM scorer, and a label-aware feedback g…
An experimental training-time context-selection prototype, SpiralCoreAttention, achieved a mean training-step speedup of 1.675× and reduced peak VRAM by 6.036 GB during QLoRA fine-tuning of Qwen2.5-7B…
Researchers found that safety alignment in Diffusion Large Language Models (DLLMs) is sparse and transferable, enabling attacks that increase attack success rates from 2.6% to 73.8% on LLaDA and from …
A study labeling 7,440 answers from four open-weight LLMs found that more than half of the hallucination labels were incorrect, and correcting them reordered the results. The author, Priyanshi Jain, r…
Researchers introduced AgentMemBench, a benchmark evaluating five long-term memory strategies for conversational AI agents across three datasets and 491 annotated question turns, finding that external…
A practical analysis of full fine-tuning, LoRA, QLoRA, and TinyLoRA shows that TinyLoRA improved mathematical reasoning in a frozen Qwen2.5-7B-Instruct model with only 13 trainable parameters under GR…
A new study from arXiv finds that linear readouts trained on the final-token hidden state of large language models can decode whether diagnostic evidence supports, challenges, or fails to address a ca…
Researchers propose Progress-conditioned Group Policy Optimization (ProGPO), a method that uses first-visit observation coverage to assign higher advantages to trajectories visiting new states when al…
Researchers propose Spectral-LSH, a training-free prompt compression method that uses Krylov subspace methods and SimHash to reduce long-prompt inference costs by grouping similar tokens into macro-to…
A developer built TrueSET, a dataset of 924 issue-diagnosis-fix examples with execution-verified proofs, and fine-tuned Qwen2.5-7B-Instruct on it, achieving a 40.0% fix rate on hard multi-file faults …
A developer compared Llamafile and vLLM for serving local LLMs on an RTX 4090, finding Llamafile faster for single-user interactive use and simpler to set up, while vLLM excels under concurrent load a…
Mariana Souza published a practical guide for fine-tuning Qwen2.5-7B-Instruct using QLoRA on custom instruction datasets, including cost estimates and a loss-masking sanity check. The tutorial covers …
Alibaba researchers unveiled SkillWeaver, a framework that uses an iterative feedback loop and execution graphs to intelligently route AI tools for complex multi-step tasks, achieving a 99.9% reductio…
Researchers released a benchmark for Arabic-Russian scientific translation, including a 27,000-sentence parallel corpus and fine-tuned multilingual models. The Qwen2.5-7B model achieved BLEU 23.15, ou…
Researchers introduced PCB-Bench, the first comprehensive benchmark for evaluating large language models on printed circuit board placement and routing tasks. The benchmark includes over 3,700 questio…
Researchers propose Strategy-Guided Policy Optimization (SGPO), a method that replaces trajectory-level imitation with reusable strategy distillation to improve LLM reasoning. SGPO extracts structured…