Researchers introduced JudgeArena, an open-source framework unifying major LLM-judge benchmarks (AlpacaEval, Arena-Hard, MT-Bench, and m-Arena-Hard) under a single interface with swappable judges and metadata logging. Th…
A new arXiv preprint (2608.02829) finds that converting a pretrained 1.4B-parameter Pythia model into a 410M-parameter sibling works best through initialization transfer, not weight projection. The authors show that dens…
A study from arXiv (2608.02621v1) found that four large language models (LLMs) spontaneously cited legal authority in 238 Taiwan bar-examination items even when not prompted, but answer correctness and authority groundin…
A new arXiv preprint (arXiv:2608.02625v1) introduces draft-then-refine decoding for diffusion language models, showing that LLaDA2.1-Flash improves GSM8K-384 accuracy from 0.848 to 0.899 and MBPP-384 from 0.545 to 0.693 …
Researchers introduced Group-Relative Preference Backpropagation (GRPB), a method that converts same-task rankings into zero-sum advantages to train long-horizon video editing agents, and released the resulting 9B Crayot…
Researchers developed a multimodal auto-regressive transformer surrogate to model variable well operations and quantify uncertainty in geological carbon storage, achieving a median saturation MAE of 0.028 and median rela…
Researchers introduced GeoID-PINN, a physics-informed neural network for regional epidemic inference, which in a four-region simulation achieved source-composition error 0.099 with a compatible distance prior, versus 0.5…
Researchers introduced MemArena, an ego-centric benchmark for on-device agentic personal memory assistants, built with the MASim agent simulator, simulating 50 agents over 15 days (10.3M dialog-text tokens, 24.1K text-on…
A new independent evaluation of OpenAI's Privacy Filter (OPF), a 1.5B-parameter PII detector, across 42 benchmarks in 22 languages and 5 domains finds that OPF outperforms Presidio and XLM-RoBERTa on PII-annotated benchm…
A new arXiv paper (2608.02975v1) introduces TQLite, a distillation framework that uses a multi-LRM jury to train small language models (SLMs) for MQM-based translation quality evaluation, achieving performance far exceed…
Researchers introduced CT-HEG, a bidirectional, timestamp-attributed event graph for ICU in-hospital mortality prediction, and their CHIRP-Net model achieved a 5-seed mean AUROC of 0.8449±0.0071 on MIMIC-IV v3.1 (31,142 …
Researchers introduced OncoTriad-QA, a patient-level radiology-pathology-genomics benchmark for pan-cancer question answering, containing 86.1k semantic questions across 9,281 TCGA patient cases from 32 cancer cohorts. T…
Researchers propose GLOBE (Gradient Local-Balanced Extraction), a trajectory-aligned coreset selection framework that formulates sample selection as a globally optimized sparse weighting problem, using gradient trajector…
Researchers introduced Neural Networks with Local Converging Inputs (NNLCI), a method that uses a neural network to correct numerical solutions for pricing multi-asset options, reducing root-mean-square error by a factor…
A new arXiv paper introduces 'evaluation blindness,' a condition where AI measurement functions produce healthy readings while systems fail silently, and finds that 53% of 50 verifiable public AI failures were silent. Th…
Researchers propose PAMT, a process-aligned reinforcement learning framework for multi-domain machine translation, which improves translation quality by using step-level process rewards to identify helpful intermediate t…
Researchers introduced OptR, an output-aware rotation method for INT2 KV-cache quantization that minimizes post-W_O attention-output error, improving long-context LLM inference efficiency. Across three models and five re…
Researchers propose NANQ, a noise-aware mixed-precision non-uniform quantization framework for analog compute-in-memory (CIM), which models magnitude-dependent weight noise from measured eFlash CIM arrays to assign finer…
Researchers introduced FLARE (Few-shot Learning-based Adaptive Reflective Engine), a framework that uses reflective mechanisms and few-shot reference examples to optimize prompts for large language models, outperforming …
Researchers introduced Local Credit Assignment (LoCA), a two-stage method for parameter-efficient fine-tuning of large language models that requires only one backward pass for calibration, after which tuning proceeds for…