Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
CARE-PPO combines PPO fine-tuning with loss prediction to jointly train LLMs for accurate numerical estimates and calibrated confidence signals in quantitative prediction tasks.
Establishes rigorous first-principles framework for identifying extractable memorization in LLMs using probability-based matching to distinguish training sequence reproduction from predictability.
Survey of metacognitive abilities in LLMs covering self-awareness, error detection, and confidence calibration as paths to more reliable AI systems.
E-VQA task requires Video LLMs to ground answers with spatio-temporal evidence including object tracking and segmentation masks.
Multilingual moral reasoning framework grounding LLMs in theory and culture-specific adaptation, with culture-aware benchmark and training methods.
Study compares production vs. perception asymmetry in Llama-3.1-8B via token probability analysis, finding LLMs lack functional production-perception distinction.
Hourglass reasoning enforces structural isolation between induction stages in LLMs, using compressed symbolic state to improve few-shot reasoning.
LLMs parameterize cumulative prospect theory for modeling human route-choice biases at scale.
LLMs applied to EDA front-end design via HDL generation and agentic AI frameworks like OpenClaw for chip design automation.
Multimodal RL reward hacking in MLLMs via visual misalignment; introduces NRFR metric to measure failures in improved-reward samples across VQA and chart tasks.
Test-time training method using guided self-learning to improve long-context LLM performance on extended input sequences.
Fine-tuning LLMs for financial natural language processing (NLP) is constrained by limited, imbalanced data. Real-world financial news overrepresents earnings... Fine-tuning LLMs for financial natural language processing (NLP) is constrained by limited, imbalanced data…
Shows quantization from 8-bit to 2-bit induces behavioral divergence in LLMs beyond accuracy/perplexity metrics; introduces correctness agreement metric.
Challenges Super Weights hypothesis by showing isolated training of critical parameters fails on OLMo models; selective parameter training ineffective.
Evaluates AMALIA (9B Portuguese model) on moral foundation coding; distinguishes agreement from validity for theoretically-grounded construct annotation.
Study calibrates LLM judges for citation verification in deep-research systems, testing whether frontier models are necessary for rubric-based reward signals.
DocMaster preserves hierarchical structure (sections, tables, figures) in document analysis using LLMs, improving performance over flat-text baselines.
Analysis of Bielik LLMs (1.5B-11B) showing activation dispersion in MLP layers predicts entity familiarity and factual reliability across domains.
DiaLLM studies English dialect generation in open-weight LLMs via continual pretraining on ICE, comparing alignment strategies across Australian, Indian, British English.
Android Bench is evolving, and developers can help guide that process.
"HalluSquatting" weaponizes LLMs' inability to say "I don't know."
First large-scale evaluation of LLM uncertainty estimation across 22 languages and model sizes, comparing nine methods on multilingual MCQA.
Paper defines prompting complexity as shortest deterministic prompt for target text, adapting Kolmogorov complexity to LLM-relative setting.
LLM4SDM evaluates open-source smaller models on clinical decision-making assessment, comparing privacy-preserving local deployment vs. commercial models.
Training LLMs at massive scale brings unique infrastructure challenges, especially as jobs span thousands of GPUs and run for extended periods. The longer these... Training LLMs at massive scale brings unique infrastructure challenges, especially as jobs span thousands of GPUs…
HeRo enables selective disclosure watermarking in LLMs, allowing verification of watermark subsets without exposing full embedded metadata.
Linear probes decode remaining output length from LLM hidden states across 7-8B open-weight models, revealing internal response-length estimation.
In the AI era, platforms have no choice but to fight fire with fire to cull spam.
MentalThink equips multimodal LLMs with SVG-based visual reasoning for structured spatial problem-solving.
Selective prediction framework enables LLMs to abstain on uncertain inputs, balancing risk-coverage tradeoff.