AI Research

A new study from Japan's AI Security Institute finds that LLM-as-judge safety evaluations are not reproducible even at temperature 0, with per-item disagreement up to 50% across runs. The researchers tested 690 API calls…

Researchers introduced HierBias, a hierarchical context-conditioned media bias detector that models document context to improve sentence-level bias classification. The system achieved 0.853 F1 and 0.723 MCC on BABE and B…

Researchers introduced Know2Guess, a contamination-aware multi-zone benchmark with 1,200 items across five domains to evaluate large language models' ability to distinguish answerable knowledge from abstention-expected u…

A new study on Llama 3.1 8B finds that helpfulness post-training (SFT and GRPO) significantly degrades animal compassion values compared to coding-domain post-training, with a 35.7% vs. 65.2% gap on the Animal Harm Bench…

Researchers released ProfileFoundry, a synthetic dataset of 100,000 person objects with 709,228 events, 40,338 households, and 52,491 employers, designed for evaluating LLM agents on memory, privacy, and tool-use tasks. …

Researchers introduced NEST-V1, a multimodal framework that generates emotion-conditioned Nepali Sign Language avatars from spoken words, achieving 81.1% ASR accuracy and 79.21% emotion recognition accuracy on a dataset …

Researchers developed a methodology to transform Hindi WordNet into 1.25 million instruction-response pairs, fine-tuning a 12B-parameter language model to create a specialized conversational AI for low-resource languages…

Researchers present the first systematic survey of unified vision-language perception in Multimodal Large Language Models (MLLMs), tracing the paradigm evolution through a five-stage taxonomy and identifying open challen…

Researchers propose AnySimLite, a lightweight similarity encoder for on-device speech-adjacent classification that achieves state-of-the-art or competitive performance in few-shot settings while using less than 1/250th t…

Researchers propose SOLAR, an auxiliary objective for supervised fine-tuning that aligns soft-token representations across languages using English as a pivot, improving cross-lingual reasoning in multilingual large langu…

Researchers propose DocArena, a fully automated pipeline that transforms raw document collections into controllable training environments for document search agents, using multimodal large language models to structure do…

Researchers propose VIGIL, a reinforcement-learning post-training framework that reduces hallucinations in multimodal large language models by maximizing mutual information between visual input and generated responses. T…

Researchers developed a multimodal deep learning model for methane-plume segmentation from hyperspectral satellite imagery, integrating a feature-guided methane enhancement mechanism into transformer-based RGB representa…

Researchers propose ForeAgent, an agentic forensics framework for AI-generated image detection that uses a Perception-Verdict architecture and a Hindsight-Driven Self-Refining strategy to iteratively improve. The system …

Researchers propose a lattice-theoretic method for AI oracles to report unbiased, self-consistent probability sets, using the Knaster-Tarski fixed-point theorem to select a canonical credal set that avoids the self-refer…

Researchers developed topology-informed neural networks that integrate topological data analysis with deep learning to improve flood detection in satellite imagery, achieving more robust and interpretable results on the …

Researchers introduced Chisao, a GPU-native parallel optimizer for multimodal black-box functions that uses a convergence-anticonvergence oscillation cycle to escape local traps while freezing confirmed modes. On the Sim…

Researchers introduced Hankel Reduced order Model (HRM) adapters, a state space model-based fine-tuning method that outperforms LoRA on long-context tasks like QuALITY and QMSum by leveraging FFT-based parallel scan for …

Researchers introduced KG-TRACE, a neuro-symbolic framework that integrates the WHO mutation knowledge graph as a biological constraint on neural genomic models for antimicrobial resistance prediction. Evaluated on the C…

A new thesis on algorithmic fairness argues that current fairness paradigms rely on deterministic point estimates and treat individuals as isolated entities, limiting their effectiveness in complex socio-technical enviro…