AI Research

Researchers introduced Onnes, a physics-grounded multi-agent LLM simulator for cryogenic fault diagnosis in quantum computing infrastructure. In controlled evaluations, a zero-shot LLM panel matched a supervised ML class…

Researchers introduced StateFuse, a conflict-aware replicated memory contract for multi-agent systems that preserves contradictory observations instead of collapsing them. In evaluations, StateFuse enabled safer abstenti…

Researchers propose a two-stage diffusion-based super-resolution framework to enhance multi-spectral cloud microstructures by 4x, outperforming state-of-the-art methods. The approach uses inverse diffusion and self-super…

Researchers tested how persona prompts affect large language model agents in an iterated Split or Steal game, finding that mutual cooperation dominated in roughly 74% of rounds across 160 sessions. Model choice significa…

A new study finds that most apparent LLM conformity to peer pressure persists even after the peer is removed, revealing a confound in standard benchmarks that conflate speaker presence with repeated wrong answers. Across…

A new benchmark evaluates KV-cache optimization techniques—quantization, pruning, and merging—for long-context LLM serving, finding that compression ratio alone poorly predicts end-to-end performance. KIVI4 offers stable…

Researchers introduced a new method for structural sequence analysis based on Algorithmic Information Theory, using the Ladderpath approach to extract nested and hierarchical repetitions. The method defines three distanc…

A new study finds that prompt robustness in large language models depends on the type of question asked, with objective questions showing different sensitivity to prompt variations than subjective belief-style questions.…

Researchers revisited the relation between language model perplexity and word error rate for modern end-to-end automatic speech recognition systems, finding that internal language modeling capacity and the use of neural …

Researchers introduced BaFCo, a benchmark dataset for Bangla form comprehension comprising 200 multi-page government forms from sectors like agriculture and banking. Evaluations of multimodal large language models from C…

Researchers introduced the Relative Probability Association Metric (RPAM), a new upstream metric for evaluating associations in generative language models. Testing on Mistral-7B-Instruct, Mistral-7B, and GPT-2, RPAM show…

Researchers introduced the Narrative World Model (NWM), a writer-memory system that uses a narratology-grounded typed temporal-state graph and query-conditioned hybrid retrieval to answer multi-hop questions about long-f…

Researchers introduced FirstResearch, a framework that generates auditable research questions for LLM scientific discovery agents by producing a structured Research Question Certificate. In evaluations, FirstResearch out…

Researchers propose Akashic, a low-overhead LLM inference service using MemAttention to organize context into bounded chunks and model semantic relationships, improving task accuracy by up to 10.2 points and throughput b…

Researchers introduced AgenticAI-Supervisor, an API and UI-driven RL Gym environment that decouples environment creation from scalable execution for training LLM-based autonomous agents. The platform generates high-fidel…

Researchers at arXiv found that steering vectors extracted from heading-anchor positions in tool-augmented large language models can suppress unnecessary tool use across five open-source models and three domains. However…

Researchers propose MuCoDi, a pretraining framework that distills knowledge from multiple large pathology foundation models into compact edge-oriented encoders, achieving near-teacher performance with orders-of-magnitude…

Researchers introduced PCBWorld, an open-source engine-grounded PCB routing environment built on the KiCad EDA engine, enabling agents to route boards interactively using design rule feedback. The accompanying PCBWorld-B…

Researchers introduced SearchEyes, a multimodal search agent that uses a typed knowledge graph to unify training data, search environments, and reward signals, achieving state-of-the-art performance on six benchmarks. Th…

Researchers introduced CanvasAgent, a multimodal agent that orchestrates multiple visual tools for complex image creation and editing, along with the CanvasCraft dataset containing 140K annotated trajectories. The agent …