Language-model gains may come from diversity, not hidden experts
A preprint challenges a weight-space explanation for ensemble gains and tests a simpler output-level alternative.
- language models
- ensembles
- inference
CHANNEL ARCHIVE // [15] ARTICLES
Artificial Intelligence
Research into artificial intelligence and intelligent systems.
CHANNEL // 01 [15]
Artificial Intelligence
Research into artificial intelligence and intelligent systems.
SpecSlot separates global image structure from local detail
An arXiv preprint describes a pixel-space image generator that routes low-frequency information through global attention while handling other spectral detail locally.
- generative transformers
- pixel-space generation
- global attention
Researchers examine multi-camera strategies for luggage re-identification
An arXiv preprint studies how an end-to-end system might detect and match luggage across video streams, but its abstract does not report results.
- luggage re-identification
- computer vision
- multi-camera video
AI agent estimates crop losses from extreme weather in Florida
A preprint describes WxCrop, an AI agent that combines weather and agricultural information to estimate crop losses, with initial tests covering three hazards in Florida.
- artificial intelligence
- agriculture
- extreme weather
Visual AI Models Show Uneven Ability to Maintain Spatial State
An arXiv preprint finds that reading a visible scene from visual model representations does not guarantee that information persists, predicts future observations or supports useful action.
- visual foundation models
- spatial representation
- object persistence
TextSeal2 aims to preserve AI text watermarks during faster decoding
An arXiv preprint reports improved detection of watermarked language-model output, including under speculative decoding, while preserving output diversity.
- AI watermarking
- large language models
- speculative decoding
A compact AI model combines when to help with what to suggest
An arXiv preprint reports that a single compact decoder performed competitively on automated measures of proactive, memory-grounded conversational hints, though human ratings favored a larger model.
- proactive assistants
- language models
- memory-grounded generation
AI agents test quantum-spin predictions and flag a drifting signal
In an arXiv preprint, researchers report that collaborating AI agents derived and tested predictions about quantum spins in diamond, revising an experiment after detecting signal drift.
- artificial intelligence
- autonomous research
- quantum spins
A probe flags impending AI-agent leaks, but struggles to identify violations
In a controlled multi-agent study, a probe detected many imminent secret disclosures before output, but also flagged permitted disclosures and generated frequent episode-level alarms.
- AI agents
- multi-agent systems
- activation probes
Deep learning targets clutter in shared wireless sensing
An arXiv preprint reports that neural-network clutter suppression improved position and velocity estimates in a modeled multistatic sensing framework, though the abstract provides no numerical results or real-world validation details.
- integrated sensing and communications
- clutter suppression
- deep learning
VeriCUA explores digital environments to check whether agent tasks are complete
An arXiv preprint describes a verifier that gathers evidence through interface and code checks, reporting better agreement with reference labels than compared baselines.
- computer-using agents
- verification
- agentic exploration
AI coding agents build a robotics simulation benchmark
An arXiv preprint introduces agent-built robotics simulations and reports that leading models failed all test tasks, though several learning approaches improved performance.
- robotics
- simulation
- coding agents
Study compares human and AI-simulated clients in counselling role-plays
A preprint reports similarities and differences in counselling communication patterns across role-plays with human peers and AI-simulated clients, but its abstract does not detail the results.
- counselling training
- large language models
- role-play
FARMS links formal mathematics to computational skills
In an arXiv preprint, researchers describe a pipeline for making computational mathematics reusable by AI systems, with selected implications checked in Lean.
- automated computational science
- formal verification
- Lean
AI agent combines executable skills for clinical genomic analysis
An arXiv preprint reports that ClinGeX matched expert classifications more often than comparators in a benchmark and was associated with higher concordance in a professional collaboration study.
- clinical genomics
- large language models
- AI agents