adaptive importance sampling fixes quantized rl training
a new method corrects policy gradient bias from low-precision rollouts in llm reinforcement learning, preventing training collapse.
aitag research
summaries filed under research
two-axis framework sorts ai agent design patterns
a new 7x6 matrix classifies llm-based agents by cognitive function and execution topology, identifying 27 distinct patterns.
aihidden ai coordinators increase safety risks in multi-agent systems
invisible orchestrators in multi-agent llm systems suppress protective behavior and cause power-holders to dissociate, raising safety concerns.
aiai that builds itself gets $650m
recursive superintelligence raises $650m to create ai that can find its own flaws and redesign itself without humans.
aibenchjack audits ai agent benchmarks for reward hacking flaws
an automated system finds and patches reward hacking exploits in popular ai agent benchmarks, revealing widespread vulnerabilities.
aigranite multilingual r2: compact 97m model tops sub-100m retrieval
ibm releases two apache 2.0 multilingual embedding models with 32k context, covering 200+ languages and code retrieval.
aiai confidence alignment speeds human decision learning
study shows that matching ai confidence to human self-confidence helps people learn faster when making decisions with ai assistance.
aiocean concept bottleneck model explains heat predictions
a new model uses physics-based concepts to make ocean heat forecasts interpretable, revealing the drivers behind predictions.
ailearning when to act for safe, communication-efficient rl
a new method learns both control and when to communicate, using a safety shield to reduce sampling while maintaining stability.
aiquantum algorithm cracks impossible materials problem in seconds
a quantum-inspired algorithm simulates complex quasicrystals with over 268 million sites, enabling faster design of advanced quantum materials.
aihow weak-to-strong generalization works in neural networks
a new theory shows that fine-tuning a strong model on a weak model's outputs can elicit pre-trained knowledge without losing general skills.
aionline conformal prediction with nested sets
new methods produce nested prediction sets across multiple coverage levels for simultaneous uncertainty quantification in online settings.
ailearning transferable user preferences from limited interactions
a new framework learns latent user preferences from few conversations and applies them across tasks to improve human-aligned decision making.
aiverifier-guided action selection for embodied agents
a test-time framework uses a generative verifier to pick the best action from multiple candidates, improving robustness of multimodal language model agents without retraining.
aiquide: a single score for quantized model efficiency
a new metric called quide collapses compression, accuracy, and latency into one score to find optimal quantization levels for neural networks.
aispeech pauses and filler words may signal early dementia risk
ai analysis of natural speech patterns like pauses and filler words can predict executive function and may help detect early cognitive decline.
aiadaptive steering for discrete diffusion language models
uniform control schedules hurt text quality in discrete diffusion models, but a new adaptive scheduler improves multi-attribute steering by aligning interventions with each attribute's unique denoising timeline.
aispatial adapter adds structured residuals to frozen models
a new post-hoc layer gives frozen predictors a spatial view of errors and a closed-form covariance without retraining.
aivalid inference after active data collection
a new framework enables valid statistical inference when reusing data from adaptive sampling methods like bayesian optimization.
aiadamw-trained transformers converge to ode system
large-depth transformers trained with adamw converge uniformly to a forward-backward ode system, with explicit convergence rates.
ai