algometrics: forecasting under algorithmic feedback
a new framework for time series where predictions influence future data, showing historical risk can mislead deployment risk.
topic
research
replicating picbreeder with vision-language models
researchers test if large vision-language models can replicate human-driven open-ended image evolution from the picbreeder system.
most llm reasoning steps are unnecessary, study finds
a large-scale study shows that 61% to 93% of reasoning steps in frontier models can be removed without changing the final answer, revealing massive redundancy in chain-of-thought.
finding metastable basins in high dimensions via trajectory discrimination
a new method identifies distinct dynamical basins in high-dimensional markov processes by comparing marginal trajectory distributions, avoiding spatial discretization.
context: proactive agents via sandboxed programs
context replaces reactive chatbots with proactive agents that advance tasks without user prompts using precomputed context, sandboxed programs, and goal-driven state machines.
optimizing latency, reliability, and cost in llm agent workflows
a study models tradeoffs in llm-based agent workflows and proposes a water-filling token allocation policy to balance speed, accuracy, and expense.
pope leo xiv's ai encyclical highlights human dignity
the vatican's new encyclical on artificial intelligence offers clear ethical guidance on interpretability, bias, accountability, and environmental impact.
latent cache flow sends model thoughts without text
a new method lets ai models share compressed internal states directly, skipping slow text generation and handling different contexts.
sciatlas: a 157m-entity knowledge graph for science
sciatlas builds a massive knowledge graph from 43 million papers to help ai navigate scientific literature with structured reasoning.
neurosymbolic framework translates natural language to temporal logic
a new architecture combines neural translation with formal verification to produce correct linear temporal logic from natural language.
fusionsense: adaptive near-sensor learning for multimodal edge ai
a new framework trains lightweight near-sensor classifiers to decide what data to transmit, reducing energy and latency in multimodal edge systems.
agentic system tackles research-level math proofs
rma uses specialized agents to solve open math problems by searching literature, building knowledge, and iteratively refining proofs.
why centerloss hurts ood detection and multi-scale mahalanobis wins
a new method called goen uses multi-scale features and mahalanobis distance to beat deep ensembles on out-of-distribution detection, while revealing that centerloss harms performance.
osctom uses rl to test nested belief conflicts in llms
a new method generates complex theory of mind scenarios where an observer's view of another agent clashes with their own belief, challenging llms with recursive reasoning.
tool-augmented agent for closed-loop cad-cae optimization
a reinforcement learning framework teaches llms to orchestrate cad generation, cae solving, and geometry revision until constraints are met.
specialization beats scale in ai procurement
a 3b specialized model outperformed frontier apis at lower cost, showing training alignment matters more than parameter count.
text degeneration costs more than you think
a small fraction of autoregressive model requests can loop into repetition, inflating batch inference time by over 40% and exposing a blind spot in standard benchmarks.
adaptive rbf-kan adds dynamic shape parameters to kolmogorov-arnold networks
a new study extends fastkan by using leave-one-out cross-validation to set kernel shape parameters and adding matern and wendland kernels for more flexible function approximation.