in-context rl under non-stationarity survey
a survey on in-context reinforcement learning methods that handle changing environments without test-time parameter updates.
topic
research
using background knowledge for faster causal discovery
a new framework integrates expert constraints directly into scalable causal discovery algorithms, improving efficiency and accuracy.
doomql: a doom-like game running entirely in sqlite
a python terminal game where sqlite handles movement, collision, enemies, combat, and pixel rendering using a recursive cte ray tracer.
deep learning improves causal mediation estimates
a two-stage method called unit uses deep representation learning to sharpen estimates of structural mediation parameters under a no essential heterogeneity assumption.
tsconet: two-stage copula model for uncertainty-aware spatio-temporal forecasting
a new model combines a cnn-lstm with a gaussian copula to jointly forecast multiple environmental variables while providing calibrated uncertainty estimates.
turing's ai assumption may be wrong
computer scientist peter j. denning argues that ai cannot capture tacit human knowledge like common sense and culture, making true human-level intelligence impossible.
format sensitivity index measures prompt wrapper impact on llm scores
a new benchmark metric reveals how small formatting changes in prompts can significantly alter model accuracy and parseability, challenging the reliability of current leaderboard rankings.
toulmin model makes ml diagnosis more transparent
researchers use argumentation theory to break down retinal disease predictions into explainable parts, helping doctors trust ai decisions.
message format effects in multi-hop agent relays depend on tier
study finds that structured message formats help or hurt llm agent relays depending on the model's capability tier, with strong models staying accurate and weak models degrading.
reward transport: steering molecular generation via noise alignment
a new method uses optimal transport coupling during flow matching training to align noise with molecular rewards, enabling controllable generation without extra models or gradients.
knowledge distillation works by sparsifying interactions in llms
a unified framework reveals that various knowledge distillation methods for large language models share a common mechanism: they force student models to use fewer interactions while zeroing out the rest.
brain decisions start earlier than thought
a new study finds decision-making activity in early sensory brain regions, challenging traditional models and offering insights for more efficient ai.
benchmark tests ai agents on long terminal tasks with partial credit
a new benchmark evaluates ai agents on 46 long-horizon terminal tasks using dense intermediate rewards to measure progress, not just final outcomes.
interval certifications for neural networks via lattice traversal
a new framework reduces adversarial robustness certification to a lattice traversal problem, enabling both sound and complete interval certifications for multilayer perceptrons.
lean formalization of vlasov equation as a strategy game
a mathematician directed an ai to formalize a mean-field derivation of the vlasov equation in lean 4, treating the process as a game with clear win conditions.
deep gaussian processes on directed acyclic graphs
a new framework extends deep gaussian processes to directed acyclic graphs, enabling better modeling of complex, partially observed processes with theoretical guarantees on information preservation.
generative patient twin guides sepsis treatment at inference time
a new framework uses a generative ehr model as a patient digital twin to optimize sepsis treatment through model predictive control, adapting to changing goals without retraining.
signed symmetric quantization for few-bit integers
signed symmetric quantization uses the extra negative value in signed integers to reduce clipping error without runtime overhead.