dynaschedbench calibrates dynamic scheduling benchmarks for llm agents
a new framework controls instance difficulty to test llm-based scheduling agents, revealing an observability paradox where more information can hurt performance.
topic
research
rsi is the new agi — and just as hard to pin down
recursive self-improvement has become the latest buzzword in ai, with startups and researchers chasing systems that can upgrade themselves without human help.
synthetic information origins via steganographic inheritance
a new paper proposes steganography as a mechanism for tracing the lineage of synthetic information, addressing the challenge of ai-generated content that diverges from its sources.
soro: a lightweight tajik language model and chatbot
soro is a family of tajik-specialized conversational language models built for low-resource deployment, outperforming same-size baselines on new tajik benchmarks.
triangular-reference bridges improve time series generation
a new method extends schrödinger bridge models for time series by using a frozen, triangular reference process across latent volatility levels, preserving the h-transform structure even with degenerate covariance.
calibrated uncertainty for treatment effects with small control groups
bayesian x-learner intervals under-cover in few-placebo settings due to nuisance model bias, but a gaussian process approach can fix calibration.
causal discovery with per-edge impossibility certificates
a new protocol for observational causal discovery attaches impossibility certificates to each edge, distinguishing data-driven orientations from those needing expert input.
llm architecture detects human values in text
a new llm-based system identifies and measures human values in text without being tied to a single value theory.
how mixture-of-experts tackles multimodal learning challenges
a survey examines how mixture-of-experts methods address key multimodal learning issues like scalability, representation, and fusion.
why llms fail at causal discovery
a new theorem shows that large language models cannot learn causal structure from observational data alone, but an agentic approach using interventions can succeed.
aircast-sr: km-scale atmospheric super-resolution model
a foundation model downscales global ai weather forecasts from 28 km to 1 km resolution, producing hourly 67-hour forecasts of eight surface variables.
google combines tee and crypto for zero-trust private analytics
google research introduces a private analytics solution that merges a single-shot cryptographic aggregation protocol with tee attestation to reduce trust in any single entity.
small models pay a constraint tax for structured output
forcing small language models to produce valid json or schemas can hurt answer quality, a tradeoff measured as constraint tax.
cart forests as sequential allocation over random opportunity sets
a new stochastic-control theory explains how cart random forests work by viewing feature subsampling as random opportunity sets and split rules as allocation policies.
neural bayesian sequential routing
a framework that models neural inference as active evidence accumulation over a hierarchical dag, enabling uncertainty-aware routing and early stopping.
silhouette scoring boosts isolation forest for fraud detection
adding a silhouette-based scoring layer to isolation forest improves unsupervised fraud detection on a large benchmark dataset.
gem: geometric entropy mixing for llm data curation
a new framework uses hypersphere geometry and entropy to find balanced data mixtures for training large language models.
curl team faces flood of ai-assisted security reports
ai tools are generating a surge in detailed, credible security reports for curl, overwhelming the team and straining work-life balance.