lean4agent brings formal verification to agent workflows
a new framework uses the lean4 proof assistant to model and verify multi-step agent behavior, catching errors before execution.
topic
research
automated rubrics for evaluating llms in education
elmes* builds fine-grained rubrics to assess how large language models teach, not just what they know, across 330 long-tail educational scenarios.
bees and chatgpt: the consciousness question
new research shifts focus from behavior to internal mechanisms when assessing consciousness in animals and ai, finding current ai likely not conscious but leaving door open for insects and future machines.
gated inference-time context optimization for time series models
gitco improves zero-shot forecast accuracy by filtering harmful patches from input context without retraining.
micropython sandbox in webassembly
a new python package runs micropython inside a wasm sandbox for safe code execution with memory and cpu limits.
shipping a multi-agent economy on a 3b model
a field report on building a tiny woodland economy with qwen2.5-3b agents, showing how small models can drive emergent market behavior when paired with designed scarcity and sharp prompting.
pycc.id brings hypothesis-driven equation discovery with identifiability checks
a new python package helps scientists find governing equations from data by using structural skeletons and checking if parameters can be uniquely determined.
gemini agentic rag adds context checks for enterprise queries
google's new agentic rag framework uses a sufficient context agent to iteratively search across data sources, improving accuracy on complex multi-hop questions.
calibrating language model confidence with three methods
a look at temperature scaling, platt scaling, and isotonic regression for fixing overconfident llms.
deterministic envelopes fix bias in tamed sgmcmc
new framework decouples taming denominator from stochastic gradient noise to eliminate stationary bias in langevin algorithms.
new clustering index uses description length for validation
a new clustering validation index called central description length uses probabilistic bounds on description length to evaluate clusters without labels, handling non-convex and irregular shapes better than traditional methods.
staged factorial screening for budget-constrained micro-pretraining
a study tests whether staged fractional-factorial experiments can identify stable early effects in micro-pretraining under tight compute budgets.
llm benchmarks miss huge capability gaps
a stereological theory shows that standard llm benchmarks have a large structural blind spot, making rankings unreliable.
hello robot puts stretch 4 in real homes
hello robot ships stretch 4, a safe home robot for research and disability assistance, focusing on real-world data collection over lab demos.
eva-bench grows to three enterprise voice agent domains
eva-bench data 2.0 adds itsm and healthcare hrsd domains, totaling 213 scenarios across 121 tools to test voice agents on realistic, domain-specific tasks.
ai emotional dependence grows from routine use, not just chatbots
new research finds emotional reliance on ai often starts incidentally during everyday tasks, not through dedicated companion apps, and can shift preferences away from human support.
pre-deployment ai agent verification with ontology and trust certificates
a framework uses ontologies to generate test scenarios and issue verifiable trust certificates for enterprise ai agents before deployment.
finite-iteration theory for tensor power iteration
new analysis of alternating power iteration for spiked tensor models gives finite-iteration error bounds and explains warm-start behavior without relying on specific initializations.