a daily read of the ML and AI papers
One prompt line — hand the LLM test-writer the spec as an explicit rule checklist — lifts final code correctness +38pp over an equally-budgeted edge-prompted baseline, replicated cross-vendor (GPT +28, Gemini +19). Doubling the test budget or ensembling 8 ungrounded suites barely moves; grounded tests recover 27/30 bugs vs 2/30 and cut false rejections of correct code from 33% to 0%.
open the interactive explainer →Controlled capacity sweeps show decomposition is the bottleneck in hierarchical search agents: spend your biggest model on planning and run retrieval and extraction sub-agents on cheap small ones — a distilled 1.7B executor matches a frontier sub-agent with 37% fewer sub-agent tokens.
78% of tau2-bench airline failures were silent policy-violating writes with no tool error; deterministic validators that check each proposed write against current state lift gpt-4o-mini success from 29.6% to 42.0% with no model changes. Gates add little where tools already enforce policy.
Promote repeatedly-validated agent traces into deterministic workflows and auto-demote on regression: a production AIOps system went from 0% to 45% deterministic execution in 8 months while volume doubled, and the paper supplies the promotion criteria, trace-extraction method, and economic model.
STRACE clusters redundant failure traces and prunes non-causal steps via a dependency graph before the LLM optimizer sees them, lifting success from 42.5% to 58.5% on VeruSAGE-Bench — a copyable recipe for any trace-driven agent-improvement loop.
EvoSOP mines recurring multi-step workflows from agent traces and registers them as composite tools, with a construct, merge, evaluate, prune lifecycle that keeps a growing toolset from bloating context. The abstract claims higher success and fewer rounds but reports no figures, and there is no code.
False-pass bias in an LLM judge silently disables skill retirement past a sharp threshold no data volume can cross, while symmetric noise is harmless. Before trusting any judge-gated loop, inject known-bad outputs and measure the judge's false-pass rate.