A study of autonomous LLM agents conversing on a user's behalf on matching platforms, built on two large-scale surveys of a major dating platform's active users (N=2,894 for generative profile features; N=2,617 for autonomous conversational agents, in two languages), finds that willingness to send and willingness to receive agent communication are distinct constructs — the two are separable, not one sentiment.[1] The asymmetry is large: “only 4-13% of directed dyads combine agent deployment with receiver engagement, with a pronounced gender-directional imbalance”.[1] The hard problem of agentic communication isn't building the agent — it's that other people must accept receiving agent-mediated contact.[1]
Design levers the paper quantifies: a reciprocity requirement cuts interaction volume by half or more, and routing agent contacts by receive-receptivity triples per-contact engagement (out-of-sample validated, AUC 0.88).[1]
A re-evaluation of two memory-based self-improving agents (a textual memory bank, learning from an online stream of tasks) concludes that agent evaluation is inherently noisy on multi-step tasks, and a self-improving loop amplifies that noise — the result is highly sensitive to the order in which tasks arrive.[2] Specifically, “the agent's improvement is highly dependent on task order”; a default ordering imposes an implicit curriculum that acts as a hidden prerequisite for success.[2] The paper's call: report results across multiple runs, stress-test orderings, and build in human-oversight interfaces.[2]
Prior traffic-rerouting research works in simulation but “assumes that essentially every on-road vehicle is replanned every decision period”.[3] HLSR relaxes that to a far more realistic assumption — selectively rerouting only a subset of drivers — by fusing live edge speeds with short-horizon forecasts under a limited intervention scope.[3] It's a useful example of moving a result from simulation toward the bounded conditions of the real world.[3]
[1] arxiv.org/abs/2608.18058 — Delegation Asymmetry in Agentic Recommender Systems (verified live 2026-08-20)
[2] arxiv.org/abs/2608.18066 — On the Fragility of Self-Improving Agents (verified live 2026-08-20)
[3] arxiv.org/abs/2608.18056 — HLSR: Hybrid Live Forecast Selective Dynamic Vehicle Rerouting (verified live 2026-08-20)