Research Digest — 2026-10-09 (Data Scientist)

Scope: arXiv cs.AI / cs.CL / cs.MA, submissions Oct 1–9, 2026. Selection lens: LocalKin multi-agent orchestration, agent memory, agent evaluation, agent safety.

ID integrity: every listed ID's YYMM prefix (2610 = Oct 2026) matches its announced submission date. Checked the submission-date field on each abstract page — this proves existence + filing month, not that titles/authors/claims match.

Multi-agent workflows usually train only the workflow generator while other agents stay frozen. FloWright makes the whole workflow a training harness — hierarchical, structure-aware reward lets roles self-evolve or co-evolve with no extra models/labels/executions. Co-evolving more roles (+5.03%) beat optimizing one role (+2.83%); small open models improved up to +7.41%. LocalKin: retrain each specialist, not just the conductor. Low cost if we have traces + a signal.

Under finite context + repeated retrieval, agent memory concentrates on a core while rare states fall into a long tail where prediction errors pile up. CTWM (rank-based memory controller, single exponent) cut tokens 5.9% and tail prediction error 13.6% on Synthetic Graph World; 24.48% token reduction on LongMemEval with accuracy parity. LocalKin: ready-made knob for token cost vs. recall; our agents carry long histories. (Submitted Jul, listed Oct — noted, not discarded.)

Enforces stateful policies via regex + stateful predicates, deferred generation, scoped semantic checks. Even fully hijacked, rules out 93–95% of AgentDojo attacks and 62–85% of Toolathlon attacks — twice allowlists on long-horizon tasks — within a few percent of unconstrained utility. LocalKin: guardrails that respect task state, not static allowlists. Good safety baseline.

8,504 NL→CLI pairs across 1,642 tools; no open-weight model exceeds 42% exact-command accuracy unrestricted. Enables runtime-free verifiable rewards — an 8B model trained on them matches a 685B MoE. LocalKin: the verifiable-reward angle transfers to any agent with deterministic tool outputs.

Multi-agent improved weighted LLM-judge 1.695→1.797 (p=0.027) on compound queries, but latency ×1.31, cost ×2.02, critical-failure rates similar (15.0% vs 13.3%). LocalKin: a sober A/B case study — define the metric first, watch the cost/latency axis.

Honorable mentions: Causal World Models (2610.00012); Reasoning-Approach vs Topic transfer (2610.00331); Rules-to-Tools executable checks (2610.00313).

Figures are author-reported [author-reported]; verified = arXiv existence + submission-date field only.