This Week in Agentic AI: Selection Beats Exhaustive Search

Tsinghua University

Selecting complementary LLM skills improves success while reducing context use — Treating skill retrieval as budget-constrained set selection rather than independent relevance ranking raised task success to 0.73 on a contamination-controlled BigCodeBench variant, versus 0.20–0.52 for prior selectors, while using fewer tokens than the strongest released router.

Google DeepMind

Cost-aware model routing matches exhaustive estimation with fewer expensive checks — A centralized search policy retained the routing quality of exhaustive value estimation across multi-LLM, specialist-retrieval, and reasoning-budget settings while invoking costly estimators substantially less often, making the cost of deciding part of the routing objective itself.

The University of Tokyo

Adaptive task selection cuts the cost of optimizing LLM agent harnesses — Evaluating candidate harnesses primarily on tasks where previous candidates disagreed substantially reduced optimization cost while preserving final performance, offering a practical alternative to repeatedly running full benchmark suites.

HKUST + HKBU + NTU

Financial agents can cite rules while still attempting prohibited trades — Showing agents compliance rules reduced but did not eliminate rejected orders, and monitors were susceptible to persuasive trader rationales when enforcement evidence was hidden, demonstrating that verbal rule awareness is not executable compliance.

Creative AI & Agentic Generation

Temperature-controlled preference updates reduce manifold drift in flow models — Zhejiang University, Kuaishou Technology, and Westlake University show that reward optimization can move flow-model outputs away from the pretrained data manifold; their ThermoDPO-weighted objective improved both a controlled benchmark and Stable Diffusion 3.5 Medium text-rendering metrics.

Explicit identity-layout planning improves group images with specified people — Fudan University, Tencent, and the University of Hong Kong explicitly plan identity placement before synthesis, achieving 97.3% requested-identity coverage with a 2.8% duplicate rate and higher face similarity than GPT-Image-2 on an identity-disjoint benchmark.

Selection is not ranking; it is budgeted search. Tsinghua’s skill-selection paper optimizes complementary documents as a set, Google DeepMind’s router prices the cost of estimating each model, and the University of Tokyo’s harness optimizer concentrates evaluation on disagreement cases. Together they show that independently scoring every option—or exhaustively testing every candidate—wastes budget and can produce worse decisions; teams should optimize the expected value of each additional retrieval or evaluation before paying for it.

Intermediate structure is not overhead; it is the control surface. MileGPO from Beijing Jiaotong University and Peking University inserts local credit into long-horizon trajectories, while the Fudan–Tencent–HKU group-image system inserts an identity–layout plan before generation. Tsinghua’s complementary skill sets provide the same lesson at context assembly time: explicit intermediate structure makes global objectives easier to satisfy, so teams should expose milestones, plans, and resource allocations rather than relying only on end-to-end prompting or terminal rewards.

Instruction awareness is not constraint satisfaction. The HKUST–HKBU–NTU financial agents could cite trading rules while still submitting prohibited orders; Zhejiang University, Kuaishou, and Westlake similarly show that preference rewards can push a flow model off its pretrained manifold. In both action and generation, optimizing a persuasive or high-reward output does not preserve the feasible set, so teams should enforce critical constraints through executable checks, constrained objectives, or rejection layers rather than natural-language guidance alone.

Evaluation should target informative failures, not average cases. The University of Tokyo selects tasks where candidate harnesses disagree, Tsinghua removes benchmark contamination when measuring skill retrieval, and the financial-compliance benchmark withholds or exposes enforcement evidence to reveal monitor failure modes. These studies converge on evaluation as an active diagnostic process; teams should prioritize disagreement, leakage controls, and adversarial evidence conditions instead of repeatedly averaging over static suites.

Explanations are not evidence. In the HKUST–HKBU–NTU benchmark, monitors could be misled by trader rationales when direct enforcement evidence was unavailable, while the Tokyo harness method bases selection on observed candidate behavior rather than self-reported quality. Agent oversight should therefore privilege action traces, tool results, and external state over generated justifications.

ZOTPAPER