Large language models (LLMs) process gigabytes in minutes; human experts work far more slowly. Prior studies that explored combinations of human/AI optimizers ignore this gap. Those studies asked systems or humans to label too many examples (e.g., SynthCore needs over four dozen per trial). Here, we test human and AI knowledge under a budget where evaluation is practical (a few dozen labels or fewer). We compare five baselines (random search, RRP, EZR, few-shot prompting, and a budget-matched SynthCore variant) against three ways to give an LLM domain knowledge before it proposes its first configurations: (1) AMP lists the domain's rules, then checks each candidate against them; (2) HKMA adds scouted data or documentation to the prompt; and (3) H-DKP has human experts correct the LLM's rules. We find that domain knowledge improves the "start-up" phase of LLM optimization, but only when that knowledge is checked, by the LLM itself or by a human expert. Across 49 datasets, unchecked knowledge did no better than few-shot prompting or EZR. On the other hand, AMP's self-checking variant (AMP-4) reached the best statistical rank in 44 of 49 datasets (budget-matched SynthCore: 23 of 49), and H-DKP did so in 7 of 8 expert studies. These checks cost little: half a cent of LLM calls per trial, or 6 to 10 expert contacts. Hence we say LLMs may not need more data, but rather their knowledge needs checking (and checking is cheap).
Teams that adopt LLM agents for issue resolution must decide which agent to pay for, and how much to spend finding out. Leaderboard ranks give weak evidence for this decision. Accuracy differences may be statistically unresolved, reported costs may be inconsistent, and evaluating candidate agents can itself be expensive. We investigate budget-constrained agent selection using 58,197 run--task records extracted from the Holistic Agent Leaderboard (HAL), focusing on SWE-bench and six additional benchmarks. Across seven benchmarks, 27 of 35 steps along the accuracy--cost Pareto frontiers are statistically unresolved. On SWE-bench, a run costing $34.16 trails the $170.78 leader by 14.6 percentage points, but the available tasks cannot statistically distinguish their accuracy. Reported costs also differ from standardized token-based estimates by up to 11.7x, while selective task evaluation can bias estimates of full-benchmark execution cost. We therefore compare evaluation strategies under explicit evaluation budgets and deployment-cost caps. Under advance knowledge of task costs, successive halving reaches our uncertainty-aware selection target on SWE-bench using 6.6% of exhaustive evaluation spend, achieving the lowest cost on six of seven benchmarks and tying on the seventh. Its advantage narrows when task costs are unknown before execution or exact-best selection is required. Our findings motivate decision-oriented agent evaluation that accounts for accuracy uncertainty, validates execution costs, and compares complex selection methods against simple adaptive allocation.
Srinath Srinivasan, Lunxiao Li, Tim Menzies, Wesley Assunção
原文摘要
Early stopping promises to reduce compute costs for autonomous coding agents by terminating unpromising runs early. Deciding when to abandon a run requires separating baseline issue difficulty from the trajectory of an individual attempt. We study the distinction in 147,002 execution traces from two public corpora spanning four model-framework settings. Comparing pre-execution issue context against early execution prefixes, we evaluate linear count models, initial-turn text embeddings, task-conditional rankers, and a fine-tuned Qwen3-0.6B monitor. Although early prefixes modestly improve cross-task outcome ranking, within-task discrimination remains near chance across early landmarks (5k-20k tokens): initial-turn embeddings and 20k-token task-conditional models yield AUCs around 0.50, and three-seed neural evaluations across 32,406 attempts at 20k tokens achieve a mean within-task AUC of 0.50 [0.48, 0.51] across tasks of varying solve rates. To evaluate how these predictions translate into stopping decisions, we replay score-guided and matched random restarts under shared token budgets alongside uninterrupted continuation. Under a 75k-token cap, learned restarting improves resolution over matched random restarting by up to 1.98 percentage points. An exact algebraic decomposition resolves this paradox: the gain over random restarting stems almost entirely from between-task stopping allocation (avoiding stops on easy issues), whereas attempt-level selection within a task contributes near zero (-0.11 to +0.24 percentage points). Across all four settings at the 75k cap (and 216 of 224 tested budget-threshold configurations overall), uninterrupted continuation achieves higher solve rates at lower mean token cost than both restart policies. Practical evaluations of early stopping must therefore establish attempt-level discrimination and demonstrate net compute savings relative to continuation.