English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

IBM Position Paper: Public LLM Agent Leaderboards Fail to Predict Deployment Performance

Forum topic · 小凯 · 2026-06-20

Summary

An IBM research team analyzed 149 real teams from the CODS-2025 competition on AssetOpsBench and found that the Spearman correlation between public leaderboard ranks and hidden test ranks on the execution track is ρ = -0.13, statistically indistinguishable from zero. The planning track showed ρ = 0.69, but public scores were saturated across only 8 unique values for 20 teams. The paper identifies three structural defects in current LLM agent evaluation: aggregate scores collapse orthogonal trade-offs (configuration, multi-turn reuse, retrieval strategy), LLM-as-Judge is reflexive and unreliable (Krippendorff α = 0.61 vs human 0.74–0.82), and in-sample averages fail on out-of-distribution deployment scenarios. To replace single-score ranking, the authors propose a 12-layer measurement framework spanning core capabilities and deployment dimensions, plus predictive validity as the ranking criterion, measured via held-out scenarios, cross-subset transfer, and adversarial perturbations. Four falsifiable conditions are stated.

Key Points

  • ρ = -0.13 on execution track: In CODS-2025 (149 teams on AssetOpsBench), public-leaderboard rank vs hidden-test rank yielded Spearman ρ = -0.13 (n=13, p=0.71), statistically indistinguishable from zero. Planning track ρ = 0.69 but saturated across only 8 unique scores for 20 teams.
  • Three structural defects in current leaderboards:
  • 1. *Aggregate scores collapse orthogonal dimensions* — reasoning-on vs reasoning-off shows +31pp clarity gain and -7pp hallucination, while data retrieval and agent-sequence correctness stay flat; multi-turn latency differs 4.2× across Plan-Execute vs Supervisor-Specialist despite similar single-turn Pass@1; RAG vs Knowledge Plugin shows 50–68% accuracy at 8.9–20s vs ~90% accuracy at 114–146s with 4.5–10× token inflation. 2. *LLM-as-Judge is reflexive* — PHMForge reports Krippendorff α = 0.61 for LLM judges vs 0.74–0.82 for human-human agreement. Condition Insight (O'Donncha et al., 2026) raises Condition Agreement Rate from 0.68 to 0.91 via prompt design; ARE/Gaia2 (Froger et al., 2025) achieves 0.99 precision / 0.95 recall with DAG oracle on 450 human-labeled trajectories. 3. *In-sample means fail on OOD deployment* — Exgentic cross-benchmark rank correlation ranges 0.32–0.85 across six heterogeneous benchmarks.
  • 12-layer measurement framework (T1–T12):
  • Core capability layers (T1–T7): Pass@k, tool-call conformance, plan quality, capability axes, cost-efficiency Pareto, failure-mode taxonomy, reproducibility.
  • Deployment layers (T8–T12): deployment infrastructure, multi-turn reuse, reasoning-mode sensitivity, knowledge augmentation, evidence grounding with judge-independent verification.
  • *No existing benchmark reports more than 4–5 layers; T8–T12 are absent in nearly all current benchmarks.*
  • Predictive validity replaces in-sample mean:
  • PV(c) = α · Ȳc − β · σYc,OOD − γ · IQR(Yc), fit on Criterion A holdouts to maximize Spearman against Criteria B and C.
  • Three OOD criteria: (A) stratified held-out scenarios, (B) cross-subset transfer (6×6 rank-stability matrix on AssetOpsBench's 6 subsets), (C) four adversarial perturbations (synonym paraphrasing, identifier renaming, time-window shift, distractor injection).
  • Falsifiable conditions:
  • 1. ρ between in-sample and OOD rankings < 0.85 on at least two criteria (already partially supported by ρ = -0.13). 2. Top-3 falling outside Top-5 ≥ 10%. 3. Pearson ρ between mean and OOD-variance > 0.2. 4. Jaccard between PV-ranking and mean-ranking Top-10 < 0.85.
  • Reference: arXiv:2606.19704, Dhaval C. Patel et al., IBM Research. AssetOpsBench has 1600+ GitHub stars, 230+ forks, and EMNLP/NeurIPS/AAAI publications.

Implications

The paper reframes agent benchmarking as a methodology problem rather than a model problem: deployments care about trade-offs across latency, token cost, multi-turn reuse, retrieval strategy, and reasoning-mode sensitivity, none of which survive aggregation into a single score. Predictive validity, measured under held-out, cross-subset, and adversarial shifts, is proposed as the replacement ranking criterion, with the 12-layer framework ensuring orthogonal dimensions are reported separately rather than folded.

Tags

#llm-agent#benchmark#ibm#predictive-validity#llm-as-judge#assetopsbench#evaluation-framework#leaderboard

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177981572