Closing the Consistency Gap: Self-Evolving Agents That Learn to Stay on Course

2609.08832 ↗ · 2026-09-08 · cs.AI

Evelyn Duesterwald, Benjamin Elder, Lilian Ngweta, Shashanka Ubaru, Malgorzata Zimon

The finding

Headline claim

A self-evolving framework that detects flip-prone steps in agent trajectories and injects prompt-level guidelines raises all-five-run consistency (Pass^5) on AppWorld by +16 points same-task and +13 points on similar tasks, with no change to the underlying model or agent architecture.

The test

ReAct agents on the 168-task AppWorld test_normal split, k=5 runs per task, comparing GPT-4.1 and GPT-OSS-120B; a black-box Consistency Analyzer resamples each inference point and an offline LLM generator converts flagged steps into episodic-memory guidelines.

Effect size & scope

Baseline consistency gap is 24.4 pp for GPT-4.1 (Mean@5 77% vs Pass^5 53%) and 23.8 pp for GPT-OSS-120B (34% vs 10%); normalized consistency is 0.69 vs 0.30. The intervention moves Pass^5 +16 pp same-task and +13 pp similar-task on GPT-4.1. Evidence holds on one benchmark, one agent architecture (plus informal CUGA checks), and 168 tasks; trajectory-consistency as outcome predictor is weak (AUROC 0.69).

Limitations

Authors concede the gap was not quantified for other architectures or models, mean-aggregation ignores dependence between steps, and guideline accumulation imposes a cold-start delay with a shrinking inconsistent-task population limiting further gains. Unstated: single benchmark, possible prompt-injection of guidelines amounts to benchmark-specific memorization rather than general reliability, no statistical significance reported, and similar-task 'generalization' uses AppWorld's own scenario variants, a narrow notion of generalization.

Counterpoint

AppWorld is a simulated app environment; guidelines retrieved by cosine similarity from the same task family may amount to test-set tuning dressed as self-evolution. No baseline comparison against simply prompting the agent to 'be careful' or against majority-voting across runs, which would attack the same problem far cheaper. n=168 tasks and no variance estimates; AUROC 0.69 is barely better than a coin-plus. Vendor-adjacent IBM research publicizing a framework with an associated open-source tool.

Economy implication

Little or nothing yet in direct economic terms, but the paper documents that headline agent accuracy numbers overstate what a production user experiences by roughly 20 to 25 points, which matters for anyone pricing or contracting on agent reliability; if the fix generalizes, reliability (not raw capability) becomes the enterprise differentiator in the next 12 to 24 months.

Who this matters for

Technical Capability: Engineers deploying agents get a quantified reminder that Mean@5 accuracy is the wrong production metric, plus a plausible prompt-level mechanism for closing the gap without retraining.

Market & Valuation: Enterprises evaluating agent vendors should demand Pass^k-style repeat-reliability numbers, since advertised accuracy can overstate delivered reliability by over 20 points on this benchmark.

Bears on (advisory · papers never write evidence)

[1.3] Agentic reliability (multi-step completion without human correction) is improving materially, not just peak single-shot.

[B1] Agentic reliability gains will translate into enterprise ROI.

Abstract

Large language model (LLM)-powered agents can be accurate on average yet unreliable in production, a discrepancy that has been observed but remains largely unaddressed. When given the same task five times, a ReAct agent on the AppWorld benchmark using GPT-4.1 succeeds in all five runs only 53% of the time, even though its per-run pass rate averages 77%. We call this 24-point shortfall the consistency gap, and we argue that addressing it is a precondition for trustworthy AI agent deployment. We present a self-evolving agent framework that reduces this gap by identifying unstable, low-consistency steps in agent trajectories and converting them into episodic memory the agent can draw on in future runs. At its core is a Consistency Analyzer that pinpoints where and why a trajectory is likely to flip across executions, and a Guideline Generator that converts the diagnosis into targeted guidelines, committed to memory and injected into future agent executions on similar tasks. On AppWorld with ReAct/GPT-4.1, our framework raises the fraction of tasks that succeed in all five runs by +16 points on same-task evaluation and +13 points on similar-task generalization.