Addressing LLM Agent Unpredictability with ALTK-Evolve
IBM Research introduces consistency guidelines and the Consistency Analyzer within ALTK-Evolve to measure and reduce reliability gaps in LLM agents.
According to a Hugging Face Blog post by IBM Research, standard AI benchmarks often hide significant performance variability. When building practical AI systems, developers frequently rely on single-run success metrics that do not reflect how an agent behaves across repeated executions.
The Consistency Gap in AI Agents
According to the source text, standard benchmarks report average success rates (Mean@k) which conceal variability across multiple runs. For example, testing a ReAct agent using GPT-4.1 on AppWorld showed a Mean@5 success rate of 77.4%. However, the Pass⁵ metric—meaning the agent succeeded on all five repeated runs—was only 53.0%. This reveals a 24.4-percentage-point consistency gap.
The consistency gap represents the difference between average success rates and complete multi-run success. For instance, if an agent resolves a user request on the first try 77% of the time, but only succeeds consistently across five identical attempts 53% of the time, it suffers from high variance.
How the Consistency Analyzer and Guidelines Work
To fix unstable behavior, the source notes that the system uses a diagnostic mechanism called the Consistency Analyzer. Instead of executing full, end-to-end task rollouts, the analyzer takes a single recorded agent trajectory and resamples each decision step k times (typically k=5) in a single model call. This black-box process requires no ground truth.
When the system flags an unstable decision step, ALTK-Evolve automatically converts it into a reusable guideline. These guidelines are then injected at inference time to steer the agent away from flip-prone choices, such as ambiguous string-counting methods.
Applying these consistency guidelines cut the consistency gap from 24.4 percentage points down to 12.0 percentage points on AppWorld test_normal. Aggregate Pass⁵ increased from 53.0% to 69.0%, and Mean@5 rose from 77.4% to 81.0%.
Why a Developer Might Care
For developers deploying AI agents to production, predictability is critical. An agent that passes a test once may fail a user's identical request later due to token probability distributions. Tracking Pass^k exposes this hidden unreliability.
Furthermore, the source states that guidelines derived from a single baseline trajectory transferred successfully to similar tasks within the same scenario. This raised Pass⁵ on similar tasks by +13.0 percentage points for GPT-4.1 and +8.7 percentage points for gpt-oss-120b. Developers can diagnose unstable behaviors without re-running expensive full environments.
Limitations
The source outlines two primary limitations for developers to keep in mind:
- Model Sensitivity: Lower-baseline models, such as
gpt-oss-120b, started from a much lower same-task Pass⁵ baseline of 10.1%. This indicates that the effectiveness of the injected guidelines depends heavily on the base model's underlying capacity to follow instructions. - Probabilistic Variance: Even when running models at temperature zero (0.0), platform-side effects and hosted endpoint probability shifts can cause near-ties in token generation to resolve differently across runs. This necessitates external stabilization mechanisms like guidelines.