Hugging Face Post Introduces Consistency Guidelines to Close AI Agent Reliability Gap

The post defines that gap using AppWorld. A ReAct agent backed by GPT-4.1 achieved a Mean@5 of 77.4% across five repetitions on test_normal, but its Pass^5 was only 53.0%, meaning it succeeded in all five runs for just over half of tasks. That 24.4-point difference is what the post calls the consistency gap. On hard tasks, the post says, the gap reaches 30 points.

Most benchmarks report Mean@k, the average pass rate over k runs, with k often set to 3 or just 1. The post says this hides variability. Pass^k is different from Pass@k: Pass@k asks whether at least one of k attempts succeeds, while Pass^k requires every attempt to succeed. The post states that Pass^k is less than or equal to Mean@k, which is less than or equal to Pass@k. A user who asks the same question again cares about Pass^k, not the average, according to the post.

The post attributes the variability to the shape of the probability distribution at each decision. A sharp distribution concentrates mass on a single token, so the same choice tends to recur. A flat distribution spreads comparable mass across near-tied tokens, making the outcome sensitive to small perturbations. Because an agent trajectory chains dozens of decisions, a small per-step chance of flipping compounds into a larger chance that a run goes differently. The post says greedy decoding and a fixed seed govern how a distribution becomes a token but say nothing about the distribution itself; on a hosted endpoint, probabilities shift slightly from run to run, so temperature zero can still resolve a near-tie one way today and another way tomorrow.

To find flip-prone steps, the post introduces the Consistency Analyzer. It resamples an agent's own recorded trajectory to find decision points where the model was one token sample away from doing something different. It needs one trace and no ground truth, and it resamples each decision point with a single call requesting k completions, with k set to 5 by default, rather than re-running the task end to end.

Consistency guidelines are a new guideline type in altk-evolve built on top of the analyzer. They turn the diagnosis into reusable guidelines injected at inference time. The post says this halves the consistency gap from 24.4 percentage points to 12.0 points, with same-task Pass^5 improving by 16.0 points and similar-task performance by 13.0 points, without costing anything in average accuracy. The full methodology and evaluations are in a technical report on arXiv.

The post says this is not a capability problem that a larger model solves. It describes consistency as an orthogonal axis: an agent can be capable and inconsistent at the same time. An earlier post introduced ALTK-Evolve, which distills an agent's past trajectories into reusable guidelines and improves task success, but those results asked the average-case question. The new consistency guidelines target the gap directly.