IBM ALTK-Evolve Measures and Fixes Non-Deterministic Agent Failures
IBM Research introduced the Consistency Analyzer in ALTK-Evolve to identify flip-prone decision steps in LLM agent execution traces. By resampling decision points offline and injecting targeted consistency guidelines, the system halved the consistency gap on AppWorld from 24.4 to 12.0 percentage points.
Impact: High
Why it matters
High average accuracy masks critical inconsistency: an agent passing 77% of runs may fail on identical inputs simply due to flat token distributions. ALTK-Evolve addresses production repeatability without retraining or model internals.
TL;DR
- 01Pass^k measures all-run reliability, uncovering substantial failure rates hidden by Mean@k averages.
- 02Flat token distributions cause execution flips at temperature 0.0 due to hosted runtime noise.
- 03ALTK-Evolve extracts targeted consistency guidelines via black-box resampling to halve the consistency gap.
Key facts
- Base Mean@5 (GPT-4.1 AppWorld)
- 77.4%
- Base Pass^5 (GPT-4.1 AppWorld)
- 53.0%
- Consistency Gap
- 24.4 percentage points
- Post-Guideline Pass^5
- 69.0%
- Post-Guideline Consistency Gap
- 12.0 percentage points
The Hidden Consistency Gap in AI Agents
Standard evaluations track Mean@k, but production systems demand Pass^k—the guarantee that an agent solves a task on every attempt. Evaluating a ReAct agent driven by GPT-4.1 on the AppWorld test_normal benchmark revealed an aggregate Mean@5 of 77.4%, but a Pass^5 of only 53.0%. This leaves a 24.4-point consistency gap where agents succeed or fail unpredictably on identical inputs.
Why Greedy Decoding Fails to Enforce Determinism
Even at temperature 0.0, agents drift due to hardware perturbations, GPU floating-point non-associativity, and hosted batching. When next-token probability distributions are flat across multiple valid tool arguments, minor numerical noise flips the execution trajectory.
Single-Trace Offline Resampling
ALTK-Evolve's Consistency Analyzer identifies fragile steps from a single execution trace without labeled ground truth: 1. Replays each decision step offline with a single model call requesting k=5 completions. 2. Computes a black-box consistency score based on output variations across resamples to pinpoint steps at risk of flipping. 3. Synthesizes targeted guidelines (e.g., anchoring regex markers or verifying multi-record queries).
Injecting these dynamic guidelines into the prompt narrowed the AppWorld consistency gap from 24.4pp to 12.0pp while raising Pass^5 to 69.0%.
✓ When to use
- Mission-critical agentic workflows like financial reconciliation or legal compliance checking where every run must succeed.
✕ When NOT to use
- Simple single-turn tasks where agent branching and multi-step tool calls are absent.
What to do today
- Audit existing ReAct agents using Pass^k across at least 5 identical runs.
- Implement black-box k-completion resampling on recorded traces to detect unstable decision points.
Sources