benchmarks evals
Agents Can Be Mapped Without Running New Rollouts
Centered action likelihoods on shared trajectories locate model-and-harness combinations by what they would do.
Summary
Centered action likelihoods on shared trajectories locate model-and-harness combinations by what they would do.
Agent Behavior Maps compare action log-likelihoods on the same reference trajectories instead of launching each target agent in a fresh environment. Across eight cross-scored models, distances followed measured trajectory divergence in rank order. Across 100 language models and four environments, ALFWorld-derived coordinates predicted performance and exposed behavioral similarities that persisted after controlling for success rate and model size. The guarantees depend on explicit assumptions, and the map is not a substitute for deployment testing.
Why it matters
Centered action likelihoods on shared trajectories locate model-and-harness combinations by what they would do.
Limits and context
- The guarantees depend on explicit assumptions, and the map is not a substitute for deployment testing.
Key claims
Centered action likelihoods on shared trajectories locate model-and-harness combinations by what they would do.
Qualification: The guarantees depend on explicit assumptions, and the map is not a substitute for deployment testing.
Evidence: source-2026-10-22-014
Sources
- arXiv preprint 2610.30014, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.