RedundancyBench (Window-to-One): leaderboard
RedundancyBench (Hu et al., 2026) asks an LLM to spot redundant tool calls (duplicated, failed, irrelevant or exploratory) in successful customer-service agent trajectories from tau2-bench, labelled step by step by experts. With the Window-to-One strategy the model judges each agent step with the three preceding and three following steps as context; trajectory-level accuracy of flagging whether a trajectory contains a redundant step, in percent averaged over the airline, retail and telecom domains, higher is better.
Metric: Trajectory-level accuracy (%; mean over the airline, retail and telecom domains of the share of successful tau2-bench trajectories (Qwen-3.6-Plus agent, synthetic redundant steps inserted, three-round expert annotation) correctly classified as containing a redundant step or not; Window-to-One: the LLM judges each step with the three preceding and three following steps as context). Source: arxiv.org. Saturation forecast: Around February 2027. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (Non-reasoning) | 70.88 | #121 (GPT-5.4) |
| 2 | GPT-4o | 64.81 | #741 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=redundancybench-window-to-one · How It Works · Data refreshed daily, snapshot 2026-11-12.