An October 14 arXiv preprint reports that CoPE, a training method that teaches language agents to forecast actions from successful runs, improved aggregate results across five benchmarks. But its results were uneven: success fell against GRPO on the Airline domain of τ²-bench. CoPE uses forecasts as a training signal; at test time, the agent still chooses one action at a time.
CoPE turns successful action sequences into a training signal
In many long tasks, such as navigating a website or calling several app interfaces, an agent receives its main reward only after the whole task ends. Group Relative Policy Optimization (GRPO), the baseline used here, adjusts a model from scores across groups of attempts. Those scores say whether an attempt succeeded, but do not directly teach the model what sequence of decisions followed each earlier point in a successful attempt.
CoPE adds that missing signal without building a separate simulator. During training, the same language-model backbone that chooses actions also predicts several actions that followed each history in successful runs. The forecast skips the observations between those actions, such as the changing page text or tool output. The agent still learns through GRPO; its additional forecast loss is trained only on successful trajectories, so the model is asked to predict action sequences associated with completed tasks.
The forecasts are not executed as a block. Once training is over, the agent uses its ordinary policy to choose one action at a time, and the forecasting instruction is removed. The authors therefore report no added inference cost, although the extra forecast pass does add work during training.
Benchmark results show uneven gains and added training time
On AppWorld, where an agent uses APIs across nine simulated consumer apps, CoPE’s Qwen2.5-14B-Instruct model completed 55.36% of tasks in the normal test split, compared with 41.67% for GRPO. On the challenge split, which holds out app-scenario combinations, success was 22.54% versus 18.71%. AppWorld scores each task as pass or fail using checks of the simulated apps’ final state; the researchers trained on 90 tasks and tested on 168 normal and 417 challenge tasks. These results show performance on that benchmark, not on live apps or customer requests.
The improvement was not uniform. On the three-domain τ²-bench customer-service benchmark, the authors' reported overall success rose from 27.5% with GRPO to 28.7% with CoPE, but Airline-domain success fell from 26.2% to 16.2% while Retail rose from 17.5% to 29.4%. The authors suggest that easy Airline tasks may skew the training signal: 38% can be solved by leaving the database unchanged, and these tasks were much easier early in training than those requiring an edit. Because CoPE learns its forecasting loss only from successful runs, those no-change cases may dominate its targets. The authors argue that a refusal learned on an easy case can carry over to a look-alike request that requires a database change. They present this as a possible explanation for the Airline decline, not a demonstrated cause.
The paper reports two timing measures from its ScienceWorld comparison. The forecasting pass added 11% to average time per training step, but the measured end-to-end step time was 9% lower than GRPO; the authors attribute that net reduction to shorter rollouts. These figures describe this ScienceWorld setup and do not establish the cost balance on other models or environments.
The authors have published a codebase on GitHub under the MIT license. The paper’s results remain benchmark experiments in an arXiv preprint; they do not show that the method improves a live assistant or generalizes beyond the tested task distributions.
Related coverage: A robot model reuses future video-and-action predictions during planning. That system recycles forecasts while a robot runs; CoPE uses action forecasts as a language-agent training signal.
Sources
- Haofei Yu et al., “Co-Evolving Acting and Forecasting for Language Agents”, arXiv v1, October 14, 2026. Method, experimental results, task setup and training-time measurements.
- CoPE code repository, public GitHub project, MIT license; README documents the implementation and environment requirements.
- Lead image: Bill Branson / National Cancer Institute, “Man uses laptop”, U.S. federal public-domain work; 1280px image. Generic illustrative photo, not part of the study.
