How much compute budget do autonomous agents waste when forced to generate lengthy natural-language thoughts for every environmental action? Engineering teams deploying autonomous agents face a stark financial reality: forcing models to output intermediate reasoning steps burns tokens and racks up compute bills without necessarily improving task outcomes.
A research paper submitted to arXiv on November 3, 2026 (arXiv:2611.05709v1) by Min-Seon Kim, Seungdong Yoa, Sanghyu Yoon, Ye Seul Sim, Suhee Yoon, Dongmin Kim, and Jung-Eun Kim from LG AI Research and North Carolina State University demonstrates that discarding teacher thoughts during agent distillation often leaves student performance intact. Standard distillation methods routinely train student models on complete teacher trajectories, supervising both intermediate thoughts and final actions. However, the study reveals that on widely adopted benchmarks, thoughts dominantly occupy the trajectory budget without providing a commensurate performance lift.
In our view, copying full teacher reasoning paths into production models risks inflating token generation costs while locking engineering teams into redundant computational overhead. The industry standard has built directly on the ReAct framework proposed by Yao and others in 2023, where operational trajectories interleave a natural-language thought, an environment-facing action, and the resulting observation. But does joint supervision of both channels actually guarantee higher task accuracy?
Discarding Teacher Thoughts
The experimental findings show that corrupting teacher thoughts preserves near-clean student performance, whereas corrupting teacher actions causes substantial degradation across model scales. In fact, a student model trained on complete teacher trajectories can follow an erroneous reasoning path and fail, while an action-only student succeeds.
"On widely adopted agent benchmarks across model scales, altering or discarding thoughts has little effect on student performance."
On a HotpotQA question asking for the founding year of the charter franchise that Marcus Wayne Garland played for, the student trained on thoughts and actions concluded that the Baltimore Orioles formed in 1903, producing an incorrect final answer. Meanwhile, the student trained strictly on actions generated no thought tokens and returned the correct founding year of 1901.
When Thought Supervision Still Matters
Thought supervision remains critical in a specific operational regime: quantitative reasoning benchmarks where information about the thought is not, or is only slightly, disclosed to the action during training. In those specific settings, discarding teacher thoughts consistently degrades student performance because the action alone fails to expose the underlying computation.
"Our central finding is that the importance of thought supervision depends on how actions are represented: specifically, on whether the information contained in the thought is already exposed in the action."
To confirm this mechanism, the researchers increased the thought information disclosed directly to the action step without altering the underlying computation, discovering that student models immediately became less reliant on thought supervision. When the action representation exposes the necessary operational information, intermediate thoughts can be omitted with little to no loss in downstream results.
Engineering workflows that default to full trajectory distillation risk allocating context budgets to reasoning steps that add no marginal accuracy. Decoupling action supervision from redundant reasoning tokens allows teams to tailor inference pipelines to whether their environment actions already expose intent.
What it signals
Why force autonomous agents to spell out every intermediate thought in natural language if it only inflates cloud inference bills? Experiments by LG AI Research and North Carolina State University clearly demonstrate that corrupting teacher actions instantly degrades student accuracy across model scales, while corrupting thoughts leaves performance virtually intact. Encoding intermediate computational data directly into the action format lets engineers bypass empty reasoning tokens entirely. Stripping intermediate thought generation in favor of action-only distillation eliminates unnecessary token costs without sacrificing compact model accuracy.