AI agents may not need every reasoning trace their teachers produce

Training a smaller AI agent to imitate a larger one often involves copying an entire interaction: a written thought, an action and the environment’s response. But does the student need supervision on every explanation? In the November 3, 2026 preprint Agents, Think Only When it Helps, Min-Seon Kim and colleagues at North Carolina State University and LG AI Research report that the answer depends on how much useful information the demonstrated actions already contain.

Their finding is not that reasoning is unnecessary. It concerns whether students must learn to reproduce a teacher’s explicit reasoning text to acquire effective behavior. A teacher may use reasoning to produce a successful action even when copying that reasoning adds little to the student’s training.

Separating explanations from actions

The researchers modified training trajectories while keeping other components fixed. They removed teacher thoughts, replaced them with thoughts from unrelated examples within the same benchmark, or paraphrased them. Separate experiments removed or replaced actions while preserving thoughts and observations.

The default setup used Qwen2.5-32B-Instruct as the teacher and Qwen2.5-3B-Instruct as the student. Only successful trajectories were retained for training. Comparisons used the same underlying examples and matched training settings, with results averaged over three independent training and evaluation runs.

On HotpotQA, FEVER, WebShop and ALFWorld—covering question answering, fact verification, shopping and text-based household tasks—thought interventions generally left performance near the full-trajectory baseline. For the default student, removing all thoughts changed HotpotQA accuracy from 48.0% to 46.8%, while WebShop success rose from 40.2% to 43.1%.

Damaging actions was much more disruptive. Removing or replacing all action targets produced zero success on WebShop and ALFWorld. The contrast suggests that, in these experiments, preserving demonstrations of what to do mattered substantially more than preserving the accompanying explanations.

The authors also tested different student sizes, model families, teachers and training objectives. They report broadly similar resilience to thought corruption, although that does not mean every intervention preserved performance exactly.

When the action already carries the message

The proposed explanation is information overlap. A search query can repeat the entity and relationship identified in the preceding thought. A household command can explicitly name the object and destination. In these “high-disclosure” settings, actions already expose much of the information that thoughts contribute.

The pattern changed on FinQA, MultiHiertt, MathQA and GeoQA. Here, the researchers constructed executable actions from dataset-provided mathematical programs. Anonymous intermediate variables showed the calculation but offered little explanation of what the quantities represented. Accompanying thoughts were generated afterward from verified computations, rather than collected as spontaneous reasoning during problem solving.

Removing thoughts reduced accuracy across every reported benchmark and student-size combination in these quantitative settings. For the 14-billion-parameter student on MathQA, accuracy fell from 41.2% to 13.5%. The paper evaluated MathQA without multiple-choice options and used a text-only GeoQA subset solvable without diagrams—important boundaries on interpreting those results.

Making actions more informative

To investigate the overlap explanation more directly, the team replaced anonymous program variables with descriptive names while preserving the numerical computation. Rewritten programs were executed to check that their outputs matched the originals.

On FinQA with the default student, action-only accuracy rose from 47.7% with anonymous variables to 61.8% with semantic names. Corresponding full-trajectory accuracy increased from 63.3% to 65.3%. Thus, action-only training retained 94.6% of its matched baseline after renaming, versus 75.4% before. These retention figures are relative performance ratios, not absolute accuracy.

Names drawn from thoughts but mismatched to the quantities did not reproduce the semantic naming benefit. Another control suppressed thought generation in fully trained students: they still outperformed action-only students on all four quantitative benchmarks. That supports a training benefit beyond simply producing extra reasoning text during evaluation.

Efficiency with conditions attached

In the four main high-disclosure benchmarks, action-only distillation reduced measured training time by 5.3–10.2% and inference time by 11.1–41.7%. Those are wall-clock savings in the reported setup, not universal deployment estimates.

The evidence supports selective supervision, not a blanket instruction to eliminate explanations. Results come from particular models, datasets and action representations; the quantitative trajectories also use retrospectively generated reasoning. Within those boundaries, the preprint suggests a practical question: what useful information would disappear if the thought were removed? When the action already preserves it, copying the explanation may be optional. When it does not, removing that supervision can be costly.