Technical Summary: AI Assistants Overassist

Problem Statement

Large language models (LLMs) are increasingly deployed as tutors and thought partners to scaffold user reasoning. However, the efficacy of such assistance depends critically on how the AI intervenes. Intervening too early or too frequently can usurp the reasoning process, hindering cognitive engagement and long-term learning, while waiting too long may leave users stuck. Current literature primarily measures the downstream outcomes of AI assistance (e.g., reduced autonomy or motivation) but lacks a granular understanding of the specific assistance behaviors—such as the timing, frequency, and content of interventions—that drive these effects. There is a need to characterize whether LLMs optimize for immediate task success at the expense of the "productive struggle" required for deeper learning.

Methodology: INT-BENCH

To address this gap, the authors introduce INT-BENCH, a simulation-based benchmark that formalizes LLM assistance as a sequential intervention game. The framework involves three primary components: a student model, a teacher model (the AI assistant), and a judge model.

Experimental Setup

1. Problem Domains: The study evaluates interventions across three domains: Code Debugging (500 problems), Mathematics (500 problems), and Brain Teasers (500 problems).
2. Student Model: Qwen2.5-7B-Instruct is used as the student, prompted to generate step-by-step reasoning traces.
3. Teacher Models: The study evaluates four teacher models: GPT-5.2, Gemini 3 Flash, GPT-OSS-120B, and DeepSeek-V3.2.
4. Intervention Conditions:
   • Standard Condition: The teacher observes the student's reasoning trace in cumulative increments (50 characters) and decides at each step whether to wait or intervene. If they intervene, a message is injected, and the student updates its reasoning. The teacher may intervene at most once.
   • Oracle Condition: The teacher receives the full baseline trace, the final answer, and the correctness verdict before deciding whether and when to intervene (post-hoc optimization).
5. Metrics:
   • Intervention Frequency ($\phi$): The proportion of episodes where the teacher intervenes.
   • Intervention Timing ($\tau$): The absolute and relative point in the reasoning trace where intervention occurs.
   • Immediate Helpfulness ($H$): The net change in answer correctness (correct $\to$ correct, incorrect $\to$ correct, etc.) following an intervention.
   • Generalization Helpfulness ($G$): The change in performance on a new, related problem (variant) when the student is provided with context from the original problem (either just the problem/solution or the full assisted episode).
6. Human Comparison: A human study was conducted with 50 participants acting as teachers on 30 brain teaser problems, using the same reasoning traces generated by the student model to allow direct comparison with LLM behavior.

Generalization Evaluation

To test transfer, the authors generate variant problems ($\tilde{q}$) that require the same underlying skill as the original question ($q$) but differ in surface form. Students are tested on $\tilde{q}$ under three context conditions:

• No-Context: Solving from scratch.
• Problem-Context: Seeing the original problem and unassisted solution.
• Intervention-Context: Seeing the original problem, the teacher's intervention, and the revised reasoning.

Key Contributions

1. Formalization of Assistance: The paper formalizes AI assistance as a sequential intervention game, moving beyond static evaluation to dynamic, step-by-step monitoring.
2. INT-BENCH Benchmark: Introduction of a simulation-based benchmark that evaluates LLM teachers on frequency, timing, immediate helpfulness, and generalization impact.
3. Human-LLM Comparison: A direct empirical comparison between LLM intervention strategies and human pedagogical strategies in a controlled setting.
4. Behavioral Characterization: Development of metrics to quantify "solution leakage" (how much the intervention reveals the answer) and functional roles of interventions (e.g., redirecting vs. hinting).

Results

1. Frequency and Timing (RQ1)

• LLMs Intervene Frequently and Early: In the Standard condition, LLM teachers intervened in approximately 90% of trials ($\phi = 0.90$) and did so very early in the reasoning process ($\tau_{rel} = 0.18$).
• Over-intervention on Correct Answers: LLMs often intervened even when the student would have solved the problem correctly without help (e.g., GPT-5.2 intervened on 98% of initially correct answers).
• Human Difference: Humans intervened significantly less often ($\phi = 0.74$) and much later ($\tau_{rel} = 0.74$). Humans were more likely to remain silent when the student was on the right track, whereas LLMs felt compelled to intervene.
• Oracle Effect: When given full information (Oracle condition), LLMs became more selective (intervening in 54% of cases) and waited longer ($\tau_{rel} = 0.56$), rarely intervening on correct answers. Humans showed minimal change in behavior between conditions.

2. Immediate Helpfulness (RQ2)

• Moderate Gains: Standard interventions resulted in a net accuracy gain of $H = 0.20$. Oracle interventions were more helpful ($H = 0.30$) and rarely harmful.
• Overly Informative Feedback: Interventions often provided complete solutions or "near-complete" scaffolding rather than targeted hints. In the "Stop-and-Answer" ablation (where the student must answer immediately after intervention without further reasoning), helpfulness remained high for code and brain teasers, suggesting the interventions often simply revealed the answer rather than guiding the reasoning process.

3. Generalization (RQ3)

• No Reliable Transfer: Interventions did not consistently improve the student's ability to solve related, unseen problems. The "Intervention-Context" condition yielded no significant accuracy gains over the "No-Context" baseline across any domain.
• Instance-Specificity: The authors suggest that LLM interventions are too specific to the original problem instance, failing to provide transferable strategies. In some cases, the added context even slightly reduced performance, likely due to distraction by irrelevant details.

4. Human vs. LLM Behavior (RQ4)

• Qualitative Differences:
  • Solution Leakage: LLMs revealed the full solution or near-complete scaffolding roughly twice as often as humans (14.2% vs. 5.4% in Standard). Humans relied more on "wrong-path rejections" and general reframing hints.
  • Pedagogical Nuance: Humans validated productive reasoning and encouraged persistence; LLMs rarely did. Humans also intervened to flag flawed intermediate steps even when the final answer was correct, whereas LLMs focused almost exclusively on the final correctness.

Significance and Claims

The paper claims that current LLM assistants tend to optimize for short-term task success rather than supporting the reasoning processes necessary for deep learning. By intervening too early, too frequently, and with overly informative feedback, LLMs may inadvertently deprive learners of the "productive struggle" required to build independent problem-solving skills.

The authors argue that this behavior contributes to a form of "gradual disempowerment," where users become increasingly dependent on AI for reasoning rather than developing their own cognitive capacities. While AI assistants are effective at correcting specific instances, they currently lack the pedagogical sophistication to balance immediate help with long-term learning goals, such as generalization and autonomy. The study highlights the need for future AI systems to be trained to maximize user empowerment and to support transferable learning strategies rather than simply providing answers.