Research map / Large language models / Reasoning and chain of thought

DICE: Decoupling Capability from Intervention Necessity in LLM Tutoring

NeurIPS 2026 · Large language models › Reasoning and chain of thought · approach: Routing, pruning and distillation for reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DICE, which decouples whether an LLM tutor should intervene from what to say, using a rollout-based Intervention Value to weight training and a cost-sensitive action selector, evaluated on the new DICE-Bench math tutoring benchmark.

Problem

LLM tutors generate a teacher response at every student turn, leading to over-intervention and false acceptance of wrong answers; tutors lack a learned decision of when not to intervene.

Approach

Factor tutoring into an explicit pedagogical action policy (including non-intervention) and action-conditioned generation; train the policy with IV-weighted SFT plus KL regularization to SFT, where IV is a rollout-based counterfactual comparing each action to non-intervention, and select actions with a log-probability intervention cost.

Key results

• DICE-Cost achieves the best All-3 session success (0.564) and fewest turns (11.940) vs. SocraticLM (0.376 success, 14.884 turns) on DICE-Bench.
• DICE-Threshold reduces over-intervention to 0.024 in session-level tutoring, versus 0.420 for SocraticLM and 0.904 for fine-tuned MathDial.
• All three DICE variants reduce unsafe final-answer acceptance to 0, while SocraticLM reaches 0.400.

Surprising findings

• Nearly half of prescribed interventions in observed tutoring data provide zero or negative marginal benefit according to IV.
• Fine-tuned MathDial has the lowest false acceptance but over-intervenes in 0.904 of sessions.

Limitations

• Evaluation relies on simulated students (Qwen-2.5-1.5B) and LLM judges rather than real students.
• IV estimates depend on a fixed student simulator, judge, and rollout budget and may be noisy.
• Action space is limited to six pedagogical actions and benchmark is restricted to math word problems.

Experimental setup

Datasets: SocraTeach GSM8K MAWPS DICE-Bench

Benchmarks: DICE-Bench

Metrics: action accuracy action macro-F1 over-intervention false acceptance unsafe final-answer acceptance success rate average turns variants solved

Baselines: StratL MathDial fine-tuned tutor Socratic prompting SocraticLM SFT policy

Builds on: SocraticLM SocraTeach MathDial advantage-weighted regression

Closest papers by meaning

• Reward Is Enough: LLMs Are In-Context Reinforcement Learners (ICLR 2026 · similarity 0.85)
• LLM-as-a-Tutor: Policy-Aware Prompt Adaptation for Non-Verifiable RL (NeurIPS 2026 · similarity 0.85)
• Get a Head Start: On-Demand Pedagogical Policy Selection in Intelligent Tutoring (AAAI 2024 · similarity 0.85)
• CurES: From Gradient Analysis to Efficient Curriculum Learning for Reasoning LLMs (ICLR 2026 · similarity 0.84)
• Simulating Students or Sycophantic Problem Solving? On Misconception Faithfulness of LLM Simulators (NeurIPS 2026 · similarity 0.84)
• Long Is More for Alignment: A Simple but Tough-to-Beat Baseline for Instruction Fine-Tuning (ICML 2024 · similarity 0.84)