Research map / Large language models / Alignment and preferences

Simulating Students or Sycophantic Problem Solving? On Misconception Faithfulness of LLM Simulators

NeurIPS 2026 · Large language models › Alignment and preferences · approach: LLM-as-judge and feedback data · 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.

Introduces a misconception-contrastive feedback protocol and Selective Flip Score to test whether LLM student simulators update selectively, finds near-zero SFS across seven LLMs due to sycophantic re-solving, and shows SFT/DPO/GRPO post-training improves it.

Problem

LLM-based student simulators are evaluated by output similarity or accuracy, not by whether they maintain a misconception-driven belief state and revise only when feedback addresses that misconception.

Approach

Compare flip rates under targeted, misaligned (different plausible misconception), and generic feedback via the Selective Flip Score (F_T minus average of F_M and F_G), and train with SFT, DPO and an SFS-aligned GRPO reward to induce selective updating.

Key results

• Across seven LLMs (4B–120B), SFS stays near zero (mostly 0.00–0.14) under base and reflective prompts on Malrule and EEDI.
• SFT yields the largest gains: Qwen3-4B SFS improves by +0.555 on Malrule and +0.250 on EEDI over the reflective baseline.
• GRPO with an SFS-aligned reward gives more consistent improvements than DPO, increasing F_T while holding or reducing F_M in three of four settings.

Surprising findings

• Stronger models produce coherent misconception-aligned first-turn reasoning yet abandon it under any feedback, a dissociation between generated reasoning and behavior.
• Failure manifests mostly as well-engaged correct flips rather than explicit sycophantic flips, which are rare.

Limitations

• Limited to mathematics domains with taxonomy-based misconceptions.
• Feedback and response classifications are generated by LLMs rather than human-authored or verified.
• SFS measures observable behavior, not latent belief states.

Experimental setup

Datasets: Malrule EEDI

Benchmarks: Malrule simulation dataset EEDI simulation dataset

Metrics: Selective Flip Score flip rate under targeted/misaligned/generic feedback response category distribution

Baselines: Llama3.1-8B-Instruct Llama3.3-70B-Instruct Qwen3-4B Qwen3-Next-80B-A3B GPT-OSS-20B GPT-OSS-120B

Builds on: DPO GRPO Malrule EEDI misconception dataset

Closest papers by meaning

• Can LLMs Reliably Simulate Human Learner Actions? A Simulation Authoring Framework for Open-Ended Learning Environments (AAAI 2025 · similarity 0.90)
• BEAGLE: Behavior-Enforced Agent for Grounded Learner Emulation (NeurIPS 2026 · similarity 0.87)
• FEEDBACK FRICTION: LLMs Struggle to Fully Incorporate External Feedback (NeurIPS 2025 · similarity 0.86)
• Towards Understanding Sycophancy in Language Models (ICLR 2024 · similarity 0.86)
• Distilling What Matters: Confidence-Aware Selective Distillation for Large Language Models (NeurIPS 2026 · similarity 0.86)
• DistiLLM-2: A Contrastive Approach Boosts the Distillation of LLMs (ICML 2025 · similarity 0.86)